literature review with artificial intelligence

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• v.13(7); 2023
• PMC10335470

Artificial intelligence in systematic reviews: promising when appropriately used

Sanne h b van dijk.

1 Health Technology & Services Research, Technical Medical Centre, University of Twente, Enschede, The Netherlands

2 Pulmonary Medicine, Medisch Spectrum Twente, Enschede, The Netherlands

Marjolein G J Brusse-Keizer

3 Medical School Twente, Medisch Spectrum Twente, Enschede, The Netherlands

Charlotte C Bucsán

4 Cognition, Data & Education, Faculty of Behavioural, Management & Social Sciences, University of Twente, Enschede, The Netherlands

Job van der Palen

Carine j m doggen.

5 Clinical Research Centre, Rijnstate Hospital, Arnhem, The Netherlands

Anke Lenferink

Associated data.

Systematic reviews provide a structured overview of the available evidence in medical-scientific research. However, due to the increasing medical-scientific research output, it is a time-consuming task to conduct systematic reviews. To accelerate this process, artificial intelligence (AI) can be used in the review process. In this communication paper, we suggest how to conduct a transparent and reliable systematic review using the AI tool ‘ASReview’ in the title and abstract screening.

Use of the AI tool consisted of several steps. First, the tool required training of its algorithm with several prelabelled articles prior to screening. Next, using a researcher-in-the-loop algorithm, the AI tool proposed the article with the highest probability of being relevant. The reviewer then decided on relevancy of each article proposed. This process was continued until the stopping criterion was reached. All articles labelled relevant by the reviewer were screened on full text.

Considerations to ensure methodological quality when using AI in systematic reviews included: the choice of whether to use AI, the need of both deduplication and checking for inter-reviewer agreement, how to choose a stopping criterion and the quality of reporting. Using the tool in our review resulted in much time saved: only 23% of the articles were assessed by the reviewer.

The AI tool is a promising innovation for the current systematic reviewing practice, as long as it is appropriately used and methodological quality can be assured.

PROSPERO registration number

CRD42022283952.

Strengths and limitations of this study

• Potential pitfalls regarding the use of artificial intelligence in systematic reviewing were identified.
• Remedies for each pitfall were provided to ensure methodological quality. A time-efficient approach is suggested on how to conduct a transparent and reliable systematic review using an artificial intelligence tool.
• The artificial intelligence tool described in the paper was not evaluated for its accuracy.

Medical-scientific research output has grown exponentially since the very first medical papers were published. 1–3 The output in the field of clinical medicine increased and keeps doing so. 4 To illustrate, a quick PubMed search for ‘cardiology’ shows a fivefold increase in annual publications from 10 420 (2007) to 52 537 (2021). Although the medical-scientific output growth rate is not higher when compared with other scientific fields, 1–3 this field creates the largest output. 3 Staying updated by reading all published articles is therefore not feasible. However, systematic reviews facilitate up-to-date and accessible summaries of evidence, as they synthesise previously published results in a transparent and reproducible manner. 5 6 Hence, conclusions can be drawn that provide the highest considered level of evidence in medical research. 5 7 Therefore, systematic reviews are not only crucial in science, but they have a large impact on clinical practice and policy-making as well. 6 They are, however, highly labour-intensive to conduct due to the necessity of screening a large amount of articles, which results in a high consumption of research resources. Thus, efficient and innovative reviewing methods are desired. 8

An open-source artificial intelligence (AI) tool ‘ASReview’ 9 was published in 2021 to facilitate the title and abstract screening process in systematic reviews. Applying this tool facilitates researchers to conduct more efficient systematic reviews: simulations already showed its time-saving potential. 9–11 We used the tool in the study selection of our own systematic review and came across scenarios that needed consideration to prevent loss of methodological quality. In this communication paper, we provide a reliable and transparent AI-supported systematic reviewing approach.

We first describe how the AI tool was used in a systematic review conducted by our research group. For more detailed information regarding searches and eligibility criteria of the review, we refer to the protocol (PROSPERO registry: CRD42022283952). Subsequently, when deciding on the AI screening-related methodology, we applied appropriate remedies against foreseen scenarios and their pitfalls to maintain a reliable and transparent approach. These potential scenarios, pitfalls and remedies will be discussed in the Results section.

In our systematic review, the AI tool ‘ASReview’ (V.0.17.1) 9 was used for the screening of titles and abstracts by the first reviewer (SHBvD). The tool uses an active researcher-in-the-loop machine learning algorithm to rank the articles from high to low probability of eligibility for inclusion by text mining. The AI tool offers several classifier models by which the relevancy of the included articles can be determined. 9 In a simulation study using six large systematic review datasets on various topics, a Naïve Bayes (NB) and a term frequency-inverse document frequency (TF-IDF) outperformed other model settings. 10 The NB classifier estimates the probability of an article being relevant, based on TF-IDF measurements. TF-IDF measures the originality of a certain word within the article relative to the total number of articles the word appears in. 12 This combination of NB and TF-IDF was chosen for our systematic review.

Before the AI tool can be used for the screening of relevant articles, its algorithm needs training with at least one relevant and one irrelevant article (ie, prior knowledge). It is assumed that the more prior knowledge, the better the algorithm is trained at the start of the screening process, and the faster it will identify relevant articles. 9 In our review, the prior knowledge consisted of three relevant articles 13–15 selected from a systematic review on the topic 16 and three randomly picked irrelevant articles.

After training with the prior knowledge, the AI tool made a first ranking of all unlabelled articles (ie, articles not yet decided on eligibility) from highest to lowest probability of being relevant. The first reviewer read the title and abstract of the number one ranked article and made a decision (‘relevant’ or ‘irrelevant’) following the eligibility criteria. Next, the AI tool took into account this additional knowledge and made a new ranking. Again, the next top ranked article was proposed to the reviewer, who made a decision regarding eligibility. This process of AI making rankings and the reviewer making decisions, which is also called ‘researcher-in-the-loop’, was repeated until the predefined data-driven stopping criterion of – in our case – 100 subsequent irrelevant articles was reached. After the reviewer rejected what the AI tool puts forward as ‘most probably relevant’ a hundred times, it was assumed that there were no relevant articles left in the unseen part of the dataset.

The articles that were labelled relevant during the title and abstract screening were each screened on full text independently by two reviewers (SHBvD and MGJB-K, AL, JvdP, CJMD, CCB) to minimise the influence of subjectivity on inclusion. Disagreements regarding inclusion were solved by a third independent reviewer.

How to maintain reliability and transparency when using AI in title and abstract screening

A summary of the potential scenarios, and their pitfalls and remedies, when using the AI tool in a systematic review is given in table 1 . These potential scenarios should not be ignored, but acted on to maintain reliability and transparency. Figure 1 shows when and where to act on during the screening process reflected by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart, 17 from literature search results to publishing the review.

Flowchart showing when and where to act on when using ASReview in systematic reviewing. Adapted the PRISMA flowchart from Haddaway et al . 17

Per-scenario overview of potential pitfalls and how to prevent these when using ASReview in a systematic review

*What is considered manually feasible is highly context-dependent (ie, the intended workload and/or reviewers available).

In our systematic review, by means of broad literature searches in several scientific databases, a first set of potentially relevant articles was identified, yielding 8456 articles, enough to expect the AI tool to be efficient in the title and abstract screening (scenario ① was avoided, see table 1 ). Subsequently, this complete set of articles was uploaded in reference manager EndNote X9 18 and review manager Covidence, 19 where 3761 duplicate articles were removed. Given that EndNote has quite low sensitivity in identifying duplicates, additional deduplication in Covidence was considered beneficial. 20 Deduplication is usually applied in systematic reviewing, 20 but is increasingly important prior to the use of AI. Since multiple decisions regarding a duplicate article weigh more than one, this will disproportionately influence classification and possibly the results ( table 1 , scenario ② ). In our review, a deduplicated set of articles was uploaded in the AI tool. Prior to the actual AI-supported title and abstract screening, the reviewers (SHBvD and AL, MGJB-K) trained themselves with a small selection of 74 articles. The first reviewer became familiar with the ASReview software, and all three reviewers learnt how to apply the eligibility criteria, to minimise personal influence on the article selection ( table 1 , scenario ③ ).

Defining the stopping criterion used in the screening process is left to the reviewer. 9 An optimal stopping criterion in active learning is considered a perfectly balanced trade-off between a certain cost (in terms of time spent) of screening one more article versus the predictive performance (in terms of identifying a new relevant article) that could be increased by adding one more decision. 21 The optimal stopping criterion in systematic reviewing would be the moment that screening additional articles will not result in more relevant articles being identified. 22 Therefore, in our review, we predetermined a data-driven stopping criterion for the title and abstract screening as ‘100 consecutive irrelevant articles’ in order to prevent the screening from being stopped before or a long time after all relevant articles were identified ( table 1 , scenario ④ ).

Due to the fact that the stopping criterion was reached after 1063 of the 4695 articles, only a part of the total number of articles was seen. Therefore, this approach might be sensitive to possible mistakes when articles are screened by only one reviewer, influencing the algorithm, possibly resulting in an incomplete selection of articles ( table 1 , scenario ③ ). 23 As a remedy, second reviewers (AL, MGJB-K) checked 20% of the titles and abstracts seen by the first reviewer. This 20% had a comparable ratio regarding relevant versus irrelevant articles over all articles seen. The percentual agreement and Cohen’s Kappa (κ), a measure for the inter-reviewer agreement above chance, were calculated to express the reliability of the decisions taken. 24 The decisions were agreed in 96% and κ was 0.83. A κ equal of at least 0.6 is generally considered high, 24 and thus it was assumed that the algorithm was reliably trained by the first reviewer.

The reporting of the use of the AI tool should be transparent. If the choices made regarding the use of the AI tool are not entirely reported ( table 1 , scenario ⑤ ), the reader will not be able to properly assess the methodology of the review, and review results may even be graded as low-quality due to the lack of transparent reporting. The ASReview tool offers the possibility to extract a data file providing insight into all decisions made during the screening process, in contrast to various other ‘black box’ AI-reviewing tools. 9 This file will be published alongside our systematic review to provide full transparency of our AI-supported screening. This way, the screening with AI is reproducible (remedy to scenario ⑥ , table 1 ).

Results of AI-supported study selection in a systematic review

We experienced an efficient process of title and abstract screening in our systematic review. Whereas the screening was performed with a database of 4695 articles, the stopping criterion was reached after 1063 articles, so 23% were seen. Figure 2A shows the proportion of articles identified as being relevant at any point during the AI-supported screening process. It can be observed that the articles are indeed prioritised by the active learning algorithm: in the beginning, relatively many relevant articles were found, but this decreased as the stopping criterion (vertical red line) was approached. Figure 2B compares the screening progress when using the AI tool versus manual screening. The moment the stopping criterion was reached, approximately 32 records would have been found when the titles and abstract would have been screened manually, compared with 142 articles labelled relevant using the AI tool. After the inter-reviewer agreement check, 142 articles proceeded to the full text reviewing phase, of which 65 were excluded because these were no articles with an original research format, and three because the full text could not be retrieved. After full text reviewing of the remaining 74 articles, 18 articles from 13 individual studies were included in our review. After snowballing, one additional article from a study already included was added.

Relevant articles identified after a certain number of titles and abstracts were screened using the AI tool compared with manual screening.

In our systematic review, the AI tool considerably reduced the number of articles in the screening process. Since the AI tool is offered open source, many researchers may benefit from its time-saving potential in selecting articles. Choices in several scenarios regarding the use of AI, however, are still left open to the researcher, and need consideration to prevent pitfalls. These include the choice whether or not to use AI by weighing the costs versus the benefits, the importance of deduplication, double screening to check inter-reviewer agreement, a data-driven stopping criterion to optimally use the algorithm’s predictive performance and quality of reporting of the AI-related methodology chosen. This communication paper is, to our knowledge, the first elaborately explaining and discussing these choices regarding the application of this AI tool in an example systematic review.

The main advantage of using the AI tool is the amount of time saved. Indeed, in our study, only 23% of the total number of articles were screened before the predefined stopping criterion was met. Assuming that all relevant articles were found, the AI tool saved 77% of the time for title and abstract screening. However, time should be invested to become acquainted with the tool. Whether the expected screening time saved outweighs this time investment is context-dependent (eg, researcher’s digital skills, systematic reviewing skills, topic knowledge). An additional advantage is that research questions previously unanswerable due to the insurmountable number of articles to screen in a ‘classic’ (ie, manual) review, now actually are possible to answer. An example of the latter is a review screening over 60 000 articles, 25 which would probably never have been performed without AI supporting the article selection.

Since the introduction of the ASReview tool in 2021, it was applied in seven published reviews. 25–31 An important note to make is that only one 25 clearly reported AI-related choices in the methods and a complete and transparent flowchart reflecting the study selection process in the Results section. Two reviews reported a relatively small number (<400) of articles to screen, 26 27 of which more than 75% of the articles were screened before the stopping criterion was met, so the amount of time saved was limited. Also, three reviews reported many initial articles (>6000) 25 28 29 and one reported 892 articles, 31 of which only 5%–10% needed to be screened. So in these reviews, the AI tool saved an impressive amount of screening time. In our systematic review, 3% of the articles were labelled relevant during the title and abstract screening and eventually, <1% of all initial articles were included. These percentages are low, and are in line with the three above-mentioned reviews (1%–2% and 0%–1%, respectively). 25 28 29 Still, relevancy and inclusion rates are much lower when compared with ‘classic’ systematic reviews. A study evaluating the screening process in 25 ‘classic’ systematic reviews showed that approximately 18% was labelled relevant and 5% was actually included in the reviews. 32 This difference is probably due to more narrow literature searches in ‘classic’ reviews for feasibility purposes compared with AI-supported reviews, resulting in a higher proportion of included articles.

In this paper, we show how we applied the AI tool, but we did not evaluate it in terms of accuracy. This means that we have to deal with a certain degree of uncertainty. Despite the data-driven stopping criterion there is a chance that relevant articles were missed, as 77% was automatically excluded. Considering this might have been the case, first, this could be due to wrong decisions of the reviewer that would have undesirably influenced the training of the algorithm by which the articles were labelled as (ir)relevant and the order in which they were presented to the reviewer. Relevant articles could have therefore remained unseen if the stopping criterion was reached before they were presented to the reviewer. As a remedy, in our own systematic review, of the 20% of the articles screened by the first reviewer, relevancy was also assessed by another reviewer to assess inter-reviewer reliability, which was high. It should be noted, though, that ‘classic’ title and abstract screening is not necessarily better than using AI, as medical-scientific researchers tend to assess one out of nine abstracts wrongly. 32 Second, the AI tool may not have properly ranked highly relevant to irrelevant articles. However, given that simulations proved this AI tool’s accuracy before 9–11 this was not considered plausible. Since our study applied, but did not evaluate, the AI tool, we encourage future studies evaluating the performance of the tool across different scientific disciplines and contexts, since research suggests that the tool’s performance depends on the context, for example, the complexity of the research question. 33 This could not only enrich the knowledge about the AI tool, but also increases certainty about using it. Also, future studies should investigate the effects of choices made regarding the amount of prior knowledge that is provided to the tool, the number of articles defining the stopping criterion, and how duplicate screening is best performed, to guide future users of the tool.

Although various researcher-in-the-loop AI tools for title and abstract screening have been developed over the years, 9 23 34 they often do not develop into usable mature software, 34 which impedes AI to be permanently implemented in research practice. For medical-scientific research practice, it would therefore be helpful if large systematic review institutions, like Cochrane and PRISMA, would consider to ‘officially’ make AI part of systematic reviewing practice. When guidelines on the use of AI in systematic reviews are made available and widely recognised, AI-supported systematic reviews can be uniformly conducted and transparently reported. Only then we can really benefit from AI’s time-saving potential and reduce our research time waste.

Our experience with the AI tool during the title and abstract screening was positive as it has highly accelerated the literature selection process. However, users should consider applying appropriate remedies to scenarios that may form a threat to the methodological quality of the review. We provided an overview of these scenarios, their pitfalls and remedies. These encourage reliable use and transparent reporting of AI in systematic reviewing. To ensure the continuation of conducting systematic reviews in the future, and given their importance for medical guidelines and practice, we consider this tool as an important addition to the review process.

Supplementary Material

Contributors: SHBvD proposed the methodology and conducted the study selection. MGJB-K, CJMD and AL critically reflected on the methodology. MGJB-K and AL contributed substantially to the study selection. CCB, JvdP and CJMD contributed to the study selection. The manuscript was primarily prepared by SHBvD and critically revised by all authors. All authors read and approved the final manuscript.

Funding: The systematic review is conducted as part of the RE-SAMPLE project. RE-SAMPLE has received funding from the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 965315).

Competing interests: None declared.

Provenance and peer review: Not commissioned; externally peer reviewed.

Ethics approval

Not applicable.

Grab your spot at the free arXiv Accessibility Forum

Help | Advanced Search

Computer Science > Artificial Intelligence

Title: artificial intelligence for literature reviews: opportunities and challenges.

Abstract: This manuscript presents a comprehensive review of the use of Artificial Intelligence (AI) in Systematic Literature Reviews (SLRs). A SLR is a rigorous and organised methodology that assesses and integrates previous research on a given topic. Numerous tools have been developed to assist and partially automate the SLR process. The increasing role of AI in this field shows great potential in providing more effective support for researchers, moving towards the semi-automatic creation of literature reviews. Our study focuses on how AI techniques are applied in the semi-automation of SLRs, specifically in the screening and extraction phases. We examine 21 leading SLR tools using a framework that combines 23 traditional features with 11 AI features. We also analyse 11 recent tools that leverage large language models for searching the literature and assisting academic writing. Finally, the paper discusses current trends in the field, outlines key research challenges, and suggests directions for future research.

Submission history

Access paper:.

• HTML (experimental)
• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 14 August 2024

Ecological footprints, carbon emissions, and energy transitions: the impact of artificial intelligence (AI)

• Qiang Wang   ORCID: orcid.org/0000-0002-8751-8093 1 , 2 ,
• Yuanfan Li 1 &
• Rongrong Li 2  

Humanities and Social Sciences Communications volume  11 , Article number:  1043 ( 2024 ) Cite this article

465 Accesses

6 Altmetric

Metrics details

• Development studies
• Environmental studies
• Science, technology and society

This study examines the multifaceted impact of artificial intelligence (AI) on environmental sustainability, specifically targeting ecological footprints, carbon emissions, and energy transitions. Utilizing panel data from 67 countries, we employ System Generalized Method of Moments (SYS-GMM) and Dynamic Panel Threshold Models (DPTM) to analyze the complex interactions between AI development and key environmental metrics. The estimated coefficients of the benchmark model show that AI significantly reduces ecological footprints and carbon emissions while promoting energy transitions, with the most substantial impact observed in energy transitions, followed by ecological footprint reduction and carbon emissions reduction. Nonlinear analysis indicates several key insights: (i) a higher proportion of the industrial sector diminishes the inhibitory effect of AI on ecological footprints and carbon emissions but enhances its positive impact on energy transitions; (ii) increased trade openness significantly amplifies AI’s ability to reduce carbon emissions and promote energy transitions; (iii) the environmental benefits of AI are more pronounced at higher levels of AI development, enhancing its ability to reduce ecological footprints and carbon emissions and promote energy transitions; (iv) as the energy transition process deepens, AI’s effectiveness in reducing ecological footprints and carbon emissions increases, while its role in promoting further energy transitions decreases. This study enriches the existing literature by providing a nuanced understanding of AI’s environmental impact and offers a robust scientific foundation for global policymakers to develop sustainable AI management frameworks.

Similar content being viewed by others

Extension of Japan’s Prefectural Emission Accounting and Enrichment of Socioeconomic Data from 1990 to 2020

Rethinking the environmental Kuznets curve hypothesis across 214 countries: the impacts of 12 economic, institutional, technological, resource, and social factors

Navigating the nexus: unraveling technological innovation, economic growth, trade openness, ICT, and CO2 emissions through symmetric and asymmetric analysis

Introduction.

The world is now facing a series of environmental degradations, from water and air pollution to the depletion of ecological resources and many of them are confirmed to be caused by human activities. According to the Global Footprint Network (GFN), the ecological pressure from human activities has exceeded the Earth’s carrying capacity by 75% (Network 2019 ). The significant impact of humans on the environment, in turn, affects human development, including economic activities, health conditions, and energy availability. Consequently, both developed and developing countries worldwide have taken various measures, such as conserving natural resources, limiting carbon emissions, and accelerating energy transitions, to mitigate the increasingly urgent climate and environmental issues. For instance, nearly all countries have signed the Paris Climate Change Agreement (PCCA) and committed to reducing their carbon emission levels. Additionally, many countries have announced their energy transition strategies, promoting a shift from fossil fuel dependence to renewable energy reliance to balance energy demand and reduce environmental impacts (Dong et al. 2022a ). Therefore, besides economic growth, protecting ecological resources, reducing carbon emissions, and increasing the utilization of renewable energy have become important development goals for governments worldwide in recent years.

The advent of artificial intelligence (AI) has brought multiple impacts on people’s lives and production methods, with its applications increasingly extending into manufacturing, energy, healthcare, logistics, finance, and other fields (Zavyalova et al. 2023 ). According to a recent study published by Forbes, the global AI market is projected to reach a staggering $71 billion by 2024. As a game-changing technology, AI’s rapid development presents unprecedented opportunities and challenges for global environmental protection efforts (Ahmad et al. 2022 ). On one hand, AI’s powerful computing capabilities are believed to help increase productivity and achieve economies of scale, thereby conserving resources and energy for the whole society (Wang et al. 2023c ). On the other hand, an increasing number of practitioners and scholars point out that the explosive growth in AI capabilities is accompanied by an exponential increase in the energy consumption required for AI training. For example, research indicates that training a single model like ChatGPT consumes 1.287 gigawatt-hours of electricity, roughly equivalent to the annual electricity consumption of 120 American households (Probst 2023 ). In short, AI is impacting the current energy system in different directions through multiple channels.

In response to the opportunities and threats brought by the rapid development of AI, more and more countries are turning their attention to regulating, guiding, and utilizing AI to protect the environment, reduce carbon emissions, or promote energy transition, aiming to align AI development with sustainable development standards. Table 1 summarizes some national strategies that emphasize harnessing AI’s potential for sustainability. However, the actual impact of AI on environmental actions remains unclear. Therefore, this study raises the question: What effects does AI development have on the environmental indicators, especially the ecological environment, carbon emissions, and energy transition? How do external conditions such as industrial structure and trade openness influence these effects? Answering these questions is crucial to ensuring AI stays on a sustainable development track and further optimizing AI policies.

From the existing literature, the factors influencing environmental degradation and energy transition have been extensively studied, providing a solid theoretical foundation for subsequent research. However, the emerging variable of AI has rarely been integrated into traditional research frameworks. Additionally, some studies have explored the potential opportunities and threats of AI concerning the ecological environment, carbon emissions, and energy transition from various perspectives but lack systematic and quantitative evaluation. To fill in this gap, this paper aims to quantitatively assess the impact of AI on the ecological environment, carbon emissions, and energy transition and explore the possible heterogeneity and nonlinearity of these impacts. First, based on panel data from 67 countries between 1993 and 2019, a set of SYS-GMM models is employed to investigate the effects of AI on ecological footprint, carbon emissions, and energy transition. Second, the potential differences in AI’s impact across developed and middle-income countries are discussed by grouping countries according to their development stages. Finally, a series of threshold models are constructed using industrial structure, trade openness, AI development level, and energy transition level as threshold variables to clarify the nonlinear characteristics of these impacts. This article aims to clarify the multifaceted environmental impacts of AI and explore its potential heterogeneity and nonlinearity, providing a scientific foundation for global policymakers. Amid the current complex and controversial relationship between AI and the environment, this research helps policymakers develop and manage AI sustainably.

The remainder of the paper is organized as follows. Section 2 reviews the literature. Section 3 briefly introduces the methodology and data. Section 4 discusses the empirical results. Conclusions and policy implications are presented in Section 5.

Literature review

Although the core variables of this paper include carbon emissions, ecological footprint, energy transition, and AI, they come from a classic and broad topic, namely the technology-environment nexus. Therefore, we first review previous discussions on the impact of technological factors on important environmental indicators. Subsequently, we focus on the qualitative discussion and empirical research on the specific environmental impact of AI. Finally, we discussed the gap in existing literature.

Literature on the nexus between technological factors and environmental indicators

Ecological footprint, carbon emissions, and energy transition are three variables that have been widely used to assess sustainability in previous research: Ecological footprint refers to the biologically productive areas that can provide the resources needed for human survival or consume the waste discharged by human beings,.and can be used to measure the pressure of human activities on the ecological environment; reducing carbon emissions is the most direct way to slow down climate change; energy transition is the key path to achieve environmental protection and climate mitigation, marking the recognition of human beings on renewable energy, a clean resource. Examining the determinants of each of these sustainability indicators is a hot topic in the empirical literature on energy economics (Dong et al. 2022b ; Ehigiamusoe and Dogan 2022 ; Wang et al. 2024b . Among the various factors that may affect the above environmental indicators, technological factors are considered to be crucial. Technology has a profound impact on many aspects of residents’ social life and economic production, and these impacts will ultimately change a country’s energy use and interaction with nature (Wang et al. 2024a ). In other words, technological factors may have a positive or negative impact on carbon emissions, ecological footprint, and energy transition.

For example, in terms of the nexus of technology and carbon emissions, Cheng et al. ( 2018 ) used the DEA-Malmquist method to measure the total factor productivity of 30 provinces in China from 1998 to 2014 as an indicator of technological progress. The findings show that technological advances reduce carbon intensity by increasing production efficiency. Gu et al. ( 2019 ) used the number of energy technology patents to represent technological progress, and found an inverted U-shaped relationship between energy technology progress and carbon emissions. Technologies that affect carbon emissions include not only technological progress or innovation in a broad sense but also specific technologies such as digital technology and ICT. Moyer and Hughes ( 2012 ) used the International Futures (IFs) comprehensive assessment system to study the relationship between ICT and global carbon emissions, and believed that ICT has brought about a decline in carbon emissions in the past 50 years. Yi et al. ( 2022 ) used the spatial panel Durbin model to study the impact of the digital economy on carbon emission reduction in 30 provinces in China. They emphasized that the development of the digital economy not only has a direct emission reduction effect but also has a significant spatial spillover effect. The ecological footprint was first brought into the field of empirical research by (Wackernagel and Rees 1998 ) as another indicator of environmental sustainability, which is considered to be a more comprehensive measure of environmental degradation than carbon emissions. Therefore, many scholars use the ecological footprint to replace carbon emissions, and re-examine the already tested nexus between carbon emissions and technology. Huang et al. ( 2022 ) compiled the ICT index through principal component analysis, and found that the development of ICT significantly increased the ecological footprint of E-7 countries, but decreased that of G-7 countries.

Regarding the impact of technology on energy transition, Zheng et al. ( 2021 ) based on China’s provincial data from 2005 to 2017, examine the impact of renewable energy technology innovation represented by patent stock on renewable energy power generation. The results show that an average increase of 1% in the level of renewable energy technology innovation directly brings about a 0.411% increase in renewable energy power generation in the province and a 3.264% increase in renewable energy power generation in neighboring provinces through technology diffusion. Tzeremes et al. ( 2023 ) and Shahbaz et al. ( 2022 ) respectively confirm the role of ICT and digital techniques in promoting energy transition. In summary, based on the above review, the use of carbon emissions, ecological footprint, and energy transition variables can characterize the different impacts of technological factors. In addition, different technologies may have different impacts on the same indicator. Therefore, in order to fully understand the environmental impact of the current new technology of AI, it is necessary to select sufficiently detailed and accurate methods to distinguish AI from previous digital technologies and examine the impact of AI on each different environmental indicator.

The environmental impact of artificial intelligence

The debate over the environmental impact of artificial intelligence (AI) can be traced back to a classic theory related to technological progress—the rebound effect. The rebound effect refers to the phenomenon where improvements in efficiency do not always reduce demand; instead, they can lead to an increase in demand. For example, when a company improves efficiency and achieves raw material savings per unit of product, it may further expand production. This expansion in production resulting from efficiency improvements can completely offset the raw material savings brought about by technological upgrades, or even lead to a dramatic increase in the company’s raw material demand. This rebound effect suggests that although technology may enhance efficiency and save resources per unit product, the ultimate resource and energy consumption, as well as its environmental impact, are uncertain.

In this context, Hilty and Hercheui ( 2010 ) proposed the three-order effect theory to summarize the multiple aspects of the environmental impact that smart technologies may have. The first-order effects of smart technologies refer to the direct environmental impacts of specific products throughout their lifecycle, including production, usage, and disposal. This encompasses the environmental footprint of acquiring raw materials for production, the electricity consumption during usage, and the environmental pollution associated with disposal. Second-order effects represent the direct environmental benefits brought about by efficiency improvements and process optimizations enabled by smart technologies. For example, the use of electronic products directly replaces some paper materials, reducing deforestation. Third-order effects represent the negative environmental impacts arising from the rebound effect based on the second-order effects. For instance, the prevalence of electronic products might lead to long-term dependence or addiction to these devices, resulting in electricity consumption (primarily produced by fossil fuels) that far exceeds the environmental impact of paper-based materials like newspapers and books.

Understanding these effects is crucial for comprehensively grasping the multifaceted debate on the environmental impact of AI. On one hand, the second-order effects of AI are widely acknowledged. In other words, AI has enormous potential to improve production efficiency, enhance production processes, or support environmental protection and energy transition initiatives, thereby promoting environmental improvement (Lei et al. 2023 ; Lyu and Liu 2021 ). On the other hand, the energy consumption throughout its lifecycle (first-order effect) and the productivity rebound (third-order effect) may exacerbate environmental issues (Taddeo et al. 2021 ; Wu et al. 2022 ).

Although it is currently challenging to quantify each specific impact of AI, utilizing detailed macroeconomic statistical data for holistic assessment is a critical research direction. Many scholars leverage global or national-level time series or panel data, selecting environmental indicators such as carbon emissions, ecological footprints, and energy consumption to evaluate the overall environmental impact of AI and understand current trends in AI development. On the one hand, using a global panel and dynamic estimation approach, Ding et al. ( 2023 ) claim that AI reduces carbon emissions significantly. On the other hand, Luan et al. ( 2022 ) found that AI contributed to air pollution through a study of 74 countries and insisted that AI has exacerbated climate warming. There is a little article that studies ecological footprint (Chen et al., 2022b ) which confirms the positive effect of AI on ecological footprint. And only (Wang et al. 2024 c) discussed the impact of AI on energy transition and claimed a positive effect. Other related research about the environmental impact of AI is summarized in Table 2 .

Research gaps and contributions

In summary, the aforementioned studies proxy environmental sustainability carbon emissions, ecological footprint, and energy transition to explore and verify the environmental impact of broad-range factors of economics, energy, and technology. However, few studies are focusing on the important AI-environment sustainability nexus. Specifically, three detailed research gaps are identified. (1) From a research perspective, existing studies discussing the environmental sustainability of AI often focus on the perspective of carbon emissions. The impact of AI on ecological footprint and energy transition has rarely been studied, and there are almost no studies combining multiple indicators. (2) Regarding the research object, China has been the most fully studied, while research on other countries or global levels has received inadequate attention. (3) In terms of research ideas and methods, most studies presuppose linear relationships between AI and corresponding environment indicators and adopt linear research methods, thus ignoring the potential nonlinearity and heterogeneity relationships.

This paper fills the above research gaps by using a global panel to evaluate and compare the impact of AI on carbon emissions, ecological footprint, and energy transition. Moreover, dynamic threshold models are developed to explore the potential nonlinear impact under different conditions or stages of industrial structure, trade openness, AI development, and energy transition. Against the backdrop of the current complex and controversial relationship between AI and the environment, this article helps to clarify the multifaceted environmental impacts of AI and explore its potential heterogeneity and nonlinearity, providing a scientific basis for global policymakers to develop and manage AI.

Methods, variables, and data

Model setting.

This study aims to evaluate the linear and nonlinear impact of AI on ecological footprint, carbon emissions, and energy transition respectively. For better modeling of these explained variables, dynamics can not be ignored. This is because in many cases, how much environmental impact a country will have or what kind of energy structure it adopts will be determined partially by its past behavior. For example, once a certain industrial production mode is determined, a lot of equipment and professionals need to be invested, and it is difficult for enterprises to switch to another production mode in a short time. Therefore, there may be inertia in carbon emissions, ecological footprint, and energy transition. Based on the above discussion, we first construct three dynamic linear models as the benchmark regression models in this paper to measure the effects of AI on ecological footprint, carbon emissions, and energy transition:

where i represents individual and t represents time. EF, CE, and ETR are the explained variables, representing ecological footprint, carbon emission, and energy transition respectively. AI is the explanatory variable, representing the level of artificial intelligence development. C represents the control variables vector, which is kept the same in the three models. \({u}_{i}\) is the individual fixed effect, and \({\varepsilon }_{{it}}\) is the disturbance term.

Based on Model L1-L3, we need to further construct a series of nonlinear models to study the influence of external variables. In the literature that uses traditional methods to discuss nonlinearity, some add quadratic terms of independent variables to the equation, and some choose to divide the research objects into different samples according to certain standards (such as income, GDP, etc.). These methods can partially solve the problem of nonlinearity, but the rationality and effectiveness of artificially setting models and dividing samples have been questioned. To address the above issues, in this paper, we adopt the dynamic panel threshold models to model the nonlinearity. In a threshold model, the division of the sample interval is endogenously determined by the sample data, and the regression coefficient is allowed to be different across intervals, which overcomes the problem of artificial selection of the model. Therefore, the following dynamic threshold models are used in this paper:

where \({{\boldsymbol{C}}}_{{\boldsymbol{it}}}\) is a vector of control variables consistent with Model L1-L3. \({q}_{{it}}\) is the threshold variable and \(\gamma\) is the corresponding threshold value. I(·) is an indicator function that takes 1 when the condition is satisfied and 0 otherwise. We specially focus on \({\delta }_{1}\) because it collects the change in the influence coefficient of AI when the threshold variable is in a higher interval.

Estimation strategy

System generalized moment estimation.

The System Generalized Method of Moments (SYS-GMM) proposed by Blundell and Bond ( 1998 ) is adopted to estimate coefficients in Model L1-L3. This method uses GMM instead of the traditional OLS estimation method to solve the problem of downward bias (Nickell bias) of the estimator caused by the introduction of the dependent variable lag term in the traditional fixed effect model (Pesaran 2020 ). In addition, SYS-GMM makes full use of the residual information in the level equation and the difference equation to construct moment conditions, which improves the efficiency of parameter estimation. In order to explain the estimation strategy of SYS-GMM, we express Model L1-L3 in vector form as follows:

where \({Y_{it}}\) and \({Y_{{it}-1}}\) are the explained variable and its first-order lag, respectively. \({\boldsymbol{X}}\) is a vector of explanatory variables. ∆ represents the first difference. ε i and \({\boldsymbol{\Delta }}{{\boldsymbol{\varepsilon }}}_{{\boldsymbol{i}}}\) denote the residual series of the level and difference equations, respectively.

In order to estimate the parameters, we need to obtain all sample moment conditions for the level and difference equations. First, for the above level equation (original equation), we have the moment condition related to the level equation:

Secondly, for the first-order difference equation of the level equation:

we also have the moment condition for the difference equation:

where \({{\boldsymbol{Z}}}_{{\boldsymbol{li}}}^{{\boldsymbol{{\prime} }}}\) and \({{\boldsymbol{Z}}}_{{\boldsymbol{di}}}^{{\boldsymbol{{\prime} }}}\) are the transposes of the level and difference instrumental variable matrix, and

Finally, estimates of the unknown parameters \({{\boldsymbol{\theta }}}_{{\boldsymbol{i}}}\)  = ( α , β 0 , β ) will be produced by solving the above sample moment conditions.

To ensure the validity of instrumental variables, two tests should be employed before analyzing the estimation result. First, the Sargan test should be adopted to prevent over-identification. Because the number of instrumental variables is a quadratic function of T in the SYS-GMM model. As T increases, the number of instrumental variables may exceed the parameters to be estimated, which is called over-identified. Second, the Arellano-Bond (AR) test needs to be employed to prevent residual serial correlation. Because SYS-GMM employs the instrumental strategy of using lagged variables, which relies on the presupposes that the residual series are uncorrelated.

DPTM with endogenous threshold variables

The dynamic panel threshold model (DPTM) newly proposed by Seo et al. ( 2019 ) with an endogenous threshold variable is adopted to estimate the nonlinear models (Model N1-N3). The key difference between DPTM and the traditional static panel threshold model is that the equation includes the lagged term of the explained variable, and relaxes the assumption that the threshold variable is completely exogenous, allowing it to become endogenous. In order to explain how DPTM solves the problem of lagged terms of explained variables and endogenous threshold variables, we express the DPTM model in the following vector form:

where \({{\boldsymbol{X}}}_{{\boldsymbol{it}}}\) is the set of explanatory variables including \({{\boldsymbol{Y}}_{{\boldsymbol{it}}}{\boldsymbol{-}}{\boldsymbol{1}}}\) , and \({\boldsymbol{(}}{1{\boldsymbol{,}}{\boldsymbol{X}}}_{{\boldsymbol{it}}}{\boldsymbol{)}}\) additionally includes a constant term. \({\boldsymbol{\delta }}{\boldsymbol{(}}{{\boldsymbol{1}}{\boldsymbol{,}}{\boldsymbol{X}}}_{{\boldsymbol{it}}}{\boldsymbol{)}}I({q}_{{it}}\ge \gamma )\) records the change in the coefficient of all the explanatory variables when the threshold variable changes from low to high cross-country threshold. According to Seo et al. ( 2019 ), the difference generalized method of moments (Diff-GMM) is used to eliminate \({u}_{i}\) and estimate unknown parameters θ  = ( β , δ , γ ). The difference form of Eq. ( 7 ) is:

Consider the following sample moment conditions:

\({{\boldsymbol{z}}}_{{\boldsymbol{it}}}\) is a set of instrumental variables consisting of lagged items, but excluding threshold variables. Footnote 1 Assume that \({\rm{E}}\left({g}_{i}\left({\boldsymbol{\theta }}\right)\right)=0\) , if and only if \({\boldsymbol{\theta }}={{\boldsymbol{\theta }}}_{{\boldsymbol{0}}}\) . Then further let \({g}_{i}={g}_{i}\left({{\boldsymbol{\theta }}}_{{\boldsymbol{0}}}\right)=\) \({\left({z}_{i{t}_{0}}^{{\prime} }\varDelta {\varepsilon }_{i{t}_{0}},\ldots ,{z}_{{iT}}^{{\prime} }\varDelta {\varepsilon }_{{iT}}\right)}^{{\prime} }\) , and let \(\varOmega ={\rm{E}}\left({g}_{i}{g}_{i}^{{\prime} }\right)\) . It can be seen that \(\varOmega\) is a positive-definite matrix. Assume \({W}_{n}\) is a positive definite weight matrix that satisfies \({W_{n}}\mathop{\rightarrow}\limits^{p}{\varOmega }^{-1}\) . Let the criterion function \({\bar{J}}_{n}({\boldsymbol{\theta }})={\bar{g}}_{n}({\boldsymbol{\theta }})^{\prime} {W_{n}}{\bar{g}}_{n}({\boldsymbol{\theta }})\) , The GMM estimator for unknown parameter \(\hat{{\boldsymbol{\theta }}}\)  =  \(\left(\hat{{\boldsymbol{\beta }}},\hat{{\boldsymbol{\delta }}},\hat{\gamma }\right)\) is given as follows:

The above minimization problem can be solved in two steps. First, let the weight matrix W n be specified as an identity matrix I or other known form, and minimize \({\bar{J}}_{n}({\boldsymbol{\theta }})\) to collect residuals \({\widehat{\varDelta \varepsilon }}_{{it}}\) . In the second step, the weight matrix is updated using the collected residuals \({W_{n}}={\left(\frac{1}{n}{\sum}_{i=1}^{n}{\hat{g}}_{i}{\hat{g}}_{i}^{{\prime} }-\frac{1}{{n}^{2}}\mathop{\sum }\nolimits_{i=1}^{n}{\hat{g}}_{i}{\sum}_{i=1}^{n}{\hat{g}}_{i}^{\prime}\right)}^{-1}\) , where \({\hat{g}}_{i}={\left({\widehat{\varDelta \varepsilon}}_{i{t}_{0}}{z}_{i{t}_{0}}^{{\prime} },\ldots ,{\widehat{\varDelta \varepsilon }}_{{iT}}{z}_{{iT}}^{{\prime} }\right)}^{{\prime} }\) , and minimize \({\bar{J}}_{n}({\boldsymbol{\theta }})\) again to solve \(\hat{{\boldsymbol{\theta }}}\) . It has been proved that the parameters estimated by the above method \(\hat{{\boldsymbol{\theta }}}=(\hat{{\boldsymbol{\beta }}},\hat{{\boldsymbol{\delta }}},\hat{\gamma })\) asymptotically obey the normal distribution, so \(t\) -statistics can be used to test the significance of parameter estimates and construct confidence intervals.

In order to test the significance of the threshold effect to see the rationality of establishing a nonlinear model, the following null hypothesis is proposed \(H0:{\boldsymbol{\delta }}=0\) , for any \(\gamma\) . The corresponding alternative hypothesis is \(H1:{\boldsymbol{\delta }}\,\ne \,0\) , for some \(\gamma\) . This null hypothesis means that the threshold effect does not exist and the model is linear. The \(\sup {Wald}\) statistic is used to test the hypothesis. A significant \(\sup {Wald}\) statistic means that the null hypothesis is rejected, that is, the threshold effect is considered to exist. The probability distributions of \(\sup {Wald}\) statistic and the p-value for rejecting the null hypothesis can be obtained by the bootstrap method according to Seo et al. ( 2019 ).

Variable definition

Explained variables: ef, ce and etr.

Ecological Footprint Consumption (EF) and Carbon Emissions (CE): Ecological Footprint Consumption (EF) can reflect the pressure of human activities on the ecological environment (Li et al., 2023 ). A country’s EF figure is calculated from the per capita amount of built-up land, carbon, arable land, fishing grounds, forest products, and pasture combined. CE is measured by annual per capita greenhouse gas emissions.

Energy transition (ETR): As stressed by Markard ( 2018 ), the modern energy transition is characterized by the gradual reduction of fossil fuels and the continuous shift to low-carbon energy such as wind energy and solar energy (Li et al., 2022a ). Thus it can be considered that the purpose of energy transition is to replace fossil energy with renewable energy in the final energy consumption structure. Following Tzeremes et al. ( 2023 ), we take the share of renewable energy in a country’s total final energy consumption Footnote 2 to represent the energy transition result.

Core explanatory variable: artificial intelligence development level (AI)

Artificial intelligence technology has experienced rapid development and multiple updates in the past forty years. With the continuous iteration of technology, the connotation of the concept of AI has also changed. Drawing on He et al. ( 2019 ), this paper defines AI as a branch of applied computer science that uses computer algorithms, including machine learning, deep learning, and natural language recognition, to be trained to perform tasks usually associated with human intelligence.

According to this definition, AI includes a variety of technologies, and the application scenarios are also rich. Therefore, how to measure the development level of AI is a major challenge of this paper. Considering that AI devices, such as robots and smart cars, often integrate hardware, software, and programming, and may integrate multiple technologies (Howard 2019 ). It can be inferred that the more advanced a country’s AI technology is, the more functions and application scenarios AI devices have, and the more they may be used. This insight led us to consider assessing AI development from the perspective of the number of devices. Inspired by the work of Acemoglu and Restrepo ( 2020 ), in this paper, we select the well-defined and easy-to-count operational stock of industrial robots as a proxy variable for a country’s AI development level and select per capita stock of industrial robots as an alternative explanatory variable for robustness test.

According to IFR, the global inventory of industrial robots in 2019 was about 2.3 million. The top five countries account for more than 75% of the total, namely China (27.5%), Japan (15.3%), South Korea (12.7%), the United States (11.5%) and Germany (8.4%), suggesting that the above-mentioned countries should have an advantage in AI development. This is supported by many reports assessing the level of AI development. For example, the 2022 Global Artificial Intelligence Innovation Index provided by the Institute of Scientific and Technological Information of China shows that the United States and China are in the first echelon of global AI innovation, while Germany, Japan, and South Korea have also entered the second echelon (ISTIC 2023 ). Therefore, we believe that the operational stock of industrial robots can well represent AI development.

Threshold variable

Industrial structure (IS), measured by the proportion of industrial added value in GDP. Various pieces of research pointed out that there are long-term and stable differences in the relationship between carbon emissions and economic development in different industries and sectors (Dong et al. 2020 ; Li et al. 2024 ). We can infer that the development and application of AI in a country will be affected by the existing industrial structure.

Trade openness (OPEN): Measured by the proportion of total trade in GDP. Trade openness will promote the international division of labor and the exchange of goods and technologies, which should be conducive to the development of AI. Therefore, AI development conditions in countries with various degrees of openness may be different.

Artificial intelligence development (AI): Recent studies have shown that AI technology is changing even faster than Moore’s Law Footnote 3 describes Taddeo et al. ( 2021 ). Thus, it is necessary to use the key independent variable AI as a threshold variable to reveal the dynamic change rules of AI’s effect, in other words, the heterogeneous environmental impact at different development levels of AI.

Energy transition (ETR): Inspired by Shahbaz et al. ( 2022 ) who found that the impact of the digital economy on ETR varies in different quantiles of ETR, we therefore use a threshold model and take ETR as a threshold to explore how the environmental effects of AI change as countries progress in energy transition.

Control variables

Based on the work of Huang et al. ( 2022 ) for EF, Li and Wang ( 2022 ) for CE, and Shahbaz et al. ( 2022 ) for ETR, the following vectors of control variables are constructed and introduced into Model L1-L3 and Model N1-N3 to model EF, CE and ETR. (1) Economic development level (ED), measured by per capita GDP; (2) Energy use (EUSE), measured by per capita energy consumption; (3) Natural resource rent (NRR), measured by the share of total natural resource rent in Footnote 4 GDP (4) Fossil fuel power generation (FFE): measured by the proportion of fossil fuels in power generation; (5) Government size (GOV): measured by the ratio of fiscal expenditure to GDP; (6) Urbanization (URB): Measured by the proportion of urban population.

Data description and preliminary test

This study covers 67 countries from 1993 to 2019. The names of countries in the research scope are shown in Appendix Table 1 . The data of the variables were sourced from GFN (Network 2019 ), WDI (WDI 2021 ), and IFR (IFR 2023 ). The descriptive statistics and data sources of variables are shown in Table 3 , and all variables are logarithmic to alleviate the problem of heteroscedasticity. The covariance matrix and the variance inflation factors of the explanatory variables are shown in Appendix Table 2 . It can be found that the largest positive correlation coefficient comes from ED and CE (0.734), and the largest negative correlation coefficient comes from ED and NRR (−0.430). The largest VIF is 3.39 of ED, indicating no obvious multicollinearity in our research case.

The results of cross-section correlation tests for the three benchmark models are reported in Appendix Table A3 . The results showed that all three models rejected the null hypothesis of no cross-section correlation at the 5% significance level. This suggests that cross-sectional correlations exist in Model L1-L3. Based on the result, second-generation unit root testing techniques such as CIPS and CADF should be adopted to overcome the issue of cross-sectional correlation (Ansari 2022 ). Appendix Table A4 shows the test results of CIPS and CADF, and it can be seen that all variables are first-order series stationary. Finally, the cointegration test proposed by Westerlund ( 2005 ) was applied to the three benchmark models to avoid spurious regressions. The results of Table A5 in the Appendix show that all three models have significantly rejected the null hypothesis of “No cointegration for all panels” at the 5% level, indicating that there is a long-term stationary relationship between the variable sequences of each model. In conclusion, the above results show that the preliminary tests are passed and the regression result is reliable.

Empirical results

Benchmark result.

In order to evaluate the impact of AI on sustainability from the three perspectives of ecological footprint, carbon emissions and energy transition, we estimate Model L1–L3 based on the SYS-GMM method of Eqs. ( 1 – 6 ). The specific results are shown in Table 4 . First, we judge the reasonableness of the model based on the results of the Sargan test and the AR test in the last three lines. The p -values of the Sargan test are 0.9945, 0.9943 and 0.9977, which are all greater than 0.1. Therefore, the null hypothesis that there is no over-identification of instrumental variables cannot be rejected, indicating that the Sargan test is passed. The p -values of the AR(2) tests in the three models are also greater than 0.1, indicating that there is no serial correlation. The above test results show that the SYS-GMM model is well set. The first two rows of the table represent the estimated coefficients for the lagged term of the dependent variable. We found that the coefficients of the lagged terms of EF, CE, and ETR are 0.8904, 0.8084, and 0.9245, respectively, which are significant at the 1% level. This means that the changes in ecological footprint, carbon emissions and energy transition have inertia and are largely affected by the previous year, which confirms the rationality of the dynamic model setting.

As for the result of the key variable, the coefficient of lnAI on lnEF is −0.0018%, which is significant at the 1% level. This means that for the 67 sample countries from 1993 to 2019, every 1% increase in the development level of artificial intelligence corresponds to an average decrease of 0.0018% in ecological footprint. This result suggests that AI contributes to sustainable development by reducing the global ecological footprint. Similar to our point of view, Holzinger et al. ( 2023 ) believe that AI technology can improve the sustainability of agriculture, forestry, and animal husbandry, especially for the preservation of biological resources. Compared with other empirical literature, Ahmad et al. ( 2020 ) and Huang et al. ( 2022 ) found that technological progress and ICT have an ecological footprint reduction effect, respectively. On this basis, our empirical results further expand their conclusions and confirm that AI also has an ecological footprint reduction effect.

According to Model L2, AI also has a significant reduction effect on CE. For every 1% increase in a country’s artificial intelligence development level, the country’s carbon emissions will decrease by 0.0013%. This means that the development of AI can help the country reduce its overall carbon emissions, thereby promoting sustainable development, which is consistent with the empirical findings of Liu et al. ( 2022b ). Note that this influence coefficient is smaller than that of AI on EF (0.0013 < 0.0016). This may be because the positive impact of AI on carbon emissions is partially offset by the negative impact of AI itself such as power consumption (Taddeo et al., 2021 ).

Model L3 shows that at the 1% significance level, AI shows a promoting effect on energy transition. Every 1% increase in the development level of artificial intelligence will boost the energy transition by 0.0025%. The coefficient is the largest among the three models (0.0025 > 0.0018 > 0.0013), indicating that AI is an important driving force for the global energy transition process, which is of great significance for sustainable development. Some existing empirical literature supports our findings. For example, Lee and He ( 2021 ) confirmed that AI can promote wind power technology innovation, and Lyu and Liu ( 2021 ) proved that AI is conducive to the development of digital energy technology. In fact, AI and energy transition are already closely integrated and AI methods such as back propagation neural networks (BPNN) have been widely used in the field of renewable energy (Jha et al. 2017 ).

Combined with the analysis of Model L1-L3 results, we conclude that AI has three positive contributions to sustainability, including ecological footprint reduction effects, carbon reduction effects, and energy transition promotion effects. Therefore, when the sustainable development of countries is hindered after the pandemic, the AI industry can become the focus of countries to accelerate economic development while improving environmental sustainability especially in promoting energy transition.

Heterogeneity analysis

Given that this study involves a large number of countries at different stages of development, it is necessary to conduct a grouped analysis. Based on the World Bank’s 2018 classification of national development levels, we divided the sample into high-income and middle-income groups Footnote 5 . The results of the grouped study are shown in Table 5 , where columns 1–3 represent the results for the high-income group and columns 4–6 for the middle-income group.

The results, with the ecological footprint as the dependent variable, indicate that the impact coefficient of AI is −0.0016 in high-income countries and −0.0028 in middle-income countries. Both coefficients are significant at the 1% level. This demonstrates that the effect of AI on reducing the ecological footprint is more pronounced in middle-income countries compared to high-income countries. Similarly, the research results on carbon emissions show that AI also significantly reduces carbon emissions in middle-income countries. These results may be because middle-income countries have more room for improvement in clean production, so the widespread application of AI technology can achieve greater reductions in carbon emissions and other ecological footprints in these countries. This contrasts with high-income countries, which have already optimized their production and resource use efficiency to a higher level, resulting in relatively smaller marginal improvements from AI.

However, the results regarding energy transition indicate that the effect of AI on energy transition is mainly concentrated in high-income countries, with an impact coefficient of 0.0053, while it is not significant in middle-income countries. This may suggest that high-income countries have more advanced infrastructure, higher technical capabilities, and more comprehensive policy support, enabling them to utilize AI technology more effectively to promote energy transition. For instance, high-income countries may have already established smart grids, renewable energy management systems, and advanced energy storage technologies, which provide a solid foundation for the application of AI in energy transition. On the other hand, middle-income countries have relatively less investment and infrastructure development in these areas, resulting in AI having a less significant role in promoting energy transition compared to high-income countries.

In summary, the grouped study results indicate that the reduction effects of AI on ecological footprint and carbon emissions are more significant in middle-income countries, whereas the effects on energy transition are mainly observed in high-income countries. This difference suggests that the developmental stage of a country may influence the environmental effects of AI, highlighting the need for countries to formulate corresponding strategies and policies based on their own developmental realities to maximize the positive effects of AI technology.

Threshold effect tests and sample distributions

Based on the benchmark results, we further use the dynamic threshold model to conduct two-pronged research on the nonlinear laws of the three environmental sustainability effects of AI: (1) Use the external variables IND and OPEN as threshold variables to discuss the impact of external macro factors (2) Taking the explanatory variable AI and the explained variable ETR as threshold variables, the impact of different AI and ETR development stages is discussed. In summary, each nonlinear model in Models N1-N3 includes 4 threshold variables (IND, OPEN, AI, ETR) and a total of 12 sub-models need to be estimated.

Before analyzing the regression results of the nonlinear model, we need to test the threshold effect and the validity of the threshold value. The results are shown in Table 6 . The result of Sup Wald test shows that all 12 sub-models reject the null hypothesis of linearity at the 1% level, indicating the existence of a threshold effect. From the result of threshold values, we can see that the threshold values of the other 11 sub-models are all significant, and all fall within the 95% confidence interval except for the sub-model in Model N3 with AI as the threshold variable. It means the threshold values in these models are accurately estimated and the estimation results of these models are reasonable. For the AI sub-model of Model N3, the insignificance of the threshold means that the division of samples may not be optimal. However, considering that the threshold effect does exist, the nonlinear regression results of Model N3 are also worth analyzing.

The sample was classified into the upper or lower regime based on whether the threshold variable was higher than the threshold value obtained. Using the threshold estimated result of Model N1 as an example, we plot the sample distributions of the four threshold variables from 1993 to 2019 in Fig. 1 . First, the proportion of samples below the IND threshold rose from 10% in 1993 to 45% in 2019 indicating that the industrial structure in the sample countries is being gradually optimized. Second, the proportion of samples below the OPEN threshold dropped from 63% to 30% from 1993 to 2007 but rose again to 37% in 2019. This suggests a slowdown in the pace of global openness after 2007 and a resurgence of reverse globalization in recent years. Third, AI has experienced explosive growth during the sample period, with the number of countries with an AI below the threshold of 6.5853 plummeting from 88% to 40%. Finally, the proportion of countries below the threshold energy transition of 3.89% only fell by 12% in 26 years, indicating a possible bottleneck in global renewable energy development.

Thresholds are 3.1903 for IND, 3.9291 for OPEN, 6.5853 for AI and 1.3581 for ETR.

Regression results of dynamic panel threshold models

According to the estimation method of DPTM shown in Eqs. 7 – 11 , we obtained the parameter estimation results of 12 nonlinear sub-models. The results of the nonlinear regressions with IND, OPEN, AI, and ETR as thresholds are shown sequentially in Tables 6 – 9 . We can see that the estimation result of each sub-model includes two parts, namely “Lower estimation” and “Threshold effect”. The former is the estimated coefficient for the low regime, while the latter represents the difference between the estimated coefficients for the high and low regimes, which can be regarded as the threshold effect when the threshold variable crosses the threshold.

Result with IND as a threshold

Table 7 shows the nonlinear regression results when industrial structure is used as the threshold variable and ecological footprint, carbon emissions and energy transition are used as dependent variables. It can be found that the “threshold effect” estimation result includes the lag term of the dependent variable, all independent variables, and constant terms, which is consistent with the DPTM model setting shown in Eq. ( 7 ), indicating that the estimated coefficients of all variables may change in the high regime. However, this paper mainly focuses on the results of the key variable AI.

In Model T1, the ecological footprint is the explained variable. The lower regime results show that the impact coefficient of AI on EF is −0.0090, significant at the 1% statistical level. This means 1% AI development reduces ecological footprint by 0.0090% when IND is lower than the threshold. However, the threshold effect estimator is significantly positive at the 1% level, which means that when IND is higher than the threshold, the impact coefficient of AI on EF will increase by 0.0089, from −0.0090 to −0.0001. This shows that in countries with a high proportion of secondary industry, the reducing effect of AI on EF is weakened. In other words, a high proportion of industry is not conducive for AI to reduce its ecological footprint.

The low regime results of Model T2 show that when IND is low, AI significantly reduces carbon emissions with a coefficient of −0.0065. However, the threshold effect estimation results show that when IND is higher, the impact coefficient of AI will significantly increase by 0.0089. Thus the influence coefficient of AI in high areas has changed from negative to positive to 0.0024, exhibiting a U -shaped nexus of AI-CE. This indicates that IND is also not conducive to AI reducing carbon emissions.

Model T3 shows that when IND is low, AI inhibits the energy transition with a coefficient of −0.0209. However, when IND is high, the coefficient increases significantly and reaches 0.0054, which means that the impact of AI on energy transition changes from negative to positive, showing a U -shaped relationship. That is to say, IND promotes the role of AI in energy transition.

In summary, the above results show that IND is not conducive to the ecological footprint reduction and carbon reduction effects of AI, but can promote the energy transition effect of AI. Judging from the magnitude of the threshold effect, IND influences the impact of AI on EF by 0.0089, CE by 0.0089, and ETR by 0.0263. This means IND has a weaker impact on the first two effects of AI, but a stronger impact on the energy transition effect. Regarding such a “double-edged sword effect” brought about by IND, we compared the existing literature and briefly summarized two aspects of possible explanations. First, an important use of AI in industrial countries is to help industrial production, but industrial production activities are not so clean compared to the primary and tertiary industries. Therefore, if AI is used in heavily polluted production scenarios, it is possible that the use of AI in countries with a high proportion of industry corresponds to a higher ecological footprint and carbon emissions, which is supported by Wang et al. ( 2019 ); Zhang and Liu ( 2015 ). Second, the industrial system can provide infrastructure and technology readiness for energy transition (Heffron et al. 2020 ) and it is therefore reasonable that IND promotion the effect of AI on energy transition.

Result with OPEN as a threshold

Table 8 includes all results with OPEN as the threshold. First, both the low regime estimator and the threshold effect estimator in Model T4 are not significant, indicating that the relationship between AI and EF under the influence of OPEN is unclear. The results of Model T5 show that when OPEN is below the threshold, 1% AI corresponds to a significant increase in CE of 0.0058%. However, when OPEN is above the threshold, this coefficient drops significantly to −0.0023, showing an inverted U -shaped relationship between AI and CE, meaning that OPEN promotes the role of AI in reducing carbon emissions. The results of Model T6 show that although the lower regime estimator of AI is not significant, the threshold effect estimator is significant. The latter is much larger than the former in magnitude, so we can say that when OPEN levels are high, AI can significantly contribute to the energy transition. From the magnitude of the threshold effect, after trade openness crosses the threshold, every 1% AI growth will reduce an additional 0.0081% carbon emission and an additional 0.0132% energy transition. We conclude that a higher degree of trade openness overall amplifies the carbon emissions reduction effect and energy promotion effect of AI to exert a more positive sustainability effect.

Compared with the research of other scholars, Demena and Afesorgbor ( 2020 ) found that openness has a positive effect on reducing carbon emissions, Jacqmin ( 2018 ); Urom et al. ( 2022 ) discussed and confirmed the significance of openness in promoting energy transition. Based on the above discussion, from the perspective of AI, our results support the pollution halo theory that trade openness enhances the environmental sustainability in host countries, which runs counter to the arguments of the trade protectionists. In this regard, the global trend of trade protectionism and anti-globalization after the 2008 financial crisis is not a wise choose. We, therefore, call for worldwide countries to strengthen international coordination and create an open environment for free trade and knowledge sharing so as to make better use of AI and accelerate the pace of sustainable development.

Result with AI as a threshold

In order to investigate the differences in the sustainability effects of AI in different stages and levels of AI development, we analyze the threshold effect of AI as a threshold variable based on Table 9 . The threshold effect estimators in Models T7–T9 show that the three sustainable effects of AI are all significantly affected by the level of AI development. Higher AI will bring an additional 0.1401% reduction in ecological footprint, 0.0804% reduction in carbon emissions, and 0.0423% increase in energy transition progress for every 1% AI growth. Therefore, we conclude that AI has a scale effect. It is worth noting that the threshold effect estimate of AI is significantly higher in magnitude than that of OPEN, which reminds us that we should pay more attention to this scale effect. Specifically, according to Models T7 and T8, the effect of AI on EF and CE is not significant in the lower regime. This insignificant effect will change into a significant reduction effect as the AI level increases. The results of Model T9 show that AI always has a significant promotion effect on energy transition, and this effect is gradually amplified as AI exceeds the threshold.

Similar scale effects are often shown in other scholars’ research on technological factors (Lahouel et al. 2021 ; Xu and Chen 2021 ). For example, (Lahouel et al. 2021 ) non-linear research results show that different ICT levels correspond to different relationships between TFP and carbon emissions. The higher the ICT, the stronger the carbon emission reduction effect of TFP. A possible explanation for this is, that as the application and research of AI technology in a country’s AI industry becomes more and more in-depth, AI-related experience, skills and talents will gradually accumulate. In turn, the cost of using AI will decrease, and the functions will become stronger, so the sustainability effects gradually appear and increase. Considering that AI has experienced explosive growth during the sample period, the continuation of this trend is of great significance for giving full play to the positive effects of AI. Recently, Indonesia, Russia, Saudi Arabia and other countries have launched medium and long-term plans to provide funding for the AI industry and improve AI development goals. Our results emphasize the significance of the uncompromising implementation of these plans for a sustainable future.

Result with ETR as a threshold

According to the threshold effect estimation results in Table 10 , only Model T10 and Model T12 provide significant results. This shows that when ETR exceeds the threshold, the effect of AI on ecological footprint and energy transition is significantly affected, but the effect of AI on carbon emissions is not significant. Specifically, the complete results of Model T10 and Model T12 demonstrate the inverted V relationship both between AI and EF and between AI and ETR. That is to say, when ETR is lower, AI increases ecological footprint but promotes energy transition; when ETR is high, AI reduces ecological footprint but inhibits energy transition. The above results show that the impact of energy transition on the sustainability of AI is also two-fold. A higher level of energy transition improves the role of AI in reducing ecological footprint but limits the promotion effect of AI on ETR.

There are mainly two comments on this finding. On the one hand, a higher level of energy transition means a larger proportion of clean energy in a country’s energy structure. Therefore, in countries with high energy transition, it is reasonable that the ecological footprint reduction effect of AI is better. This finding is consistent with some recent studies, which confirmed that energy transition can improve the environmental performance of agriculture (Sharma et al. 2021 ), ICT (Huang et al. 2022 ) and digital economy (Li et al. 2021 ). On the other hand, it is quite surprising that in countries with higher energy transition levels, the effect of AI on ETR would shift from a facilitator to a depressor. This result contradicts the findings of Dong et al. ( 2022a ), who believe that renewable energy development has a self-reinforcing effect, which means that as the energy transition progresses, the development of renewable energy will automatically become faster and faster. However, our results imply a bottleneck effect during the development of the energy transition from the perspective of AI. If renewable energy production cannot keep pace with demand growth, it is possible that the energy consumption requirement of AI may translate into demand for fossil fuels and undermine energy transition outcomes. Therefore, based on our findings, while accelerating the energy transition to improve the environmental effects of AI, countries need to address bottlenecks and structural issues in the energy transition to prevent the potential negative impacts of AI on the energy transition.

Robustness test

We test the robustness of the baseline regression results in two ways. First, we change the estimation method from system GMM to difference GMM(Diff-GMM). Second, we replace the core explanatory variable from the stock of industrial robots (AI) with the stock of industrial robots per capita (ai). The results are shown in Table 8 and key results are summarized in Fig. 2 accordingly. Figure 2 clearly shows that after changing the estimation method and replacing the core variables, the direction and relative size of AI’s influence coefficients on EF, CE, and ETR have not changed. In other words, AI reduces EF and CE and increases ETR; and AI has the largest impact coefficient on ETR, followed by EF, and finally CE. In addition, the Sargan test and AR (2) of the six robustness test models in Table 11 are both significantly greater than 0.1, indicating that the dynamic model used in the robustness test is effective. In summary, the linear research results are fully credible.

According to Table 10 , all coefficients are significant at the 1% level.

Then we adopted the method of replacing the core explanatory variable with ai to improve the robustness of the nonlinear regression results (Model T1–T12), and the results are shown in Table 12 . For all models, the Sup Wald p -value is less than 0.1, indicating that the non-linear relationship persists after replacing explanatory variables. Except for Model RT9, the estimates of other thresholds are all significant at the 5% level, and these results are in full agreement with the original model.

Subsequently, we are mainly concerned with the threshold effect estimation results of the core variable AI, which represents the impact of the threshold variable on the sustainability effect of AI after crossing the threshold. The comparison between the original results and the robustness check results is visualized in Fig. 3 . First, IND is beneficial to ai to reduce EF and CE, and is not conducive to ai to increase ETR. Second, OPEN has no significant impact on the relationship between ai and EF, but it promotes the reduction effect of ai on CE and the improvement effect of ai on ETR. Third, when ai is at a high level, the reduction effect of ai on EF and CE is amplified, indicating that ai has a scale effect. Finally, a higher energy transition process can significantly help ai decrease EF and CE; however, it inhibited the promoting effect of ai on ETR. In conclusion, the results did not change significantly after replacing the core explanatory variables, suggesting the robustness of the non-linear results.

Note: The bars represent the threshold effect estimator of AI under the influence of the corresponding threshold variable (IND, OPEN, AI or ETR). Results that are not significant at the 5% level are represented by 0. The original result coefficients collected from Tables 6 – 9 , and the robustness coefficients collected from Table 11 .

Conclusions and policy implications

Conclusions.

This article aims to comprehensively evaluate the linear and nonlinear impact of AI on environmental sustainability. First, this article uses panel data from 67 countries from 1993 to 2019 to construct a series of SYS-GMM models as benchmark models to evaluate the impact of AI on ecological footprint, carbon emissions and energy transition. On this basis, this paper further builds a series of DPTM models to explore how external variables, including industrial structure and trade openness, affect AI sustainability. Finally, we used the DPTM model to reveal the changing laws of AI’s environmental effects during the development process of AI and energy transition and obtained a series of robust research results. The fingdings are as follows.

From the real-world and multiple environmental impacts of AI, benchmark results show that AI contributes to environmental sustainability by reducing ecological footprints and carbon emissions, and promoting energy transition. Every 1% increase in the level of AI development corresponds to a 0.0025% increase in energy transition, a 0.0018% decrease in ecological footprint, and a 0.0013% decrease in carbon emissions. In other words, AI has the strongest effect in promoting energy transition, followed by the reduction effect on ecological footprint, and then the impact of AI on carbon emissions.

From the perspective of external conditions affecting AI, on the one hand, the nonlinear results of IND as a threshold variable indicate that the industrial share is a double-edged sword. More specifically, countries with a higher proportion of industries have a significantly stronger effect on AI’s promotion of energy transition, but the ecological footprint and carbon emissions corresponding to AI have also increased to a certain extent. In detail, for countries with a higher industrial share, every 1% AI growth corresponds to an additional 0.0263% energy transition, 0.0089% ecological footprint, and 0.0089% carbon emissions. On the other hand, the nonlinear results of OPEN as a threshold variable indicate that trade openness can overall promote the sustainability effect of AI. Specifically, when it is higher than the threshold of trade openness, for a 1% AI growth, carbon emissions are reduced by an additional 0.0081%, the energy transition effect is increased by an additional 0.0132%, and the ecological footprint does not change significantly.

From the nonlinear laws of AI’s effect as AI and energy transition process, there is a significant scale effect in the development process of AI. This effect is much stronger than the amplification effect brought about by trade openness. At a higher stage of AI development, every 1% of AI development will bring an additional 0.1401% ecological footprint reduction effect, 0.0804% carbon emission reduction effect, and 0.0423% energy transition promotion effect. However, there may be bottlenecks in the promotion effect of AI on energy transition. When the level of energy transition is high, the effect of AI on reducing the ecological footprint is increased but the promoting effect of AI on energy transition is suppressed. Specifically, as the energy transition process increases and exceeds the threshold, the ecological footprint corresponding to a 1% AI increase decreases by an additional 0.0319%, and the energy transition process also decreases by an additional 0.019%.

Policy implications

The findings discussed herein have significant policy implications:

Overall, the impact of AI on reducing ecological footprint, and carbon emissions and promoting energy transition shows that AI can help improve environmental sustainability. Countries should increase investment in AI R&D, propose AI development strategies and establish long-term development mechanisms. At the same time, policies should standardize the mode of the AI industry and guide AI to contribute to environmental protection actions or energy transition. At present, more than 30 countries have formulated AI development goals and action plans, aiming to encourage, fund and guide the investment and application of AI, such as China’s “The Next Generation Artificial Intelligence Development Plan” and the European Union’s “Artificial Intelligence Coordination Program”. The implementation of these goals is of great significance for promoting the development of the AI industry and moving toward sustainable development.

Countries with high industrial proportions should make full use of the advantages of their own industrial systems to develop renewable energy with the help of AI. At the same time, industrial countries cannot ignore the pollution problem in the industrial sector and should introduce policies as soon as possible to regulate the application scenarios of AI in industrial production to prevent AI from being used to accelerate pollution.

Trade openness plays an important role in exerting the positive effects of AI. The trends of anti-globalization and “trade protection” that have emerged in recent years are not conducive to the sustainable development of all countries. We call for strengthening international coordination and creating an open environment for free trade and knowledge sharing.

AI itself has scale effects, so setting binding long-term goals for AI development is of great significance to all countries, especially those countries that are lagging behind in AI. Recently, Indonesia proposed the Stranas KA plan to continue to provide funding for the AI industry and guide the country’s development of artificial intelligence from 2020 to 2045. Russia, Portugal, Saudi Arabia and other countries have also launched medium- and long-term plans to improve AI development goals. Our research supports these mid- to long-term strategies.

A higher level of energy transition can help AI better reduce its environmental footprint. However, governments need to focus on the structural problems and bottlenecks of energy transition such as rising marginal costs and difficulty in technological breakthroughs. For example, currently dozens of countries, including China, the United States, and Japan, have increased their renewable energy production targets in recent years, and have increased subsidies for renewable energy sources with high costs but huge potential, such as offshore wind power, to reduce their marginal costs and ensure supply.

Outlook and limitations

Although this study evaluated for the first time AI has linear and non-linear effects on EF, CE, and ETR, there are still some deficiencies. First, although this paper models the multiple impacts of AI, there are still some factors to be evaluated, such as the impact of AI on different types of energy or other pollutants. Future research can consider the impact of AI on sulfur dioxide, industrial wastewater, etc. Secondly, our research is based on the national level, and it is difficult to reveal specific rules at the sub-national level. In the future, this research can be expanded to a more granular level, such as the provincial or city level.

Data availability

The datasets publicly available should be through https://doi.org/10.7910/DVN/BN3PID .

The hallmark feature of DPTM is that it allows threshold variables \({q}_{{it}}\) to be endogenous. So \(E\left({q}_{{it}}{\Delta \varepsilon }_{{it}}\right){\boldsymbol{\ne }}0\) , which means that the threshold variable \({q}_{{it}}\) is not included in instrumental variable set \({{\boldsymbol{z}}}_{{\boldsymbol{it}}}\) .

According to the IEA, the scope of renewable energy statistics includes traditional utilization of biomass, modern bioenergy, hydroelectric power, wind energy, solar photovoltaic and other renewable energy sources.

Moore’s Law emphasizes the speed of information technology development, and holds that the performance of computers increases exponentially over time, doubling about every 18 months.

Total natural resource rents include oil rents, coal rents, natural gas rents, mineral rents and forest rents.

Due to data limitations of artificial intelligence, the scope of the study does not include low-income countries.

Acemoglu D, Restrepo P (2020) Robots and jobs: evidence from US labor markets. J Political Econ 128(6):2188–2244

Article   Google Scholar  

Ahmad M, Jiang P, Majeed A et al. (2020) The dynamic impact of natural resources, technological innovations and economic growth on ecological footprint: an advanced panel data estimation. Resour Policy 69:101817

Ahmad T, Zhu H, Zhang D et al. (2022) Energetics systems and artificial intelligence: applications of industry 4.0. Energy Rep 8:334–361

Ansari MA (2022) Re-visiting the environmental Kuznets curve for ASEAN: a comparison between ecological footprint and carbon dioxide emissions. Renew Sustain Energy Rev 168:112867

Article   CAS   Google Scholar  

Blundell R, Bond S (1998) Initial conditions and moment restrictions in dynamic panel data models. J Econ 87(1):115–143

Chen P, Gao J, Ji Z et al. (2022a) Do artificial intelligence applications affect carbon emission performance?—Evidence from panel data analysis of Chinese cities. Energies 15(15). https://doi.org/10.3390/en15155730

Chen Y, Cheng L, Lee C-C (2022b) How does the use of industrial robots affect the ecological footprint? International evidence. Ecol Econ 198. https://doi.org/10.1016/j.ecolecon.2022.107483

Cheng Z, Li L, Liu J (2018) Industrial structure, technical progress and carbon intensity in China’s provinces. Renew Sustain Energy Rev 81:2935–2946

Demena BA, Afesorgbor SK (2020) The effect of FDI on environmental emissions: evidence from a meta-analysis. Energy Policy 138:111192

Ding T, Li J, Shi X et al. (2023) Is artificial intelligence associated with carbon emissions reduction? Case of China. Resour Policy 85:103892

Dong B, Ma X, Zhang Z et al. (2020) Carbon emissions, the industrial structure and economic growth: evidence from heterogeneous industries in China. Environ Pollut 262:114322

Article   CAS   PubMed   Google Scholar  

Dong F, Li Y, Gao Y et al. (2022a) Energy transition and carbon neutrality: exploring the non-linear impact of renewable energy development on carbon emission efficiency in developed countries. Resour Conserv Recycling 177:106002

Dong F, Zhu J, Li Y et al. (2022b) How green technology innovation affects carbon emission efficiency: evidence from developed countries proposing carbon neutrality targets. Environ Sci Pollut Res 29(24):35780–35799

Ehigiamusoe KU, Dogan E (2022) The role of interaction effect between renewable energy consumption and real income in carbon emissions: evidence from low-income countries. Renew Sustain Energy Rev 154:111883

Gu W, Zhao X, Yan X et al. (2019) Energy technological progress, energy consumption, and CO2 emissions: empirical evidence from China. J Clean Prod 236:117666

He J, Baxter SL, Xu J et al. (2019) The practical implementation of artificial intelligence technologies in medicine. Nat Med 25(1):30–36

Article   CAS   PubMed   PubMed Central   Google Scholar  

Heffron R, Körner M-F, Wagner J et al. (2020) Industrial demand-side flexibility: a key element of a just energy transition and industrial development. Appl Energy 269:115026. https://doi.org/10.1016/j.apenergy.2020.115026

Hilty LM, Hercheui MD (2010). ICT and sustainable development IFIP International Conference on Human Choice and Computers. Berlin, Heidelberg: Springer Berlin Heidelberg 2010:227–235

Holzinger A, Keiblinger K, Holub P et al. (2023) AI for life: trends in artificial intelligence for biotechnology. N Biotechnol 74:16–24

Howard J (2019) Artificial intelligence: implications for the future of work. Am J Ind Med 62(11):917–926

Article   PubMed   Google Scholar  

Huang Y, Haseeb M, Usman M et al. (2022) Dynamic association between ICT, renewable energy, economic complexity and ecological footprint: is there any difference between E-7 (developing) and G-7 (developed) countries? Technol Soc 68:101853

IFR (2023) World Robotics Report 2023. International Federation of Robotics. https://ifr.org/wr-industrial-robots

ISTIC (2023) 2022 Global Artificial Intelligence Innovation Index Report. China Institute of Scientific and Technological Information. https://www.istic.ac.cn/html/1/284/338/1506840089869938181.html

Jacqmin J (2018) The role of market-oriented institutions in the deployment of renewable energies: evidences from Europe. Appl Econ 50(2):202–215

Article   MathSciNet   Google Scholar  

Jha SK, Bilalovic J, Jha A et al. (2017) Renewable energy: present research and future scope of Artificial Intelligence. Renew Sustain Energy Rev 77:297–317

Jiang W, Gong S-H, Li X-T (2022) Robot shock, capital-embodied technological progress and carbon reduction of manufacturing sector—theoretical analysis and empirical evidence from China. China Ind Econ 10:24–42. https://doi.org/10.19581/j.cnki.ciejournal.2022.10.002

Lahouel BB, Taleb L, Zaied YB et al. (2021) Does ICT change the relationship between total factor productivity and CO2 emissions? Evidence based on a nonlinear model. Energy Econ 101:105406

Lee M, He G (2021) An empirical analysis of applications of artificial intelligence algorithms in wind power technology innovation during 1980–2017. J Clean Prod 297:126536

Lei Y, Liang Z, Ruan P (2023) Evaluation on the impact of digital transformation on the economic resilience of the energy industry in the context of artificial intelligence. Energy Rep. 9:785–792

Li R, Li L, Wang Q (2022a) The impact of energy efficiency on carbon emissions: evidence from the transportation sector in Chinese 30 provinces. Sustain Cities Soc 82:103880. https://doi.org/10.1016/j.scs.2022.103880

Li R, Wang Q, Guo J (2024) Revisiting the Environmental Kuznets Curve (EKC) hypothesis of carbon emissions: exploring the impact of geopolitical risks, natural resource rents, corrupt governance, and energy intensity. J Environ Manag 351:119663. https://doi.org/10.1016/j.jenvman.2023.119663

Li R, Wang Q, Li L et al. (2023) Do natural resource rent and corruption governance reshape the environmental Kuznets curve for ecological footprint? Evidence from 158 countries. Resour Policy 85:103890

Li Y, Yang X, Ran Q et al. (2021) Energy structure, digital economy, and carbon emissions: evidence from China. Environ Sci Pollut Res 28:64606–64629

Li Y, Zhang Y, Pan A et al. (2022b) Carbon emission reduction effects of industrial robot applications: heterogeneity characteristics and influencing mechanisms. Technol Soc 70. https://doi.org/10.1016/j.techsoc.2022.102034

Li Z, Wang J (2022) The dynamic impact of digital economy on carbon emission reduction: evidence city-level empirical data in China. J Clean Prod 351:131570

Liu J, Liu L, Qian Y et al. (2022a) The effect of artificial intelligence on carbon intensity: evidence from China’s industrial sector. Socio-Econ Plan Sci 83. https://doi.org/10.1016/j.seps.2020.101002

Liu J, Liu L, Qian Y et al. (2022b) The effect of artificial intelligence on carbon intensity: evidence from China’s industrial sector. Socio-Econ Plan Sci 83:101002

Luan F, Yang X, Chen Y et al. (2022) Industrial robots and air environment: a moderated mediation model of population density and energy consumption. Sustain Prod Consum 30:870–888. https://doi.org/10.1016/j.spc.2022.01.015

Lyu W, Liu J (2021) Artificial Intelligence and emerging digital technologies in the energy sector. Appl Energy 303:117615

Markard J (2018) The next phase of the energy transition and its implications for research and policy. Nat Energy 3(8):628–633

Article   ADS   Google Scholar  

Moyer JD, Hughes BB (2012) ICTs: do they contribute to increased carbon emissions? Technol Forecast Soc Change 79(5):919–931

Network GF (2019) Global footprint network. Obtenido de Global Footprint Network: http://www.footprintnetwork.orgonline . Accessed: 1–10

Pesaran MH (2020) General diagnostic tests for cross-sectional dependence in panels. Empir Econ 60(1):13–50. https://doi.org/10.1007/s00181-020-01875-7

Probst D (2023) Aiming beyond slight increases in accuracy. Nat Rev Chem 7(4):227–228

Seo MH, Kim S, Kim Y-J (2019) Estimation of dynamic panel threshold model using Stata. Stata J 19(3):685–697

Shahbaz M, Wang J, Dong K et al. (2022) The impact of digital economy on energy transition across the globe: the mediating role of government governance. Renew Sustain Energy Rev 166:112620

Sharma GD, Shah MI, Shahzad U et al. (2021) Exploring the nexus between agriculture and greenhouse gas emissions in BIMSTEC region: the role of renewable energy and human capital as moderators. J Environ Manag 297:113316

Song J, Chen Y, Luan F (2023) Air pollution, water pollution, and robots: is technology the panacea. J Environ Manag 330:117170. https://doi.org/10.1016/j.jenvman.2022.117170

Taddeo M, Tsamados A, Cowls J et al. (2021) Artificial intelligence and the climate emergency: opportunities, challenges, and recommendations. One Earth 4(6):776–779. https://doi.org/10.1016/j.oneear.2021.05.018

Tzeremes P, Dogan E, Alavijeh NK (2023) Analyzing the nexus between energy transition, environment and ICT: a step towards COP26 targets. J Environ Manag 326:116598

Urom C, Abid I, Guesmi K et al. (2022) Renewable energy consumption, globalization, and economic growth shocks: evidence from G7 countries. J Int Trade Econ Dev 31(2):204–232

Wackernagel M, Rees W (1998) Our ecological footprint: reducing human impact on the earth (Vol. 9). New Society Publishers

Wang J, Wang W, Liu Y et al. (2023a) Can industrial robots reduce carbon emissions? Based on the perspective of energy rebound effect and labor factor flow in China. Technol Soc 72. https://doi.org/10.1016/j.techsoc.2023.102208

Wang Q, Hu S, Li R (2024a) Could information and communication technology (ICT) reduce carbon emissions? The role of trade openness and financial development. Telecommun Policy 48(3):102699. https://doi.org/10.1016/j.telpol.2023.102699

Wang Q, Ren F, Li R (2024b) Exploring the impact of geopolitics on the environmental Kuznets curve research. Sustainable Development. 32(3):1700–1722

Wang Q, Sun T, Li R (2023c) Does artificial intelligence promote green innovation? An assessment based on direct, indirect, spillover, and heterogeneity effects. Energy Environ, https://doi.org/10.1177/0958305X231220520

Wang Q, Zhang F, Li R et al. (2024c) Does artificial intelligence facilitate the energy transition and curb carbon emissions? The role of trade openness. J Clean Prod 447:141298

Wang S, Zeng J, Liu X (2019) Examining the multiple impacts of technological progress on CO 2 emissions in China: a panel quantile regression approach. Renew Sustain Energy Rev 103:140–150

WDI (2021) World Development Index. World Bank. https://datatopics.worldbank.org/world-development-indicators/user-guide.html

Westerlund J (2005) New simple tests for panel cointegration. Econom Rev 24(3):297–316

Wu SR, Shirkey G, Celik I et al. (2022) A review on the adoption of AI, BC, and IoT in sustainability research. Sustainability 14(13):25. https://doi.org/10.3390/su14137851

Xu B, Chen J (2021) How to achieve a low-carbon transition in the heavy industry? A nonlinear perspective. Renew Sustain Energy Rev 140:110708

Yao S, Zhang S, Zhang X (2019) Renewable energy, carbon emission and economic growth: a revised environmental Kuznets Curve perspective. J Clean Prod 235:1338–1352

Yi M, Liu Y, Sheng MS et al. (2022) Effects of digital economy on carbon emission reduction: new evidence from China. Energy Policy 171:113271

Yu L, Zeng C, Wei X (2022) The impact of industrial robots application on air pollution in China: mechanisms of energy use efficiency and green technological innovation. Sci Prog 105(4):368504221144093. https://doi.org/10.1177/00368504221144093

Zavyalova EB, Volokhina VA, Troyanskaya MA et al. (2023) A humanistic model of corporate social responsibility in e-commerce with high-tech support in the artificial intelligence economy. Human Soc Sci Commun 10(1):274. https://doi.org/10.1057/s41599-023-01764-1

Zhang C, Liu C (2015) The impact of ICT industry on CO2 emissions: a regional analysis in China. Renew Sustain Energy Rev 44:12–19

Zhang X, Liu P, Zhu H (2022) The impact of industrial intelligence on energy intensity: evidence from China. Sustainability 14(12):721914. https://doi.org/10.3390/su14127219

Zheng S, Yang J, Yu S (2021) How renewable energy technological innovation promotes renewable power generation: evidence from China’s provincial panel data. Renew Energy 177:1394–1407

Zhong J, Zhong Y, Han M et al. (2023) The impact of AI on carbon emissions: evidence from 66 countries. Appl Econ 56(25):2975–2989

Download references

Acknowledgements

This work is supported by the “Youth Innovation Team Project” of the Higher Education Institutions under the Shandong Provincial Department of Education (No. 2023RW015), the National Natural Science Foundation of China (Grant No. 71874203).

Author information

Authors and affiliations.

School of Economics and Management, China University of Petroleum, Qingdao, 266580, People’s Republic of China

Qiang Wang & Yuanfan Li

School of Economics and Management, Xinjiang University, Wulumuqi, 830046, People’s Republic of China

Qiang Wang & Rongrong Li

You can also search for this author in PubMed   Google Scholar

Contributions

QW: conceptualization, methodology, software, data curation, writing—original draft preparation, supervision, writing—reviewing and editing. YL: methodology, software, data curation, investigation writing—original draft, writing—reviewing and editing. RL: conceptualization, methodology, data curation, investigation writing—original draft, writing—reviewing.

Corresponding authors

Correspondence to Qiang Wang or Rongrong Li .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Ethical statements

This article does not contain any studies with human participants performed by any of the authors.

Informed consent

Additional information.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Wang, Q., Li, Y. & Li, R. Ecological footprints, carbon emissions, and energy transitions: the impact of artificial intelligence (AI). Humanit Soc Sci Commun 11 , 1043 (2024). https://doi.org/10.1057/s41599-024-03520-5

Download citation

Received : 29 January 2024

Accepted : 29 July 2024

Published : 14 August 2024

DOI : https://doi.org/10.1057/s41599-024-03520-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Deep dive into RNA: a systematic literature review on RNA structure prediction using machine learning methods

• Open access
• Published: 15 August 2024
• Volume 57 , article number  254 , ( 2024 )

Cite this article

You have full access to this open access article

• Michał Budnik 1 , 2 ,
• Jakub Wawrzyniak 2 ,
• Łukasz Grala 2 ,
• Miłosz Kadziński 1 &
• Natalia Szóstak 2 , 3  

133 Accesses

Explore all metrics

The discovery of non-coding RNAs (ncRNAs) has expanded our comprehension of RNAs’ inherent nature and capabilities. The intricate three-dimensional structures assumed by RNAs dictate their specific functions and molecular interactions. However, the limited number of mapped structures, partly due to experimental constraints of methods such as nuclear magnetic resonance (NMR), highlights the importance of in silico prediction solutions. This is particularly crucial in potential applications in therapeutic drug discovery. In this context, machine learning (ML) methods have emerged as prominent candidates, having previously demonstrated prowess in solving complex challenges across various domains. This review focuses on analyzing the development of ML-based solutions for RNA structure prediction, specifically oriented toward recent advancements in the deep learning (DL) domain. A systematic analysis of 33 works reveals insights into the representation of RNA structures, secondary structure motifs, and tertiary interactions. The review highlights current trends in ML methods used for RNA structure prediction, demonstrates the growing research involvement in this field, and summarizes the most valuable findings.

Similar content being viewed by others

Advances and opportunities in RNA structure experimental determination and computational modeling

Big data and deep learning for RNA biology

Integrating end-to-end learning with deep geometrical potentials for ab initio RNA structure prediction

Explore related subjects.

• Artificial Intelligence

Avoid common mistakes on your manuscript.

1 Introduction

Ribonucleic acid (RNA) is a polymer molecule essential for converting genetic information from deoxyribonucleic acid (DNA) into proteins. For a while, this was thought of as the sole role of RNA. However, studies have unraveled other significant functions. One of the first such RNAs discovered by Stark et al. ( 1978 ) was RNAse P, a ribozyme that cleaves a precursor sequence of RNA in tRNA molecules. This was followed by Yang et al. ( 1981 ), who described small nuclear RNAs (snRNAs), other non-coding RNA that exists in a nucleus and is responsible for splicing. Over the years, the number of functional non-coding RNAs (ncRNAs) discovered has expanded vastly (Wilusz et al. 2009 ; Wang and Chang 2011 ; Ulitsky and Bartel 2013 ; Fu 2014 ; Kopp and Mendell 2018 ).

Non-coding RNA has been found to control protein synthesis, regulate transcription and translation, modify and stabilize RNA, and regulate gene expressions at different levels (Doudna and Cech 2002 ; Meister and Tuschl 2004 ; Garst et al. 2011 ; Serganov and Nudler 2013 ; Mortimer et al. 2014 ). These diverse functions of ncRNAs take part in many complex biological processes vital for human health, such as immune cell, neural, or muscle development (Sun and Kraus 2015 ; Mehta and Baltimore 2016 ; Andersen and Lim 2017 ; Constantin 2018 ).

However, RNA functions are not determined solely by information in the nucleotide chain but, similarly to proteins, by the three-dimensional shape into which the given sequence folds (Graf and Kretz 2020 ). This shape allows RNA to interact with DNA, proteins, lipids (Mańka et al. 2021 ; Czerniak and Saenz 2021 ), and other molecules. Therefore, it is essential to know and understand the tertiary structure of RNA as a foundation for the design of potential targeted drugs based on RNA (Childs-Disney et al. 2022 ). Although numerous ncRNA sequences are available and their numbers increase rapidly (Stephens et al. 2015 ), their structures are still poorly determined. Along with the underestimation of the role of RNA, this is one of the reasons why, for a long time, the central part of research on predicting biological structures has been focused not on RNA but on the problem of protein structure prediction.

Accurately predicting RNA tertiary structures provides valuable insight into their biological functions. By solving RNA 3D structures, researchers can identify areas crucial for catalysis, regulation, and protein interaction. These functional sites frequently include highly precise arrangements of nucleotides inside the folded structure. Unveiling RNA’s intrinsic ability to generate specialized three-dimensional shapes and selective interactions with other biomolecules allows researchers to use it for therapeutic drug discovery. By knowing the RNA structure, researchers can alter RNA activity in a targeted manner. It enables the rational design of compounds that target functional RNA structures, marking a paradigm shift from traditional protein-centric drug discovery. This significantly widens the possibilities for fighting a wide range of diseases, from neurodegenerative disorders to various types of cancers (Sun and Kraus 2015 ; Schmitt and Chang 2016 ).

Some experimental methods to obtain RNA’s atomic coordinates include X-ray crystallography or nuclear magnetic resonance (NMR). These methods, though quite reliable even for long sequences with multiple possible conformations, have several constraints. These include the long time required to gather data, the costs of running the apparatus, or the need for specialized equipment and personnel to perform the experiments (Kotar et al. 2020 ). Knowing these limitations, numerous works have turned to in silico methods of predicting RNA’s structures. The leading paradigm for early solutions was to identify a structure with minimum free energy (MFE), as it was the most likely state in which a molecule would exist, similar to proteins (Anfinsen 1973 ). Following this approach, various algorithms have been proposed and tested to solve both secondary and tertiary structure prediction problems. These approaches are based on thermodynamic simulations evaluated by dynamic programming approaches (Mathews and Zuker 2004 ; Eddy 2004 ; Havgaard et al. 2005 ), statistical mechanics (Ding and Lawrence 2003 ; Mathews 2004 ; Ding et al. 2005 ), or genetic algorithms (Shapiro and Navetta 1994 ; Chen et al. 2000 ; Taneda 2012 ), among others.

Although they achieved satisfactory results, especially for secondary structure prediction, these methods have failed to achieve significant improvements in the accuracy and speed of predicting the RNA structure. As an alternative, with development in both optimization methods and computational possibilities, machine learning (ML) based methods started being utilized for various parts of the prediction pipeline. At first, these methods have not received much attention due to comparatively low prediction scores. This was partially due to the lack of higher volumes of data available for the training part required by ML algorithms. However, the amount of available data has progressively grown over the years. In recent years, ML-based methods have surpassed the capabilities of classical algorithms and are now the main active area of research on RNA structure prediction. Taking this into account, the main contribution of this review is to describe and compare developments in the ML-based RNA structure prediction field, with a particular focus on deep learning (DL) based methods. The specific machine learning approaches used, including classical ML methods, recurrent neural networks, and reinforcement learning, have been gathered and described, showing a clear shift toward utilizing different types of neural network architectures in recent years.

In this context, it is worth mentioning that both modern protein and RNA structure predictions share deep learning techniques, utilizing models like CNNs and transformers for capturing spatial dependencies and sequence-structure relationships. Both fields benefit from pre-training on large datasets and leveraging evolutionary information for feature extraction. This similarity can be seen in solutions like DeepFold (Pearce et al. 2022a ) and DeepFoldRNA (Pearce et al. 2022b ), which use common methodologies for both problems. However, RNA structure prediction is uniquely challenging due to the dynamic nature of RNA chains, their varied structures, and the limited high-resolution structural data compared to proteins. While protein secondary structures consist of alpha-helices and beta-sheets, RNA secondary structures comprise various structural elements such as hairpins, bulges, internal loops, pseudoknots, and multi-branch loops. This plethora of RNA structural motifs means that, in practice, there may be an insufficient number of suitable templates for efficient conformational sampling due to the limited size of the available resolved RNA structures.

Several recent studies provide insight into the growing landscape of RNA structure prediction methods and their potential for drug discovery. In particular, Sato and Hamada ( 2023 ) gives a compelling overview of the challenge and its relevance in drug development but lacks a thorough analysis of specific approaches. Further, Zhang et al. ( 2022 ) provides a detailed description of the problem, encompassing biological and chemical considerations, but does not discuss particular algorithms or results. Then, Zhao et al. ( 2021 ) presents machine learning algorithms for RNA structure prediction in an organized manner but does not cover recent deep learning solutions. In contrast, Yu et al. ( 2022 ) provides a thorough analysis of deep learning solutions, but it lacks a qualitative comparison of the works. Finally, Wang et al. ( 2023b ) provides a broad perspective of RNA-related approaches though solely addressing 3D structure prediction.

The purpose of this review is to contribute to the area by presenting a systematic approach as well as knowledge updates on machine and deep learning methods. It concentrates on algorithmic features and provides a comprehensive examination of selected methods and architectures. Furthermore, a full comparison of the addressed subproblems, methodology, and achieved results is provided. This comprehensive approach will provide researchers with significant insights into the possibilities of machine learning and deep learning in RNA structure prediction.

This review organizes the analyzed works based on the specific structure prediction problem they tackle, whether secondary or tertiary. By exploring the research presented, this review aims to compare the architectures and results of the methodologies used and identify potential research gaps. The subsequent sections of this paper are organized as follows. Section  2 introduces the details of the RNA structure and its representations, including a discussion on the secondary structure and the pairing of the bases, as well as the tertiary structure. Section  3 introduces the underlying methods and algorithms utilized for RNA structure prediction. Section  4 presents an overview of the studies together with their availability. Section  5 provides a detailed discussion of the advantages and disadvantages of solutions for the prediction of secondary and tertiary structures. The paper is then concluded in Sect.  6 , summarizing key findings and potential research gaps.

2 RNA structure and representations

This section is intended to familiarize the reader with the structure of RNA and its common representations. In particular, the secondary and tertiary structures are described separately to underlie their specific natures. The goal of explaining these differences is to highlight the vastly differing nature of the prediction problem to be solved, and thus demonstrate the need to separate the methods used in analyses.

2.1 Secondary structure and base pairing

RNAs are molecules created from a chain, arranged in the 5′ to 3′ direction, of four nucleotides distinguished by their nitrogenous bases—guanine (G), uracil (U), adenine (A), and cytosine (C). Similar to DNA, the secondary structure of RNA is defined by canonical base pairing. These include the Watson–Crick pairs (A-U and G-C) and the wobble base pair (G-U). These pairs are established via hydrogen bonds and form a structure in which subsections of paired nucleotides form a helix, while unpaired bases can form various secondary motifs, distinguishing RNA’s from DNA’s structures. The secondary structure of RNA can be represented as a 2D figure of connected base pairs, as shown in Fig.  1 .

An example of the RNA secondary structure based on 6OPE tRNA molecule. Different structural motifs are colored for visual distinction (in blue—stems, in yellow—multiloops, in orange—hairpin loops, and in gray—dangling ends). Structure and visualization obtained from bpRNA (Danaee et al. 2018 ). (Color figure online)

However, it has been observed that the RNA structure, unlike that of DNA, can consist of a wider variety of base pairs (Zhao et al. 2018 ). Three groups of special base pairs can be distinguished, the most commonly occurring being non-canonical base pairs. Up to 40% of all base pairs in an RNA molecule can consist of base pairs other than Watson–Crick pairs or the wobble pair (Leontis and Westhof 2001 ). Another type of atypical base pairs are triples—clusters of three RNA nucleobases that interact edge to edge by hydrogen bonding, mostly creating base pairs from the central base (Almakarem et al. 2012 ). Additionally, G-quadruplexes are increasingly important—structures that consist of four Hoogsteen-bound guanines as planar assemblies (Lorenz et al. 2013 ).

These various interactions between nucleotides describe the secondary structure of the RNA strand. Due to the resulting shape of the 2D view of the RNA molecule, various reoccurring motifs have been identified (Hendrix et al. 2005 ), namely:

Single-stranded regions—sequences of unpaired nucleotides;

Helices—RNA is composed in large part of Watson–Crick pairs creating A-form double helices, though other helical forms have been observed;

Hairpin loops—by the SCOR database classification, hairpin loops must close with a Watson–Crick pairing and have a length between 2 and 14 nucleotides;

Internal/bulge loops—separate helical RNA into two segments with residues not paired canonically in at least one strand of the stem;

Junction loops/multiloops—formed at the intersection of at least three double helices separated by single-strand sequences;

Pseudoknots—structures formed when a single-stranded region of RNA in the loop creates a base pair with complementary nucleotides elsewhere in the RNA (Brierley et al. 2007 ).

Performing in silico calculations to predict these characteristics of the secondary structure requires appropriate data representations for both traditional and machine learning-based methods. Let n denote the length of an RNA molecule. The simplest form is to compose a set of base pairs indices ( i ,  j ) where \(0 \le i< j < \le n\) . However, one of the most popular representations is the so-called “dot-bracket” notation introduced by the ViennaRNA package (Lorenz et al. 2011 ). It is a plain text format using ‘.’ to represent unpaired bases and matched parentheses for canonical base pairs. The format was later extended to cover pseudoknots by introducing square, curly, and angle brackets. A minimalistic representation like this often speeds up computations. A more graphical method to represent the data is to create a contact table (CT table). It is an \(n \times n\) square matrix, where each cell represents an interaction between nucleotides at given indices. For certain algorithms, the RNA structure may also be represented as a graph, where nucleotides are treated as nodes, and edges display base pairing. A visualization of the aforementioned representations is shown in Fig.  2 .

An example of the RNA secondary structure representations based on 6OPE tRNA molecule. The dot-bracket notation (left) displays connections between nucleotides using dots that represent unpaired bases, and matching pairs of opening and closing brackets that represent a connection between given nucleotides in the chain. A contact table (right) is a matrix where x -axis and y -axis represent indices of nucleotides in the sequence. The connection between a given pair of bases is marked by putting a “true” value at a cell indicated by their indices intersection (yellow color on the figure). (Color figure online)

2.2 Tertiary structure

Due to secondary interactions, the RNA molecules fold onto themselves and create three-dimensional conformations. Therefore, the tertiary structure refers to defining the spatial coordinates of atoms in the RNA molecule and the spatial relationships between them (tertiary interaction), as represented in Fig.  3 .

An example of the RNA tertiary structure based on 6OPE tRNA molecule. The curved line represents the sugar-phosphate backbone, while nucleobases are portrayed by their nitrogenous rings. Visualization obtained from PDB (Berman and Henrick 2003 )

The tertiary conformation of an RNA molecule is stabilized by networks of various interactions, and numerous factors play a role in molecule folding, especially osmolytes and ligands, including metal ions and proteins. However, the most critical factors for the final shape of an RNA molecule are stacking interactions. The bases of aromatic nucleic acids are planar, allowing them to stack at contact distance ( \(\sim\) 3.4 Å), maximizing van der Waals interactions. Base stacking interactions are more important than hydrogen bonds for the structural stability of nucleic acids in aqueous solution (Yakovchuk et al. 2006 ).

Analysis of the tertiary structure of RNA has shown that certain shapes of molecules appear to be reoccurring. These shapes, or motifs, are independent of the context in which they occur (Moore 1999 ), and studies have shown that they often define specific functions of the molecule (Ferhadian et al. 2018 ; Ross and Ulitsky 2022 ; Xu et al. 2022 ). Some motifs are widely recognized, including U-turns, tetraloops, or ribose zippers.

An essential feature of the RNA structure is that it is dynamic. The shape acquired by a specific chain of nucleotides is, in theory, the most stable (or thermodynamically favored) structure, also known as the minimum free energy (MFE) structure. However, the so-called folding landscapes are rugged and exhibit multiple local energy minima (Shcherbakova et al. 2008 ). Because of this, an RNA molecule can fold into different conformations depending on the environment.

In silico representation of the structure of RNA molecules often originates from a PDB file format, stored in Protein Data Bank (PDB) (Berman and Henrick 2003 ), one of the most extensive databases for large biological molecules. This extensible plain text format stores information about atomic orthogonal coordinates and polymer division, among others. Based on the choice of the prediction method, further representations are obtained, including graph representations (Townshend et al. 2021 ), distance and angle maps (Pearce et al. 2022b ), or modeling the molecule as a 3D box (Li et al. 2018 ).

3 Methods and algorithms

This section introduces the underlying methods and algorithms that govern RNA structure prediction. It is divided into five parts—an overview of classical algorithms utilized for RNA structure prediction (Sect.  3.1 ), an overview of the machine and deep learning methods utilized in analyzed solutions (Sects.  3.2 and  3.3 ), a description of interpretability of the methods (Sect.  3.4 ) and a computational complexity description of RNA structure prediction problem and its solutions (Sect.  3.5 ).

3.1 Classical methods of predicting RNA structure

Classical methods have adopted a few paradigms that led the structure prediction of RNA. One way of searching for RNA structures is to utilize dynamic programming methods, as did Zuker in one of the most popular solutions in the field—Mfold (Zuker 2003 ). Dynamic programming can solve complex problems by breaking them down into simpler subproblems, solving each of them once, and storing the solutions to avoid redundant computations. This method requires a certain optimization goal, and in the case of RNA structure prediction, one of the most popular goals is to find a structure with minimal free energy in a molecule. The solutions based on dynamic programming generally calculate the free energies of certain substructures, store the results, and repeat the process for increasing the size of the substructures. The results are stored in a \(n \times n\) matrix, where n is the number of nucleotides in a structure, and the minimum free energy structure is determined by analyzing the minimum energy pairings in the matrix.

Another classical approach applied to RNA secondary structure prediction is the Greedy Randomized Adaptive Search Procedure (GRASP), which involves iteratively constructing and refining potential structures. The procedure starts with constructing basic structures using a greedy heuristic to select energetically favorable base pairings, followed by randomization to ensure diversity. Local search algorithms are then used to refine these structures, iteratively making minor alterations to further minimize free energy. Several iterations of this mix of local optimization and greedy and randomized construction produce a range of possible structures. The projected RNA secondary structure is chosen among the best structures, usually the ones with the lowest free energy. This strategy increases the possibility of finding the most stable RNA structure by balancing effective optimization and in-depth solution space search. One of the research works using this algorithm is the work of Fatmi et al. ( 2017 ).

In the same article, another class of methods for RNA structure prediction, namely genetic algorithms (GA), are introduced. These approaches predict RNA secondary structure by modeling evolution: creating an initial population of structures, assessing their fitness based on free energy, and iteratively applying selection, crossover, and mutation to build new generations. This approach balances exploring many structures and exploiting the best solutions, eventually settling on the structure with the lowest free energy. GAs help optimize RNA folding predictions because of their capacity to handle a large search space and avoid local minima. These methods are still used in recent solutions, as displayed by Shahidul Islam and Rafiqul Islam ( 2022 ).

3.2 Machine learning algorithms predicting RNA structure

A multitude of different machine and deep learning approaches have been proposed for solving RNA structures. This section will briefly introduce the architectures used in the works we cover in this paper.

From classical machine learning methods, a passive-aggressive online learning algorithm (PA) was the earliest used approach. As an online learning algorithm, it operates incrementally, processing one training example at a time. The update rule for the PA algorithm involves adjusting the weight vector \(\textbf{w}\) based on the example \((x_t, y_t)\) at a specific training step t . If the example is predicted correctly, the model remains passive with no change. However, upon a poorly predicted example, the algorithm updates its weight vector aggressively to correct the mistake. The prediction itself happens by multiplying the weight vector by the input \(x_t\) vector, while the loss functions used are mostly hinge loss (for classification) or squared loss (for regression). This method has been introduced in Crammer et al. ( 2006 ).

Two other classical approaches were used by Yonemoto et al. ( 2015 ). First, stochastic context-free grammar (SCFG) is an extension of context-free grammar that adds probabilities to the production rules. It is formally defined as \(G = (V, \alpha , S, R, P_p)\) , where:

V —non-terminal alphabet, in other words, symbols that generate the next set of symbols;

\(\alpha\) —terminal alphabet, in case of RNA that could be its bases (A, C, G, T);

S —a sequence start symbol;

R —a set of rewrite rules called production rules. It specifies how certain symbols from the non-terminal alphabet can produce the next set of symbols;

\(P_p\) —set of probabilities associated with each production rule.

Thus defined algorithm allows for a generative process of building the RNA structure representation. The second classical approach used in this work is the conditional random field (CRF), which enhances the solution by being a discriminative model. It is an undirected probabilistic graphical model representing the conditional probability of a specific sequence of labels Y, given a sequence of observations X. This allows them to capture contextual dependencies among the labels, making them particularly effective for tasks like part-of-speech tagging, named entity recognition, and image segmentation. A detailed introduction for CRFs has been created in Sutton and McCallum ( 2010 ).

The last classical solution was introduced by Su et al. ( 2019 ), and utilized Positive-Unlabeled (PU) Learning algorithm with a logistic regressor. PU Learning algorithm is a type of semi-supervised learning where a machine learning model is trained using a dataset that contains only positive and unlabeled examples, without explicit negative examples. The algorithm typically follows a two-step process, where it first identifies a reliable subset of negative examples from the mixed set U using the information from the positive set P. Then, it iteratively constructs predictive models using these positive and “negative” examples, ultimately selecting the best-performing model from these iterations. This solution was first introduced for building text classifiers in Liu et al. ( 2003 ).

3.3 Deep learning architectures unraveling the RNA data

Long-Short Term Memory (LSTM) architecture was the most commonly used deep learning technique for RNA structure prediction. Its original use came from the natural language processing field and was introduced by Hochreiter and Schmidhuber (Hochreiter and Schmidhuber 1997 ) as a remedy for forgetting long-term information along with vanishing and exploding gradient problems in gradient back-propagation through time. This architecture consists of a cell state that acts like a memory, holding important information from past inputs, and specialized gates that control information flow. The “Forget Gate” decides which information to forget from the cell state, while the “Input Gate” is responsible for adding new information to the cell. Additionally, the “Output Gate” decides what information should be given out at the current time-step. This combination allows the LSTM architecture to not only understand the input, but also to remember long-term relationships between the inputs. A version of these networks, called Bi-LSTM, is also capable of iterating through the input sequence from the end to the beginning.

Another frequently used deep-learning component in the RNA structure prediction problem is the Convolutional Neural Network (CNN). Widely popularized by its use in LeCun et al. ( 1989 ), CNNs, at their core, utilize convolution layers that act as filters to scan the image, identifying patterns and features like edges, lines, and shapes. This happens by sliding the convolution window across the image, looking for specific patterns at each location. Additionally, pooling layers downsample the data, reducing its size and complexity while preserving important features. By stacking these convolutional and pooling layers, the network can learn increasingly complicated patterns. These components are widely used for RNA prediction in two popular architectures—a ResNet (He et al. 2016 ), which groups the convolutional blocks into residual blocks that add their input to the output, and a U-Net (Ronneberger et al. 2015 ) that first downsamples and then upsamples back the data with additional “skip connections” at each depth.

The transformer architecture, introduced in Vaswani et al. ( 2017 ), is characterized by its reliance on attention blocks, allowing the model to focus on specific parts of an input sequence relevant to the current part being processed. An attention block consists of three sets of vectors derived from the input sequence—queries that represent the model’s current focus, keys that represent different parts of the input sequence, and values that carry the actual information from each part of the input sequence. The attention block assigns a score to each possible relationship between a query and a key. This score indicates how significant a specific portion of the input (represented by the value) is to the current focus (represented by the query). Transformers use these attention blocks as an encoder-decoder architecture. The encoder uses the input sequence to generate a contextual representation for each word. In this case, attention allows the encoder to understand how each word in the sequence relates to the others, capturing long-range dependencies. The decoder generates the target sequence based on the encoder’s output. Attention in the decoder allows it to focus on relevant parts of the encoded context while generating each word in the target sequence. Building on top of the transformer architecture, large language models (LLMs) emerged. These models utilize the same ideas, but are defined by very complex structures with billions of parameters, and are trained on massive collections of text data. During this training, LLMs acquire sophisticated statistical models of the data that capture subtle connections between the learned tokens.

Learning on graphs is another interesting deep learning approach. Graph Neural Networks (GNN) provide a powerful approach to deep learning challenges, in which data is organized as nodes (entities) connected by edges (relationships). Using a message-passing paradigm, GNNs take advantage of graphs’ inbuilt connection. Each layer’s nodes collect information from their near neighbors along the edges. This information may include node characteristics as well as messages exchanged between connected nodes. Aggregation can include summing, averaging, or more advanced neural network layers. Based on the aggregated data, each node updates its internal representation to include not just its own qualities but also the contextual effect of its neighbors. There exist a few paradigms on how to train those networks, like Graph Convolutional Networks (GCNs) (Kipf and Welling 2017 ), Graph Attention Networks (GATs) (Veličković et al. 2018 ), or Message Passing Neural Networks (MPNNs) (Gilmer et al. 2017 ), which is more of a general framework.

While it is more of a paradigm than a network design, Deep Reinforcement Learning (DRL) combines the capability of deep neural networks with reinforcement learning, allowing agents to handle complicated decision-making tasks in high-dimensional state spaces. Unlike traditional reinforcement learning, which uses handcrafted features, DRL uses deep neural networks to directly transfer raw sensory inputs (for example, picture pixels from a camera) to value or policy functions. These networks can be taught using approaches like deep Q-learning (Mnih et al. 2015 ) or policy gradient (Sutton et al. 1999 ), where the agent receives scalar rewards for its actions and attempts to maximize expected future rewards. DRL’s power stems from its capacity to discover detailed correlations within the environment via function approximation with deep neural networks.

3.4 Interpretability

While powerful, deep learning methods lack interpretability. They often function as so-called black boxes, making it difficult to understand how they arrive at their predictions (Saeed and Omlin 2023 ). This lack of explanation hinders the ability to validate predictions, improve model design, and gain valuable biological insights, which are often more important than just the predictions (Zhou and Troyanskaya 2015 ).

Methods based on different architectures may be partially interpreted through various approaches. Deeper layers of convolutional neural networks capture higher-level visual constructs and naturally retain spatial information. Visually interpreting a network can be achieved by creating a Class Activation Map (CAM) to get the features from the last convolution layer and measure their activity when predicting the output probabilities. LSTMs and other recurrent neural networks (RNNs) were particularly known for their lack of interpretability. A significant step forward was the invention of the attention mechanism (Vaswani et al. 2017 ), which assigns values corresponding to the importance of the different parts of the time series according to the model. While such approaches explain basic DL architectures, advanced DL models pose several challenges due to their complexity and the complicated connection structure (Choo and Liu 2018 ).

Although there are generic strategies for explaining model outputs, such as LIME (Ribeiro et al. 2016 ) and SHAP (Lundberg and Lee 2017 ), RNA structure prediction could benefit from an approach comparable to in silico mutagenesis. This entails selecting a specific data point X and systematically changing each feature (e.g., modifying individual nucleotides) while maintaining all others fixed and monitoring how the network’s output changes. This is simple to grasp but computationally expensive because the model must be rerun after each mutation. It is critical to understand that these explanation approaches are not causal models which seek to uncover cause-and-effect relationships. While interpreting a model can highlight critical aspects and provide hypotheses, actual causal knowledge requires experimental validation.

3.5 Computational challenges

The accurate prediction of RNA structure is a severe computational challenge. Classical structure prediction approaches that use dynamic programming algorithms, such as the Nussinov algorithm (Nussinov and Jacobson 1980 ) or Zucker’s algorithm (Zuker 1989 ) have a time complexity of \(O(n^3)\) , where n is the RNA sequence length. This cubic complexity results from evaluating all possible base pairings inside the sequence using dynamic programming techniques. Then, Frid and Gusfield ( 2010 ) employed the Four-Russians Speedup approach, which reduced the complexity to \(O\big (\frac{n^3}{\log n}\big )\) . However, the problem becomes considerably more prominent when the system must also predict pseudoknots. In particular, Rivas and Eddy ( 1999 ) developed such a solution with the worst time complexity of \(O(n^6)\) .

While dynamic programming has a well-defined complexity, other techniques, such as the Greedy Randomized Adaptive Search Procedure (GRASP), typically include iterative procedures with evaluations at each step. The computational complexity of Genetic Algorithms (GAs) is more complicated than that of other algorithms since it is determined by many factors, including population size ( p ), number of generations ( g ), and fitness function assessment ( f ( x )), which together give \(O(p \cdot g \cdot f(x))\) complexity. The last part of the equation represents the complexity of evaluating the “fitness” of a single individual (solution) within the population. The fitness function’s complexity can vary depending on the specific problem. In RNA structure prediction, the fitness function typically assesses how well a predicted structure folds using parameters such as base-pairing probabilities and minimum free energy. This function may be more or less computationally expensive at the cost of overall accuracy, so determining the total complexity is impossible.

Deep learning models for RNA structure prediction offer high accuracy but come with their own computational costs. The complexity of the model is determined by its architecture (number of layers and neurons), training method (epochs and batch size), optimization algorithm, and the activation functions used. This complexity may exceed classic methods like dynamic programming due to the required training. However, researchers continually develop strategies for more efficient network topologies, hardware advancements, and optimized training algorithms to overcome this difficulty. Furthermore, after a model has been trained, the computational cost for inference is significantly decreased, as is the time complexity, thanks to high parallelism and the usage of GPUs. Deriving a standard “big O” notation formula is not really feasible for deep learning algorithms, as their end of execution for training depends on the algorithm’s convergence rather than the raw number of data. Modern deep learning architectures’ complexity is commonly evaluated by the total number of parameters of the network or the number of mathematical operations involved. A more in-depth investigation of deep learning models complexity has been presented in Hu et al. ( 2021 ).

4 Overview of the studies

This section gives an overview of the works found for this SLR. The description has been divided into four sections—an aggregated analysis of the studies (Sect.  4.1 ), works on secondary structure prediction (Sect.  4.2 ), works on tertiary structure prediction and scoring (Sect.  4.3 ), and methods availability (Sect.  4.4 ).

4.1 Aggregated analysis

The primary search, which encompassed papers published before 30.09.2023, resulted in 52 works found using the Web of Science and 243 works found via Google Scholar. The subsequent filtering steps were divided into three steps. The initial selection yielded 74 works by removing duplicated entries and evaluating them based on metadata exclusion criteria. The intermediate selection focused on removing secondary studies and evaluating substantive exclusion criteria by reading titles, keywords, and abstracts of the works, which resulted in 44 papers. The final selection consisted of reading and evaluating the studies using the substantive exclusion criteria, which provided the final list of 33 works analyzed in this review. The full description of eligibility criteria and search strategy is available as supplementary material (Online Resource 1). The supplement enlists trusted publishing sources, outlines the search string criteria, describes the works selection procedure and data extraction strategy, and presents the exclusion criteria. The selection pipeline is presented in Fig.  4 .

The paper selection pipeline displaying the process of searching and filtering works relevant for this review. Each step of the pipeline shows the used filtering procedures, the relevant exclusion criteria, and the number of resulting works in gray boxes (see Supplementary material for a full description of eligibility criteria and search strategy)

Various approaches and methodologies have been used to utilize machine learning for the prediction of RNA structures, including convolutional neural networks, recurrent neural networks, and graph-based algorithms. As seen in Fig.  5 , the machine learning methods used have completely shifted from classical ones (such as conditional random fields (CRF) and passive-aggressive online learning) towards various deep learning-based methodologies, with recurrent neural networks (RNN) and convolutional neural networks (CNN) dominating the field in recent years.

Figure  6 displays the cumulative sums of articles published over the years, split by the problem being solved—secondary or tertiary structure prediction and scoring. Aside from the significant overall increase in the number of works published, methods trying to predict the actual 3D structure started being developed. This has become possible due to higher volumes of data available, as well as due to the increase in computing power and the developments in the field of machine learning itself.

Analyzed works often share the datasets used for training and validation. Throughout the literature, five datasets were identified as being commonly used. The most reoccurring dataset, bpRNA (Danaee et al. 2018 ), contains descriptions of loops, stems, and pseudoknots, along with the positions, sequence, and flanking base pairs of every structural feature. It contains over 100,000 single-molecules with their secondary structures, and at the time, it was approximately 20 times bigger than already existing datasets. The second most commonly used dataset, RNAStralign (Tan et al. 2017 ), contains over 37,000 structures divided into homologous families based on the classifications in the source databases. RNAStralign is a structure database and a multiple sequence alignment database, allowing broader analysis of the dependencies within structures. Other datasets include ArchiveII (Sloma and Mathews 2016 ), commonly used for benchmarking due to covering multiple RNA families, RNAStrand (Andronescu et al. 2008 ), which adds a user-friendly webserver for searching and analyzing structures, and Rfam (Griffiths-Jones et al. 2005 ) that provides a comprehensive resource for understanding and classifying ncRNAs based on their sequence, structure, and function. The availability of these datasets and the underlying data volume allowed for using deep learning methods for secondary structure prediction. Additional information, such as multiple sequence alignment, allows for a more profound and unified understanding of relations within the structures. However, the datasets come with certain limitations, the biggest one being the RNA family coverage. While datasets like bpRNA and ArchiveII contain ten different families, most structures in all the datasets are from ribosomal RNAs (rRNAs) and transfer RNAs (tRNAs). This limits the potential of machine learning algorithms to discover the RNA structure space. Moreover, the datasets may contain redundant structures, which, when unfiltered, may yield false results on trained methods.

A cumulative sum graph illustrating the number of papers used for RNA structure prediction grouped by the specific machine learning methods, showing the development over time. The y -axis denotes the total number of publications accumulated, while the x -axis denotes the year. The total number exceeds the number of papers found, as some solutions may use several machine learning methods

A cumulative sum graph illustrating the number of papers used for RNA structure prediction grouped by the problem tackled, showing the development over time. The y -axis denotes the total number of publications accumulated, while the x -axis denotes the year

4.2 Secondary structure prediction

The quest for accurate prediction of the secondary structure is still dominating the field, since the literature search provided a total of 28 works (out of 33 analyzed works) focused on utilizing machine learning for this problem. The research found has been summarized in Table 1 ; however, the reader should be aware that the findings of these articles cannot be directly compared due to variations in their test datasets and methodologies.

The paper by Zakov et al. ( 2011 ) was one of the initial studies that was found to achieve promising results by using ML methods. It proposed using a modified version of Collins ( 2002 ) discriminative structured-prediction learning framework based on Hidden Markov Models (HMM), which was primarily used for natural language processing (NLP). The modification was implemented by coupling a passive-aggressive online learning algorithm proposed by Crammer et al. ( 2006 ), whose function was an appropriate weight update for cost-sensitive learning with structured outputs. The main reason for using these algorithms was their ability to adapt well to large feature sets, as this was the main obstacle to RNA structure prediction at the time. The models created by Zakov’s team induced up to 205 thousand features (70 thousand after ridding of zero-valued parameters after training). An important point to mention is that the secondary structure prediction problem defined by the team was omitting pseudoknots and non-canonical base pairs in the task. The dataset used by Zakov’s team came from the work of Andronescu et al. ( 2010 ) and contained 3245 distinct structures of lengths between 10 and 700 nucleotides, which in turn was based on RNA STRAND dataset (Andronescu et al. 2008 ). On the development set, the best-obtained results were sensitivity at 83. 8%, precision (referred to as PPV) at 83.0%, and F 1 -score equal to 83.2%, which at the time of publication were state-of-the-art results.

A more recent example of the use of classical machine learning methods can be found in Su et al. ( 2019 ). It introduced a Positive-Unlabelled (PU) data-driven framework called ENTRNA. The team considers not only the sequence itself but also expands the input features by free energy, sequence, and structural motifs, and a new feature called Sequence Segment Entropy (SSE), which measures the diversity of RNA sequences. PU learning requires two datasets: a positive labeled set P and a mixed, unlabeled set U . The learning process generally involves two steps. First, an algorithm is trained to identify negative examples in the set U based on knowing positive examples from the set P . Having a true set P and self-labeled negative examples from the set U , the second step is to build complete predictive models iteratively and choose the best-performing one. In their work, Su’s team generated the unlabeled set U by computationally creating synthetic RNAs. Positive data was prepared as three separate datasets for algorithm training, cross-validation, and blind testing. For each secondary structure, 100 sequences were generated by three different RNA inverse folding algorithms—RNAinverse (Hofacker et al. 1994 ), incaRNAtion (Reinharz et al. 2013 ), and antaRNA (Kleinkauf et al. 2015 ). Two separate experiments were conducted, one on pseudoknot-free structures and the other on pseudoknotted RNAs. The underlying classifier of ENTRNA is a logistic regressor that predicts the foldability of molecules using 11 features (different for the two experiments). The researchers report that the sensitivity of their solution is equal to 80.6% for the first experiment and 71.0% for the second experiment (on test datasets).

Deep learning methods for predicting the secondary structure of RNA have seen the highest increase in popularity in recent years, partially due to previously limited knowledge, hardware, and lack of data. Wang et al. ( 2019a ) introduced DMfold, which outperformed previous state-of-the-art machine learning-based algorithms. The core of the solution, called the Prediction Unit (PU), is a sequence-to-sequence model based on a bidirectional LSTM network used as the encoder with fully connected layers used as the decoder. As a second step of the solution, the authors have introduced the so-called Correction Unit (CU), which reduces the errors produced by the PU. The final sequence is a dot-bracket style notation of the secondary RNA structure. The data used for training and testing comes from the public database of Mathews lab, ArchiveII (Sloma and Mathews 2016 ), comprising 2975 known RNA sequences and structure pairs, and the problem definition included pseudoknots. The prediction results of the test set were divided by RNA families, in terms of F 1 -score, the method achieved 93.7% for tRNAs, 92.7% for 5sRNAs, 70.6% for tmRNAs, and 61.9% for RNaseP, which at the time of publication exceeded the results of previous methods.

In the same year, Zhang et al. ( 2019 ) proposed a solution called CDPfold based on a convolutional neural network (CNN) paired with Dynamic Programming (DP). The network consists of three convolution layers, each utilizing sixteen 3 \(\times\) 3 convolution kernels. The output layer is then mapped to three labels of the dot-bracket representation using two fully connected layers. The data comes from the public database of Mathews lab, NNDB (Turner and Mathews 2009 ), and is first represented as n \(\times\) n matrix, where n is the length of the RNA sequence. The matrix values are set according to the number of hydrogen bonds between bases, that is, 2 for paired A and U, 3 for paired G and C, x \((0< x < 2)\) for the wobble pair, and 0 otherwise. Due to the size of the convolutions and RNA sequences, the matrix is split into windows of length d , that is, matrices of size d \(\times\) n . Additionally, the resulting dot-bracket sequence is then corrected using a probability sum algorithm. The prediction results of the test set were divided by RNA families, and in terms of F 1 -score, the method achieved 90.5% for tRNAs, 91.1% for 5sRNAs, and 82.3% for srpRNAs, which makes the method slightly worse performing than that of Wang’s team.

In their paper, Fu et al. ( 2022 ) proposed UFold—a method utilizing a different CNN architecture to solve secondary structure prediction problems. Due to the variable length of the RNA sequences, the team has decided to use a U-Net architecture, which is a fully convolutional neural network (Ronneberger et al. 2015 ). The most important feature of this architecture is the ability to produce an output matrix of the same size as the input matrix without setting a fixed input length. Fu and the team used this fact to input the model with 17 channels of size \(n \times n\) , where 16 channels come for a Kronecker product of \(n \times 4\) representation of the sequence (each base in a separate row), with the 17th channel added to overcome the sparsity as the base pairing possibilities used in CDPfold. The output is a one-channel n \(\times\) n matrix containing probabilities of pairing between each pair of bases and is scored against the ground truth of the paired bases. The algorithm was trained and tested on a few datasets, including RNAStralign (Tan et al. 2017 ), ArchiveII (Sloma and Mathews 2016 ), and bpRNA-1 m (Danaee et al. 2018 ), among others. The model achieved, in terms of F 1 -score, between 65.4% on bpRNA-1 m to 91% on the ArchiveII test set. It is worth noting that both results, which may seem lower than previously described works, are measured on more extensive and complicated datasets, reaching results higher than previous state-of-the-art methods, such as MXfold2 or SPOT-RNA.

Among the models based on convolutional neural networks, the highest reported scores were achieved by Booy et al. ( 2022 ). The team utilized a ResNet architecture (He et al. 2016 ) that consists of N residual blocks containing convolution layers, batch normalization, and an activation function (leaky ReLU in the case of Booy’s model). Additionally, within each residual block, the input size of the matrix does not change its shape, which allows for so-called “skip connections”—a concatenation of the input to the convolution blocks to their output. The secondary structure target was therefore formulated as a binary \(L \times L\) matrix, where L is the length of an RNA sequence, and each output cell of the matrix represents binary information on whether two bases are paired. The input to the network is represented as an \(L \times L \times 8\) tensor, where the additional dimension is a one-hot representation of eight potential relations between bases. Six channels show possible combinations of bases (A, U), (U, A), (U, G), (G, U), (G, C), (C, G). One channel indicates the same nucleotide in a sequence, or in other words, is a diagonal line for every index \(i = j\) , and one channel represents pairs of bases that cannot form a bond due to their short distance, non-valid base combination, or other constraints. Due to the visual nature of this representation, an overview of Booy’s team solution is shown in Fig.  7 .

A general illustration of the solution created by Booy et al. ( 2022 ). An input sequence is converted into a defined representation that a CNN-based prediction model can process. The output target matrix displays a contact map that is post-processed for obtaining the final secondary structure. (Own work based on Booy et al. ( 2022 ))

Different standard datasets were used to train and evaluate the architecture. The first model was trained on 80% of the RNAStralign dataset (Tan et al. 2017 ), with a randomly chosen secondary structure when more than one was available for a given RNA chain. The other 20% of the data was split into validation and test sets with stratification over RNA families. This enabled a fair comparison of the results achieved with Chen et al. ( 2020 ). The second dataset utilized in this work was AchiveII (Sloma and Mathews 2016 ), which was additionally used for the first trained model for evaluation. Other datasets used in this work included bpRNA (Danaee et al. 2018 ) and bpRNA-new (Sato et al. 2021 ), which was used for family-wise cross-validation. The best results on these datasets are shown in detail in Table 2 .

As seen, most recent works have shifted their focus towards using convolutional neural networks. One of the exceptions is the work of Castro et al. ( 2020 ) that utilized graph neural networks in combination with variational autoencoders. The main problem to solve in this approach was to find an embedding given an initial set of graphs that satisfies the given properties: faithfulness—graphs near each other (in terms of graph edit distance) should be close to each other in the embedding space; smoothness—the embedding should be smooth in terms of a real-valued meta-property \(M = \{m_1, m_2,\) ... \(, m_n\}\) , where \(m_i \in \mathbb {R} ^n\) ; invertibility—new graphs should be possible to be generated from interpolated points in the embedded space. To satisfy these conditions, the team proposed a framework based on geometric scattering obtained from a set of Diracs that are further processed by a graph autoencoder. To train the network for this approach, four RNA secondary structure datasets were generated via ViennaRNA’s RNAsubopt program (Lorenz et al. 2011 ), and the evaluation was based on Gibbs free energy. The testing was then carried out on four specific sequences that were identified to have specific structures, namely SEQ3 (an artificial RNA sequence of 32 nucleotides designed to be bistable); SEQ4 (similarly to SEQ3, also an artificial RNA sequence of 32 nucleotides that is bistable); HIVTAR (an ensemble generated from the transactivation response element (TAR) RNA of HIV consisting of 61 nucleotides); TWBOWN (a designed bistable sequence of 72 nucleotides that was later described as a “faulty riboswitch” with three or more dominant states). Some features of the embeddings and structures generated by the model were tested, such as energy prediction, energy smoothness, and reconstruction error. For the secondary structure prediction problem, the reconstruction score can be utilized to measure the model performance. The score is provided as a mean squared error (MSE) of the adjacency matrices generated and is shown in Table 3 .

Another architecture that has gained traction in recent years, a large language model (LLM), was tested by Wang et al. ( 2023a ). The team had to gather, construct, align, and refine a massive dataset for the pre-training. The data was collected from diverse sources such as RNAcentral (Consortium 2020 ), NCBI (Sayers et al. 2020 ), and Genome Warehouse (Chen et al. 2021 ), which resulted in approximately 1 billion RNA sequences. These sequences were aligned to a standardized DNA alphabet, analyzed and filtered using statistical analyses, and clustered using the mmseqs2 algorithm (Steinegger and Söding 2017 ). Their model, called Uni-RNA, incorporates advanced deep learning techniques such as rotary embedding, flash attention, and fused layernorm to optimize performance. Pre-training utilized a masked nucleic acid modeling framework, enabling Uni-RNA to capture robust representations of RNA’s biological structures. The training of Uni-RNA required significant computational resources, including 128 A100 GPUs. The model was scaled to various sizes to address different downstream tasks, to which the models were subsequently fine-tuned. Wang’s solution was evaluated on tasks such as splice site identification, non-coding RNA classification, and secondary structure prediction. In their tests, the model achieved 82.1% in terms of F 1 -score, 89.4% in terms of precision, and 80.1% in terms of recall, which yielded state-of-the-art results compared to other methods. The authors also mentioned using the model to predict contact maps and RNAs’ tertiary structures. However, no detailed comparison to other methods was performed.

4.3 Tertiary structure prediction and scoring

As a result of secondary interactions, the RNA molecules fold onto themselves, creating three-dimensional conformations. Therefore, the tertiary structure refers to defining spatial coordinates of atoms in the RNA molecule and spatial relationships between them (tertiary interaction). In silico prediction and scoring of these structures is still an ongoing challenge, as the problem complexity is far surpassing the secondary structure prediction, bound together with much smaller volumes of data (crystallographically solved structures). The literature search provided a total of 5 works (out of 33 analyzed) focused on using machine learning for this problem (Table 5 ).

Li et al. ( 2018 ) proposed a complex neural network-based solution called RNA3DCNN to evaluate the RNA structure using 3D convolutions. The VGG-like network (Simonyan and Zisserman 2014 ) inputs a 32 \(\times\) 32 \(\times\) 32 tensor, where each value is a dimensional representation of the distance of 1Å. The network follows with convolutional and maxpooling layers, ending with one fully connected layer to output a single value, which is described as “unfitness score”. To allow for comparison with other methods, the datasets used for testing purposes come from different sources. Test dataset I was introduced in the RASP paper (Capriotti et al. 2011 ), with 85 non-redundant RNAs and 500 structural decoys for each sample. Test dataset II came from the KB paper (Bernauer et al. 2011 ), which was produced using the normal-mode perturbation approach for 15 RNAs and position-restrained dynamics and REMD simulations for 5 RNAs. Depending on the method, between 490 and 3500 decoys were generated for each RNA structure. Test dataset III comes from RNA-Puzzles rounds I–III consisting of 18 target RNAs. The training dataset was constructed using only non-redundant RNA structures obtained from the NDB website . The team used Enrichment Score (ES) as the main metric for evaluation (Bernauer et al. 2011 ; Wang et al. 2015 ). The results achieved vary depending on the dataset, where, in some cases, the model outperformed previously known methods, as shown in Table  4 .

Wang et al. ( 2019b ) proposed a scoring function based on multi-layer neural networks. Two networks (called NET1 and NET2) were trained on differently defined input tensors. However, their architectures are similar in that they both contain an input layer (of sizes 291 and 5524, respectively), one hidden layer (of sizes 30 and 10, respectively), and a single node for their outputs. The team built train/validation/test datasets using the NDB website , obtaining non-redundant structures of 462 RNAs with lengths varying between 8 and more than 200 nucleotides that excluded complexes with other molecules and RNAs with non-standard nucleotides. Each RNA was paired with 300 decoys generated using molecular dynamics simulations, and the data was split by the real sequences into 322/70/70 RNAs, respectively. The loss function used for training was the mean squared error measured between an RMSD sample difference from the native structure and the score given by the network. The results showed that the trained networks were able to correctly predict 39 and 49 of the 70 structures closest to the native structure for NET1 and NET2, respectively. In contrast, the Ribonucleic Acids Statistical Potential (RASP) based on all-atom knowledge (Capriotti et al. 2011 ) correctly predicted 26 of 70 structures.

Townshend et al. ( 2021 ) introduced a scoring method called Atomic Rotationally Equivariant Scorer (ARES) that achieved excellent results in scoring RNA structures in proximity to native ones. The solution does not incorporate any RNA-specific information in its predictions. Instead, the team uses only 3D coordinates and the chemical element type of each atom in the structure. The underlying machine learning components of the solution can be defined as a graph neural network, utilizing graph convolution layers, rotational and translational equivariance, and other dense layers. The specific design of the layers is built on recent techniques, in particular, the tensor field networks (Thomas et al. 2018 ) and the PAUL method (Eismann et al. 2021 ). The solution first aims at identifying local structural motifs by computing several features for each atom based on the geometry of surrounding atoms and features computed by previous layers. The remaining layers then aggregate generated information across all atoms, which allows for predicting the accuracy of the whole structural model. Figure  8 presents an overview of the solutions.

A general illustration of the solution created by Townshend et al. ( 2021 ). A structural model of a given RNA is first converted into a graph on which the ARES solution learns features by repeated information sharing between neighboring atoms’ positions and element types. The features are then averaged across all atoms and fed into an artificial neural network (ANN). The final output is the predicted RMSD from a native structure. (Own work based on Townshend et al. ( 2021 ))

ARES was trained using only 18 RNA structures solved before 2007, ranging between 17 and 47 nucleotides in chain length, with a median of 26 (Das and Baker 2007 ). For each RNA, 1000 structural models were generated using the Rosetta FARFAR2 sampling method (Watkins et al. 2020 ) without using the original structure. The solution’s parameters were then optimized to match the RMSD of each generated model to the corresponding structure. For testing purposes, the team used a benchmark consisting of all RNAs that were included in the RNA-Puzzles structure prediction challenge, with experimentally determined structures published between 2010 and 2017 (Miao et al. 2020 ) with at least 1500 structural models generated for each RNA. One of the challenges for ARES is that these structures comprise a much larger number of nucleotides than the training set structures—between 112 and 230 nucleotides, with a median of 152.5. The results of the model were compared with three other state-of-the-art scoring functions, namely Rosetta (ver. 2020) (Watkins et al. 2020 ), Ribonucleic Acids Statistical Potential (RASP) (Capriotti et al. 2011 ), and 3dRNAscore (Wang et al. 2015 ). For each RNA in the described benchmark set, the team determined the rank of the best-scoring near-native structural model. This can be thought of as searching through a list ranked by scoring solutions to find a nearly native structure (RMSD \(< 2\) Å). Across the dataset, the mean rank of the best-scoring near-native model is 3.6 for ARES, compared to 73.0, 26.4, and 127.7 for Rosetta, RASP, and 3dRNAscore, respectively. Additionally, the team has chosen four RNAs from the recent rounds of RNA-Puzzles (whose structures are now in the Protein Data Bank with IDs 6OL3, 6PMO, 6POM, and 6UFM), generated sets of candidates with FARFAR2 and submitted the best structures as solutions to the puzzle. The comparison of this method with the best previous submissions is shown in Table  6 .

Yet another approach was proposed by Deng et al. ( 2022 ) with a graph convolutional network solution for tertiary structure scoring. This solution first represents the RNA structure as a graph, although, due to varying lengths of RNA chains, the whole structure is split into so-called “local environments”. Each local environment is defined by a central nucleotide at the position i along the chain and its neighboring nucleotides (nucleotides within the threshold of 14Å). For an RNA chain of length \(N_s\) , this creates \(N_s\) environments represented by \(N_s\) graphs. Each atom in such a graph was represented as a one-hot vector of length \(N_t\) , where \(N_t\) is the total number of atom types (54, based on the AMBER99SB force field). Graph edges were created by connecting the fourteen nearest atoms in space and contained five features, namely one for the distance between atoms, three for direction features, and a binary value of 0 or 1, depending on the presence of chemical bonds between atoms. The data used by the team was built based on the non-redundant set of RNA 3D Hub (Leontis and Zirbel 2012 ), from which complexes of RNA with other molecules were removed. Additionally, RNA chains shorter than eight nucleotides were also removed, and the remaining 610 RNAs were split into training, validation, and testing datasets with the use of the infernal program (Nawrocki and Eddy 2013 ) to ensure that there is no overlap in RFAM families between the testing and training datasets. The network architecture consisted of five serially connected graph convolution layers with adopted residual modules and skip connections to solve the vanishing gradient problem (Li et al. 2019 , 2021 ). The output of those five layers went to a \(1 \times 1\) kernel convolution layer, followed by a MaxPooling layer, and then produced a single score using a fully connected network. The final score is a scalar, indicating the quality of the input graph. During training, this scalar measured the difference between the input structure and the experimental structure in terms of RMSD. Therefore, the inference output can be viewed as a predicted RMSD score between the input structure and the unknown to the network’s native structure. The architecture of the solution has been shown in Fig.  9 .

A general illustration of the solution created by Deng et al. ( 2022 ). A part of a structural model (local neighborhood) of given RNA is first converted into a graph and fed through multiple graph convolution layers (GCL). A schema for the inner working of a graph convolution layer used is presented on the right. The final fully connected network predicts an RMSD of an input structure to the native one. (Own work based on Deng et al. ( 2022 ))

The first test included evaluating the quality of the structures in the test dataset, which included 92 native RNA structures associated with 200 decoys for each. The team used both the Top-1 and Top-5 criteria to reflect whether the experimental RNA was ranked first or among the best five, respectively. The results were compared with four other popular solutions (shown in Table 7 )—namely RASP (Capriotti et al. 2011 ), Rosetta (Watkins et al. 2020 ), 3dRNAscore (Wang et al. 2015 ), and rsRNASP (Tan et al. 2022 ). It is worth acknowledging that the RNAGCN model achieved the highest average score for enrichment score (ES) and Pearson correlation coefficient (PCC) for near-native structure (<4Å)—scores that indicate the strength of correlations between ground truth and prediction.

Pearce et al. ( 2022b ) proposed in 2022 a de novo method called DeepFoldRNA that uses geometric potentials from deep learning. The body of the network architecture is built with 48 multiple sequence alignment (MSA) Transformer blocks utilizing self-attention layers. The output embedding is then processed by four Sequence Transformer blocks, after which the whole process is repeated for four cycles. The last step is predicting the distance and orientation restraints from the final pair representation, and the backbone angles prediction from a linear projection of the sequence representation. To obtain the final RNA model, those restraints are converted into a negative-log likelihood potential to guide L-BFGS simulations. The solution concept has been visualized in Fig.  10 .

The model was trained using 2986 RNA chains gathered from the PDB, which were sure to be non-redundant to the 122 test RNAs. For each chain, an MSA representation is generated using rMSA (Zhang et al. 2023 ), and fed to the network along with the paired sequence. From these structures, some labeled features were extracted, including the native C4′, N1/N9, and P distance maps, the inter-residue \(\Omega\) and \(\lambda\) orientations, and the backbone \(\eta\) and \(\theta\) pseudo-torsion angles. The trained model was tested on two datasets—one consisting of 105 non-redundant RNAs from 32 Rfam families, and the second containing 17 targets from RNA-Puzzles. On the first benchmark, the method generated models with an average RMSD of 2.68Å and a TM-score of 0.757. In the second benchmark, DeepFoldRNA has produced models of higher quality than the best models submitted by the community for 15 of 17 cases.

A general illustration of the solution created by Pearce et al. ( 2022b ). For the input RNA sequence, a multiple sequence alignment (MSA) is generated via rMSA (Zhang et al. 2023 ) and fed into an MSA transformer together with a secondary structure predicted by PETfold (Seemann et al. 2008 ). The outputs of MSA blocks are then fed into a sequence transformer that predicts the distance and orientation maps together with torsion angles. Using L-BFGS optimization on the negative log-likelihood potentials, the method predicts the final RNA model. (Own work based on Pearce et al. ( 2022b ))

4.4 Availability

Most described works can be accessed via model checkpoints shared with the community or via appropriate webservers. Additionally, most works either use publicly accessible datasets or share their data. A handful of works (Lu et al. 2019 ; Zhang et al. 2019 ; Willmott et al. 2020 ; Calonaci et al. 2020 ; Castro et al. 2020 ; Deng et al. 2022 ; Fei et al. 2022 ) made only their code public, which makes it theoretically possible to train own models on available data. Few works do not share or mention their models and codebases (Wu et al. 2018 ; Wang et al. 2019b ; Quan et al. 2020 ; Wang et al. 2020 , 2023a ). One research provides information on accessing its codebase and models; however, the provided hyperlinks no longer work (Zakov et al. 2011 ). In another work (Yonemoto et al. 2015 ), the authors do not publicly share their solution; yet, they mention that the software can be accessed on request.

5 Results and discussion

This section summarizes the results, advantages and disadvantages, and methods used in the reviewed works. In Sect.  5.1 , we provide an aggregated summary of the results found. In Sect.  5.2 , we discuss the advantages and disadvantages of methods used, alongside comparing certain works.

5.1 Results overview

The works analyzed in this review show some trends and conventions used in research using machine learning-based algorithms for RNA structure predictions. The works on secondary structure prediction display a clear shift towards deep learning algorithms over time. The works of Zakov et al. ( 2011 ), Yonemoto et al. ( 2015 ), and Su et al. ( 2019 ) are the only ones to use classical machine learning approaches. The newer works converge on using one of two deep learning paradigms, either convolutional neural networks or recurrent neural networks, yielding state-of-the-art results. Some recent works, such as Zhao et al. ( 2023 ) or Qiu ( 2023 ), utilize both paradigms to enhance the results further. As for the prediction and scoring of tertiary structures, in addition to using CNNs and MLP, approaches based on graph neural networks bring promising results. However, the newest work uses the transformer architecture with input consisting of the MSA and the RNA secondary structure.

Almost all secondary works focus on canonical and Watson–Crick base pairs as their prediction goal. Most works also include pseudoknot pairings, which makes the prediction problem more complex, thus initially scoring worse than the pseudoknot-free solutions. Only a few works (Zhao et al. 2023 ; Booy et al. 2022 ; Mao et al. 2022 ; Singh et al. 2019 , 2021 , 2022 ) consider multiplets. In the tertiary structure section, only the work of Pearce et al. ( 2022b ) is focused on predicting the actual RNA structure. Other works concentrate on creating a DL-based scoring function. However, pairing such a function with thousands of RNAs generated using other methods (like Rosetta’s FARFAR2 sampling in the case of Townshend et al. ( 2021 )) can also tackle the prediction problem.

Regarding the metrics used in secondary structure prediction research, as it is a classification problem, the articles evaluate implemented solutions using standard classification metrics such as accuracy or F 1 -score. Some works also report Matthew’s correlation coefficient (MCC), which measures the quality of binary classification rather than the overall performance, thus providing a better overview of all possible binary classification results. Due to the varying nature of the prediction goals and metrics used, it is impossible to directly compare the results between solutions. However, the results obtained in the analyzed articles show improved prediction over the years. Depending on the datasets used to test the models, deep learning methods have achieved between 0.62 and 0.97 in terms of F 1 -score for the secondary structure prediction. The commonly used metric in the tertiary structure prediction and scoring problem is the root-mean-square deviation (RMSD), which reflects the average distance between the atoms. In the case of structure prediction, it is used to calculate the error of the predicted atomic positions relative to the true structure. In the scoring problem, it can be used as the model’s prediction of the difference between the provided structure and the theoretical true structure.

It is also worth noting that most works use common datasets to train and evaluate the results. Some of those datasets come from teams that have first aggregated them for use in their research (such as Tan et al. ( 2017 )). Others are created to provide a standard benchmark for the algorithms (such as Andronescu et al. ( 2008 ) or RNA-Puzzles).

5.2 Discussion

Machine learning has become prominent in RNA structure prediction. The multitude of approaches and architectures used in the works yield various outcomes, even while utilizing the same core mechanisms. This section compares and evaluates those approaches through the lens of the results, complexity, and implementation details.

Classical machine learning methods generally have a computational advantage over deep learning architectures. One of the first works using machine learning as a solution for secondary structure prediction was Zakov’s team approach (Zakov et al. 2011 ). It opened pathways for new research in this domain by demonstrating the feasibility of using machine learning. At the time of the publication, pseudoknots and non-canonical base pairs were still beyond reach. While the high number of features used might have caused the curse of dimensionality, the experimental results yielded state-of-the-art results. The much later approach proposed in Su’s work (Su et al. 2019 ) may have a great computational advantage over previous works as well as deep learning methods by operating on only 11 input features and using a lightweight logistic regressor at its core. However, the solution may suffer in its training performance and results because of the need for synthetic RNA data.

Deep learning-based solutions for the secondary structure prediction most commonly used LSTM architectures working on the nucleotide chains, CNN architectures taking specifically crafted matrices of nuclei dependencies, or a combination of both of these architectures. However, there is no clear winner in the case of the underlying DL mechanisms, as the results varied greatly. For example, Wang’s team DMfold (Wang et al. 2019a ), which uses a standard bi-directional LSTM architecture coupled with their own correction unit for predictions, achieves higher scores than another LSTM-based work by Lu’s team, published in the same year (Lu et al. 2019 ). There may be a variety of reasons for the differences in results of these works, like the datasets used, own modules, etc. However, the main difference between them lies in the output of the LSTM modules. While Wang’s solution is a sequence-to-sequence model outputting the secondary structure in dot-bracket notation, Lu’s method predicts pairing probabilities between nucleotides that are further optimized by the energy filter.

In recent years, however, CNNs have gained traction. Most works use either fully CNN-based architectures or CNNs coupled with LSTMs, but the use of fully LSTM-based architectures has almost vanished. This may be caused by two factors—lower computational cost of architectures like ResNet or U-Net, and by achieving generally better results, perhaps through the greater potential for spatial features extraction. One such work was presented by Booy’s team (Booy et al. 2022 ), which tackled all prediction goals (canonical and Watson–Crick pairs, pseudoknots, and multiplets) while achieving state-of-the-art results. This method uses a standard ResNet architecture that takes as its input a specifically crafted matrix containing potential possible pairings between nucleotides. Similarly, other best-performing convolutional methods (Fu et al. ( 2022 ), Chen and Chan ( 2023 )) also include the possible pairings between the bases, either as a simple indication in the input matrices by putting a specific value at the possible pairing position, or by appending the input with pairing probability matrix. This information alone greatly enhances the prediction possibilities of the networks.

Against the convention of using CNNs and LSTMs, some works explored other deep learning techniques. In particular, Castro et al. ( 2020 ) explored deep graph embeddings in RNA structures. This work, however, does not solely focus on predicting the correct secondary structures but displays a potential for a generative process manipulated by desired properties in a low-dimensional space. By embedding the graphs representing RNA secondary structures in the Euclidean space, the team can predict the folding landscape of given RNA molecules. Tackling a similar problem, the work of Mao and Xiao ( 2021 ) aims to learn the fastest RNA folding path by using deep reinforcement learning. The network begins with an open RNA strand and aims to fold it into its native structure. Through a combination of value and policy neural networks, along with Monte Carlo tree search, the algorithm selects base pairs step by step. By learning from reward signals generated by the comparison between predicted and native structures, the solution adjusts its strategy episode by episode. The sequence of selected base pairs at each step represents the predicted folding path. Chen et al. ( 2020 ), Wang et al. ( 2020 ), Fei et al. ( 2022 ) incorporated the transformer architecture in their solutions. The transformer networks are fed the sequence as one-hot-encoded vectors, weighted and positionally encoded vectors, or embedded inputs. The transformer architecture acts as an information encoder, which is later decoded into a pairing matrix by CNN networks (like U-Net). The internal attention mechanism of transformer networks is a promising solution for encoding the importance of nuclei interactions. These methods achieve good results, however, simpler architectures display higher state-of-the-art performance. Coupling this with the high computational power required for training transformer networks answers why only a few works explore them further.

On the topic of transformer networks, it is worth mentioning the work of Wang et al. ( 2023a ), which introduces a large language model for secondary structure prediction, along with other capabilities. LLMs use transformer blocks as part of their overall architecture, making them computationally expensive. This, however, is justified by the flexibility and versatility of these models. Utilizing this pre-train model paradigm of LLMs seems to embed RNA information that is further applicable to many downstream tasks, as the solution achieved state-of-the-art results in most of them. In the end, the trained model does not outperform other solutions for the secondary structure prediction published in the same year. However, it displays a deep and general understanding of RNA properties, even for more niche types like ncRNAs.

For the tertiary structure, only one work tried to tackle the structure prediction itself. The solution proposed by Pearce et al. ( 2022b ) uses the transformer architecture in two configurations—an MSA Transformer consisting of 48 blocks and a Sequence Transformer consisting of 4 blocks. The architecture is similar to the initial AlphaFold solution (Jumper et al. 2021 ), which implies a computationally heavy model. That, in turn, increases the potential inference time, limits the possibility of reproducing and validating the solution, and expands the overall use costs. However, this computational power required yielded high prediction results, beat other established models for most tested cases, and outperformed classical methods based on Monte Carlo sampling. Recent works on the tertiary structure scoring use mainly Geometric Deep Learning, displaying its potential in the field as exemplified by Townshend et al. ( 2021 ). By smart incorporation of spacial properties like rotational and translational equivariance, the model achieves superb results without incorporating any domain-specific knowledge. Moreover, this was achieved by the model being trained only using 18 RNA structures, displaying the strength of graph neural networks.

6 Conclusions

Over the years, machine and deep learning have proven their ability to partially solve the RNA structure prediction problem. Not unlike in other fields, this shift has yielded state-of-the-art results compared to classical solutions. The architectures used vary mainly between sequential methods like LSTM networks and convolution-based solutions utilizing CNNs that have an advantage in both the results achieved and the computational resources necessary for training and inference.

A great limitation for predicting secondary structures with machine learning methods is the overfitting of the training data. This is especially true for complex methods based on deep learning with numerous parameters to be learned and adjusted. Although there are a variety of techniques to mitigate this issue, the RNA structure prediction is prone to overfitting by inappropriate diversification of testing data by RNA families. For example, Sato et al. ( 2021 ) showed that E2Efold (Chen et al. 2020 ) achieved a much lower F 1 -score for unseen families ( \(F=0.0361\) ). This showcases the overfitting problem even in published papers and highlights the importance of unified and standardized datasets to solve the RNA structure prediction problem.

One of the significant limitations of using machine-learning-based methods, especially in the prediction of tertiary structures, is the volume of available data. There is an insufficiency of structures mapped and uploaded to PDB that limits the possible knowledge to be obtained by a model, thus leading to sub-optimal solutions. This makes the methods prone to overfitting, resulting in learning the training RNA dataset instead of extracting and generalizing the knowledge (Das 2023 ). Despite huge progress in the area, the achieved results are not yet ideal for realistic use, as showcased by the RNA-Puzzles. More importantly, however, most of the works focused on creating a scoring function that can be used for structure modeling by evaluating a large number of generated structures. In this problem, deep learning yields mixed results compared to established classical methods, depending on the nature of the RNA families tested. Additionally, the need to generate large quantities of structures affects the potential performance of this kind of method.

A major challenge in predicting RNA structures comes directly from the extensive use of deep learning methods. These methods are generally regarded as “black-box models” that disallow interpretation or explanation of the achieved results. Although methods based on deep learning can achieve state-of-the-art results, with growing complexity, it becomes increasingly difficult to distinguish whether the models gather actual knowledge and follow some discovered folding mechanisms or just memorize the general idea of the structures available in the training dataset. The knowledge extraction problem itself can be partially alleviated by increasing the volume of data and diversifying the represented RNA families. However, even with recent advances in explainable AI (Saeed and Omlin 2023 ), extracting this knowledge from such models to expand our understanding of RNA structure remains unfeasible.

However, the future seems promising as the number of experimentally solved RNA structures is constantly increasing. This growth, coupled with the continuous development of ML/DL methods, indicates that it is only a matter of time before a breakthrough of AlphaFold’s magnitude will occur in solving 3D RNA structures.

An appealing idea for prospective research involves reviewing practical applications rather than methodological advancements. Such a review should incorporate various case studies, emphasizing practical insights and demonstrating the better performance and increased effectiveness of some methods for RNA structure prediction.

Data availability

No datasets were generated or analysed during the current study.

Akiyama M, Sakakibara Y, Sato K (2022) Direct inference of base-pairing probabilities with neural networks improves prediction of RNA secondary structures with pseudoknots. Genes 13:2155. https://doi.org/10.3390/genes13112155

Article   Google Scholar  

Almakarem ASA, Petrov AI, Stombaugh J et al (2012) Comprehensive survey and geometric classification of base triples in RNA structures. Nucleic Acids Res 40:1407–1423. https://doi.org/10.1093/nar/gkr810

Andersen RE (2017) Forging our understanding of lncRNAs in the brain. Cell Tissue Res 371:55–71. https://doi.org/10.1007/S00441-017-2711-Z

Andronescu M, Bereg V, Hoos HH et al (2008) RNA STRAND: the RNA secondary structure and statistical analysis database. BMC Bioinform 9:1–10. https://doi.org/10.1186/1471-2105-9-340

Andronescu M, Condon A, Hoos HH et al (2010) Computational approaches for RNA energy parameter estimation. RNA 16:2304–2318. https://doi.org/10.1261/rna.1950510

Anfinsen CB (1973) Principles that govern the folding of protein chains. Science 181:223–230. https://doi.org/10.1126/science.181.4096.223

Berman HM, Henrick HNK (2003) Announcing the worldwide Protein Data Bank. www.wwpdb.org . Accessed 27 Dec 2022

Bernauer J, Huang X, Sim AY et al (2011) Fully differentiable coarse-grained and all-atom knowledge-based potentials for RNA structure evaluation. RNA 17(6):1066–1075. https://doi.org/10.1261/rna.2543711

Booy MS, Ilin A, Orponen P (2022) RNA secondary structure prediction with convolutional neural networks. BMC Bioinform 23:58. https://doi.org/10.1186/s12859-021-04540-7

Brierley I, Pennell S (2007) Viral RNA pseudoknots: versatile motifs in gene expression and replication. Nat Rev Microbiol 5:598–610. https://doi.org/10.1038/nrmicro1704

Calonaci N, Jones A, Cuturello F et al (2020) Machine learning a model for RNA structure prediction. NAR Genomics Bioinform 2:Iqaa090. https://doi.org/10.1093/nargab/lqaa090

Capriotti E, Norambuena T, Marti-Renom MA et al (2011) All-atom knowledge-based potential for RNA structure prediction and assessment. Bioinformatics 27(8):1086–1093. https://doi.org/10.1093/bioinformatics/btr093

Castro E, Benz A, Tong A, et al (2020) Uncovering the folding landscape of RNA secondary structure using deep graph embeddings. In: 2020 IEEE international conference on big data (Big Data). Institute of Electrical and Electronics Engineers Inc., pp 4519–4528. https://doi.org/10.1109/bigdata50022.2020.9378305

Chen CC, Chan YM (2023) REDfold: accurate RNA secondary structure prediction using residual encoder-decoder network. BMC Bioinform 24:122. https://doi.org/10.1186/s12859-023-05238-8

Chen JH, Le SY, Maizel JV (2000) Prediction of common secondary structures of RNAs: a genetic algorithm approach. Nucleic Acids Res 28:991–999. https://doi.org/10.1093/nar/28.4.991

Chen X, Li Y, Umarov R et al (2020) RNA secondary structure prediction by learning unrolled algorithms. arXiv:2002.05810

Chen M, Ma Y, Wu S et al (2021) Genome warehouse: a public repository housing genome-scale data. Genomics Proteomics Bioinform 19(4):584–589. https://doi.org/10.1016/j.gpb.2021.04.001

Childs-Disney JL, Yang X, Gibaut QM et al (2022) Targeting RNA structures with small molecules. Nat Rev Drug Discov 21:736–762. https://doi.org/10.1038/s41573-022-00521-4

Choo J, Liu S (2018) Visual analytics for explainable deep learning. IEEE Comput Graphics Appl 38(4):84–92. https://doi.org/10.1109/mcg.2018.042731661

Collins M (2002) Discriminative training methods for hidden markov models: theory and experiments with perceptron algorithms. In: Proceedings of the 2002 conference on empirical methods in natural language processing (EMNLP 2002). Association for Computational Linguistics, pp 1–8. https://doi.org/10.3115/1118693.1118694

Consortium R (2020) RNAcentral 2021: secondary structure integration, improved sequence search and new member databases. Nucleic Acids Res 49(D1):D212–D220. https://doi.org/10.1093/nar/gkaa921

Constantin L (2018) Circular RNAs and neuronal development. Adv Exp Med Biol 1087:205–213. https://doi.org/10.1007/978-981-13-1426-1_16/cover

Crammer K, Dekel O, Keshet J et al (2006) Online passive-aggressive algorithms. J Mach Learn Res 7(19):551–585

MathSciNet   Google Scholar  

Czerniak T, Saenz JP (2021) Lipid membranes modulate the activity of RNA through sequence-dependent interactions. Proc Natl Acad Sci. https://doi.org/10.1073/pnas.2119235119

Danaee P, Rouches M, Wiley M et al (2018) bpRNA: large-scale automated annotation and analysis of RNA secondary structure. Nucleic Acids Res 46(11):5381–5394. https://doi.org/10.1093/nar/gky285

Das R (2023) Assessment of three-dimensional RNA structure prediction in CASP15. Proteins. https://doi.org/10.1101/2023.04.25.538330

Das R, Baker D (2007) Automated de novo prediction of native-like RNA tertiary structures. Proc Natl Acad Sci USA 104:14664–14669. https://doi.org/10.1073/pnas.0703836104

Das R, Karanicolas J (2010) Atomic accuracy in predicting and designing noncanonical RNA structure. Nat Methods 7:291–294. https://doi.org/10.1038/nmeth.1433

Deng C, Tang Y, Zhang J et al (2022) RNAGCN: RNA tertiary structure assessment with a graph convolutional network. Chin Phys B. https://doi.org/10.1088/1674-1056/ac8ce3

Ding Y, Lawrence CE (2003) A statistical sampling algorithm for RNA secondary structure prediction. Nucleic Acids Res 31:7280–7301. https://doi.org/10.1093/nar/gkg938

Ding Y, Chi YC, Lawrence CE (2005) RNA secondary structure prediction by centroids in a Boltzmann weighted ensemble. RNA 11:1157–1166. https://doi.org/10.1261/rna.2500605

Doudna JA (2002) The chemical repertoire of natural ribozymes. Nature 418:222–228. https://doi.org/10.1038/418222a

Eddy SR (2004) How do RNA folding algorithms work? Nat Biotechnol 22:1457–1458. https://doi.org/10.1038/nbt1104-1457

Eismann S, Townshend RJ, Thomas N et al (2021) Hierarchical, rotation-equivariant neural networks to select structural models of protein complexes. Proteins 89:493–501. https://doi.org/10.1002/prot.26033

Fatmi AE, Chentoufi A, Bekri MA, et al (2017) A heuristic algorithm for RNA secondary structure based on genetic algorithm. In: 2017 Intelligent systems and computer vision (ISCV), pp 1–7. https://doi.org/10.1109/isacv.2017.8054964

Fei Y, Zhang H, Wang Y et al (2022) LTPConstraint: a transfer learning based end-to-end method for RNA secondary structure prediction. BMC Bioinform 23:354. https://doi.org/10.1186/s12859-022-04847-z

Ferhadian D, Contrant M, Printz-Schweigert A et al (2018) Structural and functional motifs in influenza virus RNAs. Front Microbiol 9:559. https://doi.org/10.3389/fmicb.2018.00559/bibtex

Frid Y, Gusfield D (2010) A simple, practical and complete O-time Algorithm for RNA folding using the Four-Russians Speedup. Algorithms Mol Biol 5(1):13. https://doi.org/10.1186/1748-7188-5-13

Fu XD (2014) Non-coding RNA: a new frontier in regulatory biology. Natl Sci Rev 1:190–204. https://doi.org/10.1093/nsr/nwu008

Fu L, Cao Y, Wu J et al (2022) UFold: fast and accurate RNA secondary structure prediction with deep learning. Nucleic Acids Res 50:E14. https://doi.org/10.1093/nar/gkab1074

Garst AD, Edwards AL, Batey RT (2011) Riboswitches: structures and mechanisms. Cold Spring Harb Perspect Biol 3:1–13. https://doi.org/10.1101/CSHPERSPECT.A003533

Gilmer J, Schoenholz SS, Riley PF et al (2017) Neural message passing for quantum chemistry. arXiv:1704.01212

Graf J, Kretz M (2020) From structure to function: route to understanding lncRNA mechanism. BioEssays 42:2000027. https://doi.org/10.1002/BIES.202000027

Griffiths-Jones S, Moxon S, Marshall M et al (2005) Rfam: annotating non-coding RNAs in complete genomes. Nucleic Acids Res 33:D121–D124. https://doi.org/10.1093/NAR/GKI081

Grigorashvili EI, Chervontseva ZS, Gelfand MS (2022) Predicting RNA secondary structure by a neural network: what features may be learned? PeerJ 10:e14335. https://doi.org/10.7717/peerj.14335

Havgaard JH, Lyngsø RB, Gorodkin J (2005) The foldalign web server for pairwise structural RNA alignment and mutual motif search. Nucleic Acids Res 33:W650. https://doi.org/10.1093/NAR/GKI473

He K, Zhang X, Ren S et al (2016) Deep residual learning for image recognition. In: 2016 IEEE conference on computer vision and pattern recognition (CVPR), pp 770–778. https://doi.org/10.1109/CVPR.2016.90

Hendrix DK, Brenner SE, Holbrook SR (2005) RNA structural motifs: building blocks of a modular biomolecule. Q Rev Biophys 38:221–243. https://doi.org/10.1017/S0033583506004215

Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9:1735–80. https://doi.org/10.1162/neco.1997.9.8.1735

Hofacker IL, Fontana W, Stadler PF et al (1994) Fast folding and comparison of RNA secondary structures. Monatshefte für Chemie Chem Monthly 125:167–188. https://doi.org/10.1007/BF00818163/METRICS

Hu X, Chu L, Pei J et al (2021) Model complexity of deep learning: a survey. Knowl Inf Syst 63(10):2585–2619. https://doi.org/10.1007/s10115-021-01605-0

Jumper J, Evans R, Pritzel A et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589. https://doi.org/10.1038/s41586-021-03819-2

Kipf TN, Welling M (2017) Semi-supervised classification with graph convolutional networks. arXiv:1609.02907

Kleinkauf R, Mann M, Backofen R (2015) antaRNA: ant colony-based RNA sequence design. Bioinformatics 31:3114–3121. https://doi.org/10.1093/BIOINFORMATICS/BTV319

Kopp F, Mendell JT (2018) Functional classification and experimental dissection of long noncoding RNAs. Cell 172:393–407. https://doi.org/10.1016/J.CELL.2018.01.011

Kotar A, Foley HN, Baughman KM et al (2020) Advanced approaches for elucidating structures of large RNAs using NMR spectroscopy and complementary methods. Methods 183:93–107. https://doi.org/10.1016/J.YMETH.2020.01.009

LeCun Y, Boser BE, Denker JS et al (1989) Handwritten digit recognition with a back-propagation network. In: Neural information processing systems. https://doi.org/10.5555/109230.109279 , https://api.semanticscholar.org/CorpusID:2542741

Leontis NB, Westhof E (2001) Geometric nomenclature and classification of RNA base pairs. RNA 7:499–512. https://doi.org/10.1017/S1355838201002515

Leontis NB, Zirbel CL (2012) Nonredundant 3D structure datasets for RNA knowledge extraction and benchmarking, Springer, Berlin, pp 281–298. https://doi.org/10.1007/978-3-642-25740-7_13

Li J, Zhu W, Wang J et al (2018) RNA3DCNN: local and global quality assessments of RNA 3D structures using 3D deep convolutional neural networks. PLoS Comput Biol 14:e1006514. https://doi.org/10.1371/journal.pcbi.1006514

Li G, Muller M, Thabet A et al (2019) DeepGCNs: can GCNs go as deep as CNNs? In: Proceedings of the IEEE/CVF international conference on computer vision (ICCV)

Li G, Müller M, Ghanem B et al (2021) Training graph neural networks with 1000 layers. In: Meila M, Zhang T (eds) Proceedings of the 38th international conference on machine learning, proceedings of machine learning research, vol 139. PMLR, pp 6437–6449

Liu B, Dai Y, Li X et al (2003) Building text classifiers using positive and unlabeled examples. In: Third IEEE international conference on data mining, pp 179–186. https://doi.org/10.1109/ICDM.2003.1250918

Lorenz R, Bernhart SH, zu Siederdissen CH (2011) ViennaRNA Package 2.0. Algorithms Mol Biol 6:1–14. https://doi.org/10.1186/1748-7188-6-26/TABLES/2

Lorenz R, Bernhart SH, Qin J et al (2013) 2D meets 4G: G-quadruplexes in RNA secondary structure prediction. IEEE/ACM Trans Comput Biol Bioinform 10:832–844. https://doi.org/10.1109/TCBB.2013.7

Lu W, Tang Y, Wu H et al (2019) Predicting RNA secondary structure via adaptive deep recurrent neural networks with energy-based filter. BMC Bioinform 20:1–10. https://doi.org/10.1186/s12859-019-3258-7

Lundberg SM, Lee SI (2017) A unified approach to interpreting model predictions. In: Guyon I, Luxburg UV, Bengio S et al (eds) Advances in neural information processing systems, vol 30. Curran Associates Inc., Glasgow, pp 4765–4774

Google Scholar  

Mańka R, Janas P, Sapoń K et al (2021) Role of RNA motifs in RNA interaction with membrane lipid rafts: implications for therapeutic applications of exosomal RNAs. Int J Mol Sci 22:9416. https://doi.org/10.3390/ijms22179416

Mao K, Xiao Y (2021) Learning the fastest RNA folding path based on reinforcement learning and monte carlo tree search. Molecules 26:4420. https://doi.org/10.3390/molecules26154420

Mao K, Wang J, Xiao Y (2022) Length-dependent deep learning model for RNA secondary structure prediction. Molecules 27:1030. https://doi.org/10.3390/molecules27031030

Mathews DH (2004) Using an RNA secondary structure partition function to determine confidence in base pairs predicted by free energy minimization. RNA 10:1178–1190. https://doi.org/10.1261/RNA.7650904

Mathews DH, Zuker M (2004) Predictive methods using RNA sequences. Bioinformatics 143–170

Mehta A (2016) MicroRNAs as regulatory elements in immune system logic. Nat Rev Immunol 16:279–294. https://doi.org/10.1038/nri.2016.40

Meister G (2004) Mechanisms of gene silencing by double-stranded RNA. Nature 431:343–349. https://doi.org/10.1038/nature02873

Miao Z, Adamiak RW, Antczak M et al (2020) RNA-Puzzles Round IV: 3D structure predictions of four ribozymes and two aptamers. RNA 26:982–995

Mnih V, Kavukcuoglu K, Silver D et al (2015) Human-level control through deep reinforcement learning. Nature 518(7540):529–533. https://doi.org/10.1038/nature14236

Moore PB (1999) Structural motifs in RNA. Annu Rev Biochem 68:287–300. https://doi.org/10.1146/annurev.biochem.68.1.287

Mortimer SA, Kidwell MA, Doudna JA (2014) Insights into RNA structure and function from genome-wide studies. Nat Rev Genet 15:469–479

Nawrocki EP, Eddy SR (2013) Infernal 1.1: 100-fold faster RNA homology searches. Bioinformatics 29:2933–2935. https://doi.org/10.1093/bioinformatics/btt509

Nussinov R, Jacobson AB (1980) Fast algorithm for predicting the secondary structure of single-stranded RNA. Proc Natl Acad Sci 77(11):6309–6313. https://doi.org/10.1073/pnas.77.11.6309

Pearce R, Li Y, Omenn GS et al (2022a) Fast and accurate Ab Initio Protein structure prediction using deep learning potentials. PLoS Comput Biol 18(9):1–22. https://doi.org/10.1371/journal.pcbi.1010539

Pearce R, Omenn GS, Zhang Y (2022b) De novo RNA tertiary structure prediction at atomic resolution using geometric potentials from deep learning. bioRxiv. https://doi.org/10.1101/2022.05.15.491755

Qiu X (2023) Sequence similarity governs generalizability of de novo deep learning models for RNA secondary structure prediction. PLoS Comput Biol 19:e1011047. https://doi.org/10.1371/journal.pcbi.1011047

Quan L, Cai L, Chen Y et al (2020) Developing parallel ant colonies filtered by deep learned constrains for predicting RNA secondary structure with pseudo-knots. Neurocomputing 384:104–114. https://doi.org/10.1016/j.neucom.2019.12.041

Reinharz V, Ponty Y, Waldispühl J (2013) A weighted sampling algorithm for the design of RNA sequences with targeted secondary structure and nucleotide distribution. Bioinformatics 29:i308–i315. https://doi.org/10.1093/bioinformatics/btt217

Ribeiro MT, Singh S, Guestrin C (2016) “Why should I trust you?”: explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, San Francisco, CA, USA, August 13–17, 2016, pp 1135–1144

Rivas E, Eddy SR (1999) A dynamic programming algorithm for RNA structure prediction including pseudoknots11 Edited by I Tinoco. J Mol Biol 285(5):2053–2068. https://doi.org/10.1006/jmbi.1998.2436

Ronneberger O, Fischer P, Brox T (2015) U-Net: convolutional networks for biomedical image segmentation. Springer, New York, pp 234–241. https://doi.org/10.1007/978-3-319-24574-4_28

Book   Google Scholar  

Ross CJ, Ulitsky I (2022) Discovering functional motifs in long noncoding RNAs. Wiley Interdiscip Rev 13:e1708. https://doi.org/10.1002/wrna.1708

Saeed W, Omlin C (2023) Explainable AI (XAI): a systematic meta-survey of current challenges and future opportunities. Knowl-Based Syst 263:110273. https://doi.org/10.1016/j.knosys.2023.110273

Sato K, Hamada M (2023) Recent trends in RNA informatics: a review of machine learning and deep learning for RNA secondary structure prediction and RNA drug discovery. Brief Bioinform 24(4):bbad186. https://doi.org/10.1093/bib/bbad186

Sato K, Akiyama M, Sakakibara Y (2021) RNA secondary structure prediction using deep learning with thermodynamic integration. Nat Commun 12:941. https://doi.org/10.1038/s41467-021-21194-4

Sayers EW, Beck J, Bolton EE et al (2020) Database resources of the National Center for Biotechnology Information. Nucleic Acids Res 49(D1):D10–D17. https://doi.org/10.1093/nar/gkaa892

Schmitt AM, Chang HY (2016) Long noncoding RNAs in cancer pathways. Cancer Cell 29:452–463. https://doi.org/10.1016/j.ccell.2016.03.010

Seemann SE, Gorodkin J, Backofen R (2008) Unifying evolutionary and thermodynamic information for RNA folding of multiple alignments. Nucleic Acids Res 36(20):6355–6362. https://doi.org/10.1093/nar/gkn544

Serganov A, Nudler E (2013) A decade of riboswitches. Cell 152:17–24. https://doi.org/10.1016/j.cell.2012.12.024

Shahidul Islam M, Rafiqul Islam M (2022) A hybrid framework based on genetic algorithm and simulated annealing for RNA structure prediction with pseudoknots. J King Saud Univ-Comput Inf Sci 34(3):912–922. https://doi.org/10.1016/j.jksuci.2020.03.005

Shapiro BA, Navetta J (1994) A massively parallel genetic algorithm for RNA secondary structure prediction. J Supercomput 8:195–207. https://doi.org/10.1007/bf01204728

Shcherbakova I, Mitra S, Laederach A et al (2008) Energy barriers, pathways, and dynamics during folding of large, multidomain RNAs. Curr Opin Chem Biol 12:655–666. https://doi.org/10.1016/j.cbpa.2008.09.017

Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. In: 3rd international conference on learning representations, ICLR 2015—conference track proceedings. arXiv:1409.1556

Singh J, Hanson J, Paliwal K et al (2019) RNA secondary structure prediction using an ensemble of two-dimensional deep neural networks and transfer learning. Nat Commun 10:5407. https://doi.org/10.1038/s41467-019-13395-9

Singh J, Paliwal K, Zhang T et al (2021) Improved RNA secondary structure and tertiary base-pairing prediction using evolutionary profile, mutational coupling and two-dimensional transfer learning. Bioinformatics 37:2589–2600. https://doi.org/10.1093/bioinformatics/btab165

Singh J, Paliwal K, Litfin T et al (2022) Predicting RNA distance-based contact maps by integrated deep learning on physics-inferred secondary structure and evolutionary-derived mutational coupling. Bioinformatics 38:3900–3910. https://doi.org/10.1093/bioinformatics/btac421

Sloma MF, Mathews DH (2016) Exact calculation of loop formation probability identifies folding motifs in RNA secondary structures. RNA 22:1808–1818. https://doi.org/10.1261/rna.053694.115

Stark B, Kolet R, Bowman E et al (1978) Biochemistry Ribonuclease P: An enzyme with an essential RNA component (endoribonuclease/precursor tRNA substrates/RNA subunit). Proc Natl Acad Sci USA 75:3717–3721

Steinegger M, Söding J (2017) MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol 35(11):1026–1028. https://doi.org/10.1038/nbt.3988

Stephens ZD, Lee SY, Faghri F et al (2015) Big Data: astronomical or genomical? PLoS Biol 13:e1002195. https://doi.org/10.1371/journal.pbio.1002195

Su C, Weir JD, Zhang F et al (2019) ENTRNA: a framework to predict RNA foldability. BMC Bioinform 20:1–11. https://doi.org/10.1186/s12859-019-2948-5

Sun M, Kraus WL (2015) From discovery to function: the expanding roles of long NonCoding RNAs in physiology and disease. Endocr Rev 36:25–64. https://doi.org/10.1210/er.2014-1034

Sutton C, McCallum A (2010) An introduction to conditional random fields. arXiv:1011.4088

Sutton RS, McAllester D, Singh S et al (1999) Policy gradient methods for reinforcement learning with function approximation. In: Proceedings of the 12th international conference on neural information processing systems. MIT Press, Cambridge, MA, USA, NIPS’99, pp 1057–1063

Tan Z, Fu Y, Sharma G et al (2017) TurboFold II: RNA structural alignment and secondary structure prediction informed by multiple homologs. Nucleic Acids Res 45:11570–11581. https://doi.org/10.1093/nar/gkx815

Tan YL, Wang X, Shi YZ et al (2022) rsRNASP: a residue-separation-based statistical potential for RNA 3D structure evaluation. Biophys J 121:142–156. https://doi.org/10.1016/j.bpj.2021.11.016

Taneda A (2012) Multi-objective genetic algorithm for pseudoknotted RNA sequence design. Front Genet 3:36. https://doi.org/10.3389/fgene.2012.00036/bibtex

Thomas N, Smidt T, Kearnes S et al (2018) Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds. arXiv:1802.08219

Townshend RJL, Eismann S, Watkins AM et al (2021) Geometric deep learning of RNA structure. Science 373:1047–1051. https://doi.org/10.1126/science.abe5650

Turner DH, Mathews DH (2009) NNDB: the nearest neighbor parameter database for predicting stability of nucleic acid secondary structure. Nucleic Acids Res 38:D280–D282. https://doi.org/10.1093/nar/gkp892

Ulitsky I, Bartel DP (2013) lincRNAs: genomics, evolution, and mechanisms. Cell 154:26. https://doi.org/10.1016/j.cell.2013.06.020

Vaswani A, Shazeer N, Parmar N et al (2017) Attention is all you need. arXiv:1706.03762

Veličković P, Cucurull G, Casanova A et al (2018) Graph attention networks. arXiv:1710.10903

Wang KC, Chang HY (2011) Molecular mechanisms of long noncoding RNAs. Mol Cell 43:904–914. https://doi.org/10.1016/j.molcel.2011.08.018

Wang J, Zhao Y, Zhu C et al (2015) 3dRNAscore: a distance and torsion angle dependent evaluation function of 3D RNA structures. Nucleic Acids Res 43:e63–e63. https://doi.org/10.1093/nar/gkv141

Wang L, Liu Y, Zhong X et al (2019a) DMFold: a novel method to predict RNA secondary structure with pseudoknots based on deep learning and improved base pair maximization principle. Front Genet 10:143. https://doi.org/10.3389/fgene.2019.00143

Wang YZ, Li J, Zhang S et al (2019b) An RNA scoring function for tertiary structure prediction based on multi-layer neural networks. Mol Biol 53:118–126. https://doi.org/10.1134/S0026893319010175

Wang Y, Liu Y, Wang S et al (2020) ATTfold: RNA secondary structure prediction with pseudoknots based on attention mechanism. Front Genet 11:612086. https://doi.org/10.3389/fgene.2020.612086

Wang X, Gu R, Chen Z et al (2023a) UNI-RNA: universal pre-trained models revolutionze RNA Res. https://doi.org/10.1101/2023.07.11.548588

Wang X, Yu S, Lou E et al (2023b) RNA 3D structure prediction: progress and perspective. Molecules 28(14):5532. https://doi.org/10.3390/molecules28145532

Watkins AM, Rangan R, Das R (2020) FARFAR2: improved de novo Rosetta prediction of complex global RNA folds. Structure 28:963-976.e6. https://doi.org/10.1016/j.str.2020.05.011

Willmott D, Murrugarra D, Ye Q (2020) Improving RNA secondary structure prediction via state inference with deep recurrent neural networks. Comput Math Biophys 8:36–50. https://doi.org/10.1515/cmb-2020-0002

Article   MathSciNet   Google Scholar  

Wilusz JE, Sunwoo H, Spector DL (2009) Long noncoding RNAs: functional surprises from the RNA world. Genes Dev 23:1494. https://doi.org/10.1101/gad.1800909

Wu H, Tang Y, Lu W et al (2018) RNA secondary structure prediction based on long short-term memory model. In: Intelligent computing theories and application, vol 10954 LNCS. Springer, New York, pp 595–599. https://doi.org/10.1007/978-3-319-95930-6_59

Xu J, Liu Y, Li H et al (2022) Transcriptional and functional motifs defining renal function revealed by single-nucleus RNA sequencing. Proc Natl Acad Sci USA 119:e2203179119. https://doi.org/10.1073/pnas.2203179119/-/dcsupplemental

Yakovchuk P, Protozanova E, Frank-Kamenetskii MD (2006) Base-stacking and base-pairing contributions into thermal stability of the DNA double helix. Nucleic Acids Res 34:564–574. https://doi.org/10.1093/nar/gkj454

Yang VW, Lerner MR, Steitz JA et al (1981) A small nuclear ribonucleoprotein is required for splicing of adenoviral early RNA sequences. Proc Natl Acad Sci USA 78:1371. https://doi.org/10.1073/pnas.78.3.1371

Yonemoto H, Asai K, Hamada M (2015) A semi-supervised learning approach for RNA secondary structure prediction. Comput Biol Chem 57:72–79. https://doi.org/10.1016/j.compbiolchem.2015.02.002

Yu H, Qi Y, Ding Y (2022) Deep learning in RNA structure studies. Front Mol Biosci 9:869601. https://doi.org/10.3389/fmolb.2022.869601

Zakov S, Goldberg Y, Elhadad M et al (2011) Rich parameterization improves RNA structure prediction. J Comput Biol 18:1525–1542. https://doi.org/10.1089/cmb.2011.0184

Zhang H, Zhang C, Li Z et al (2019) A new method of RNA secondary structure prediction based on convolutional neural network and dynamic programming. Front Genet 10:467. https://doi.org/10.3389/fgene.2019.00467

Zhang J, Fei Y, Sun L et al (2022) Advances and opportunities in RNA structure experimental determination and computational modeling. Nat Methods 19(10):1193–1207. https://doi.org/10.1038/s41592-022-01623-y

Zhang C, Zhang Y, Pyle AM (2023) rMSA: a sequence search and alignment algorithm to improve RNA structure modeling. J Mol Biol 435:167904. https://doi.org/10.1016/j.jmb.2022.167904

Zhao Y, Wang J, Zeng C et al (2018) Evaluation of RNA secondary structure prediction for both base-pairing and topology. Biophys Rep 4:123–132. https://doi.org/10.1007/S41048-018-0058-Y

Zhao Q, Zhao Z, Fan X et al (2021) Review of machine learning methods for RNA secondary structure prediction. PLoS Comput Biol. https://doi.org/10.1371/journal.pcbi.1009291

Zhao Q, Mao Q, Zhao Z et al (2023) RNA independent fragment partition method based on deep learning for RNA secondary structure prediction. Sci Rep 13:2861. https://doi.org/10.1038/s41598-023-30124-x

Zhou J, Troyanskaya OG (2015) Predicting effects of noncoding variants with deep learning-based sequence model. Nat Methods 12(10):931–934. https://doi.org/10.1038/nmeth.3547

Zuker M (1989) Computer prediction of RNA structure. In: RNA Processing Part A: general methods, methods in enzymology, vol 180. Academic Press, New York, pp 262–288. https://doi.org/10.1016/0076-6879(89)80106-5

Zuker M (2003) Mfold web server for nucleic acid folding and hybridization prediction. Nucleic Acids Res 31(13):3406–3415. https://doi.org/10.1093/nar/gkg595

Download references

Acknowledgements

M. Budnik acknowledges support from the Polish Ministry of Science and Higher Education, Project no. DWD/6/0059/2022. M. Kadziński was supported by the Polish Ministry of Science and Higher Education, Grant no. 0311/SBAD/0742.

Author information

Authors and affiliations.

Institute of Computing Science, Poznan University of Technology, Piotrowo 2, 60-965, Poznań, Poland

Michał Budnik & Miłosz Kadziński

TIDK Sp. z o.o., Bałtycka 6, 61-013, Poznań, Poland

Michał Budnik, Jakub Wawrzyniak, Łukasz Grala & Natalia Szóstak

Institute of Bioorganic Chemistry, Polish Academy of Sciences, Zygmunta Noskowskiego 12/14, 61-704, Poznań, Poland

Natalia Szóstak

You can also search for this author in PubMed   Google Scholar

Contributions

MB prepared the study protocol, performed the research, discussed the results, prepared tables and figures, and wrote the manuscript; JW and ŁG discussed the study protocol and the results, and oversaw the article creation process; MK discussed the results and helped with manuscript writing; NS participated in the study design, coordinated the study, discussed results, and helped with manuscript writing. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Natalia Szóstak .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary file 1 (pdf 123 KB)

Rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Budnik, M., Wawrzyniak, J., Grala, Ł. et al. Deep dive into RNA: a systematic literature review on RNA structure prediction using machine learning methods. Artif Intell Rev 57 , 254 (2024). https://doi.org/10.1007/s10462-024-10910-3

Download citation

Accepted : 06 August 2024

Published : 15 August 2024

DOI : https://doi.org/10.1007/s10462-024-10910-3

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• RNA structure prediction
• Machine learning
• Deep learning
• Systematic literature review
• Find a journal
• Publish with us
• Track your research

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Analysing near-miss incidents in construction: a systematic literature review.

1. Introduction

• Q 1 —Are near-miss events in construction industry the subject of scientific research?
• Q 2 —What methods have been employed thus far to obtain information on near misses and systems for recording incidents in construction companies?
• Q 3 —What methods have been used to analyse the information and figures obtained?
• Q 4 —What are the key aspects of near misses in the construction industry that have been of interest to the researchers?

2. Definition of Near-Miss Events

3. research methodology, 4.1. a statistical analysis of publications, 4.2. methods used to obtain information about near misses, 4.2.1. traditional methods.

• Traditional registration forms
• Computerized systems for the recording of events
• Surveys and interviews

4.2.2. Real-Time Monitoring Systems

• Employee-tracking systems
• Video surveillance systems
• Wearable technology
• Motion sensors

4.3. Methods Used to Analyse the Information and Figures That Have Been Obtained

4.3.1. quantitative and qualitative statistical methods, 4.3.2. analysis using artificial intelligence (ai), 4.3.3. building information modelling, 4.4. key aspects of near-miss investigations in the construction industry, 4.4.1. occupational risk assessment, 4.4.2. causes of hazards in construction, 4.4.3. time series of near misses, 4.4.4. material factors of construction processes, 4.5. a comprehensive overview of the research questions and references on near misses in the construction industry, 5. discussion, 5.1. interest of researchers in near misses in construction (question 1), 5.2. methods used to obtain near-miss information (question 2), 5.3. methods used to analyse the information and data sets (question 3), 5.4. key aspects of near-miss investigations in the construction industry (question 4), 6. conclusions.

• A quantitative analysis of the Q 1 question has revealed a positive trend, namely that there is a growing interest among researchers in studying near misses in construction. The greatest interest in NM topics is observed in the United States of America, China, the United Kingdom, Australia, Hong Kong, and Germany. Additionally, there has been a recent emergence of interest in Poland. The majority of articles are mainly published in journals such as Safety Science (10), Journal of Construction Engineering and Management (8), and Automation in Construction (5);
• The analysis of question Q 2 illustrates that traditional paper-based event registration systems are currently being superseded by advanced IT systems. However, both traditional and advanced systems are subject to the disadvantage of relying on employee-reported data, which introduces a significant degree of uncertainty regarding in the quality of the information provided. A substantial proportion of the data and findings presented in the studies was obtained through surveys and interviews. The implementation of real-time monitoring systems is becoming increasingly prevalent in construction sites. The objective of such systems is to provide immediate alerts in the event of potential hazards, thereby preventing a significant number of near misses. Real-time monitoring systems employ a range of technologies, including ultrasonic technology, radio frequency identification (RFID), inertial measurement units (IMUs), real-time location systems (RTLSs), industrial cameras, wearable technology, motion sensors, and advanced IT technologies, among others;
• The analysis of acquired near-miss data is primarily conducted through the utilisation of quantitative and qualitative statistical methods, as evidenced by the examination of the Q 3 question. In recent years, research utilising artificial intelligence (AI) has made significant advances. The most commonly employed artificial intelligence techniques include text mining, machine learning, and artificial neural networks. The growing deployment of Building Information Modelling (BIM) technology has precipitated a profound transformation in the safety management of construction sites, with the advent of sophisticated tools for the identification and management of hazardous occurrences;
• In response to question Q 4 , the study of near misses in the construction industry has identified several key aspects that have attracted the attention of researchers. These include the utilisation of both quantitative and qualitative methodologies for risk assessment, the analysis of the causes of hazards, the identification of accident precursors through the creation of time series, and the examination of material factors pertaining to construction processes. Researchers are focusing on the utilisation of both databases and advanced technologies, such as real-time location tracking, for the assessment and analysis of occupational risks. Techniques such as Analytic Hierarchy Process (AHP) and clustering facilitate a comprehensive assessment and categorisation of incidents, thereby enabling the identification of patterns and susceptibility to specific types of accidents. Moreover, the impact of a company’s safety climate and organisational culture on the frequency and characteristics of near misses represents a pivotal area of investigation. The findings of this research indicate that effective safety management requires a holistic approach that integrates technology, risk management and safety culture, with the objective of reducing accidents and enhancing overall working conditions on construction sites.

7. Gaps and Future Research Directions, Limitations

• Given the diversity and variability of construction sites and the changing conditions and circumstances of work, it is essential to create homogeneous clusters of near misses and to analyse the phenomena within these clusters. The formation of such clusters may be contingent upon the direct causes of the events in question;
• Given the inherently dynamic nature of construction, it is essential to analyse time series of events that indicate trends in development and safety levels. The numerical characteristics of these trends may be used to construct predictive models for future accidents and near misses;
• The authors have identified potential avenues for future research, which could involve the development of mathematical models using techniques such as linear regression, artificial intelligence, and machine learning. The objective of these models is to predict the probable timing of occupational accidents within defined incident categories, utilising data from near misses. Moreover, efforts are being made to gain access to the hazardous incident recording systems of different construction companies, with a view to facilitating comparison of the resulting data;
• One significant limitation of near-miss research is the lack of an integrated database that encompasses a diverse range of construction sites and construction work. A data resource of this nature would be of immense value for the purpose of conducting comprehensive analyses and formulating effective risk management strategies. This issue can be attributed to two factors: firstly, the reluctance of company managers to share their databases with researchers specialising in risk assessment, and secondly, the reluctance of employees to report near-miss incidents. Such actions may result in adverse consequences for employees, including disciplinary action or negative perceptions from managers. This consequently results in the recording of only a subset of incidents, thereby distorting the true picture of safety on the site.

Author Contributions

Institutional review board statement, informed consent statement, data availability statement, conflicts of interest.

• Occupational Risk|Safety and Health at Work EU-OSHA. Available online: https://osha.europa.eu/en/tools-and-resources/eu-osha-thesaurus/term/70194i (accessed on 28 June 2023).
• Guo, S.; Zhou, X.; Tang, B.; Gong, P. Exploring the Behavioral Risk Chains of Accidents Using Complex Network Theory in the Construction Industry. Phys. A Stat. Mech. Its Appl. 2020 , 560 , 125012. [ Google Scholar ] [ CrossRef ]
• Woźniak, Z.; Hoła, B. The Structure of near Misses and Occupational Accidents in the Polish Construction Industry. Heliyon 2024 , 10 , e26410. [ Google Scholar ] [ CrossRef ]
• Li, X.; Sun, W.; Fu, H.; Bu, Q.; Zhang, Z.; Huang, J.; Zang, D.; Sun, Y.; Ma, Y.; Wang, R.; et al. Schedule Risk Model of Water Intake Tunnel Construction Considering Mood Factors and Its Application. Sci. Rep. 2024 , 14 , 3857. [ Google Scholar ] [ CrossRef ]
• Li, X.; Huang, J.; Li, C.; Luo, N.; Lei, W.; Fan, H.; Sun, Y.; Chen, W. Study on Construction Resource Optimization and Uncertain Risk of Urban Sewage Pipe Network. Period. Polytech. Civ. Eng. 2022 , 66 , 335–343. [ Google Scholar ] [ CrossRef ]
• Central Statistical Office Central Statistical Office/Thematic Areas/Labor Market/Working Conditions/Accidents at Work/Accidents at Work in the 1st Quarter of 2024. Available online: https://stat.gov.pl/obszary-tematyczne/rynek-pracy/warunki-pracy-wypadki-przy-pracy/wypadki-przy-pracy-w-1-kwartale-2024-roku,3,55.html (accessed on 17 July 2024).
• Manzo, J. The $ 5 Billion Cost of Construction Fatalities in the United States: A 50 State Comparison ; The Midwest Economic Policy Institute (MEPI): Saint Paul, MN, USA, 2017. [ Google Scholar ]
• Sousa, V.; Almeida, N.M.; Dias, L.A. Risk-Based Management of Occupational Safety and Health in the Construction Industry—Part 1: Background Knowledge. Saf. Sci. 2014 , 66 , 75–86. [ Google Scholar ] [ CrossRef ]
• Amirah, N.A.; Him, N.F.N.; Rashid, A.; Rasheed, R.; Zaliha, T.N.; Afthanorhan, A. Fostering a Safety Culture in Manufacturing through Safety Behavior: A Structural Equation Modelling Approach. J. Saf. Sustain. 2024; in press . [ Google Scholar ] [ CrossRef ]
• Heinrich, H.W. Industrial Accident Prevention ; A Scientific Approach; McGraw-Hill: New York, NY, USA, 1931. [ Google Scholar ]
• Near Miss Definition Per OSHA—What Is a Near Miss? Available online: https://safetystage.com/osha-compliance/near-miss-definition-osha/ (accessed on 17 August 2024).
• Cambraia, F.B.; Saurin, T.A.; Formoso, C.T. Identification, Analysis and Dissemination of Information on near Misses: A Case Study in the Construction Industry. Saf. Sci. 2010 , 48 , 91–99. [ Google Scholar ] [ CrossRef ]
• Tan, J.; Li, M. How to Achieve Accurate Accountability under Current Administrative Accountability System for Work Safety Accidents in Chemical Industry in China: A Case Study on Major Work Safety Accidents during 2010–2020. J. Chin. Hum. Resour. Manag. 2022 , 13 , 26–40. [ Google Scholar ] [ CrossRef ]
• Wu, W.; Gibb, A.G.F.; Li, Q. Accident Precursors and near Misses on Construction Sites: An Investigative Tool to Derive Information from Accident Databases. Saf. Sci. 2010 , 48 , 845–858. [ Google Scholar ] [ CrossRef ]
• Janicak, C.A. Fall-Related Deaths in the Construction Industry. J. Saf. Res. 1998 , 29 , 35–42. [ Google Scholar ] [ CrossRef ]
• Li, H.; Yang, X.; Wang, F.; Rose, T.; Chan, G.; Dong, S. Stochastic State Sequence Model to Predict Construction Site Safety States through Real-Time Location Systems. Saf. Sci. 2016 , 84 , 78–87. [ Google Scholar ] [ CrossRef ]
• Yang, K.; Aria, S.; Ahn, C.R.; Stentz, T.L. Automated Detection of Near-Miss Fall Incidents in Iron Workers Using Inertial Measurement Units. In Proceedings of the Construction Research Congress 2014: Construction in a Global Network, Atlanta, GA, USA, 19–21 May 2014; pp. 935–944. [ Google Scholar ] [ CrossRef ]
• Raviv, G.; Fishbain, B.; Shapira, A. Analyzing Risk Factors in Crane-Related near-Miss and Accident Reports. Saf. Sci. 2017 , 91 , 192–205. [ Google Scholar ] [ CrossRef ]
• Zhao, X.; Zhang, M.; Cao, T. A Study of Using Smartphone to Detect and Identify Construction Workers’ near-Miss Falls Based on ANN. In Proceedings of the Nondestructive Characterization and Monitoring of Advanced Materials, Aerospace, Civil Infrastructure, and Transportation XII, Denver, CO, USA, 4–8 March 2018; p. 80. [ Google Scholar ] [ CrossRef ]
• Santiago, K.; Yang, X.; Ruano-Herreria, E.C.; Chalmers, J.; Cavicchia, P.; Caban-Martinez, A.J. Characterising near Misses and Injuries in the Temporary Agency Construction Workforce: Qualitative Study Approach. Occup. Environ. Med. 2020 , 77 , 94–99. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• What Is OSHA’s Definition of a Near Miss. Available online: https://www.osha.com/blog/near-miss-definition (accessed on 4 August 2023).
• Martins, I. Investigation of Occupational Accidents and Diseases a Practical Guide for Labour Inspectors ; International Labour Office: Geneva, Switzerland, 2015. [ Google Scholar ]
• National Safety Council. Near Miss Reporting Systems ; National Safety Council: Singapore, 2013. [ Google Scholar ]
• PKN PN-ISO 45001:2018-06 ; Occupational Health and Safety Management Systems—Requirements with Guidance for Use. CRC Press: Boca Raton, FL, USA, 2019.
• PKN PN-N-18001:2004 ; Occupational Health and Safety Management Systems—Requirements. CRC Press: Boca Raton, FL, USA, 2004.
• World Health Organisation. WHO Draft GuiDelines for Adverse Event Reporting and Learning Systems ; World Health Organisation: Geneva, Switzerland, 2005. [ Google Scholar ]
• International Atomic Energy Agency IAEA Satety Glossary. Terminology Used in Nuclear Safety and Radiation Protection: 2007 Edition ; International Atomic Energy Agency: Vienna, Austria, 2007. [ Google Scholar ]
• Marks, E.; Teizer, J.; Hinze, J. Near Miss Reporting Program to Enhance Construction Worker Safety Performance. In Proceedings of the Construction Research Congress 2014: Construction in a Global Network, Atlanta, GA, USA, 19 May 2014; pp. 2315–2324. [ Google Scholar ] [ CrossRef ]
• Gnoni, M.G.; Saleh, J.H. Near-Miss Management Systems and Observability-in-Depth: Handling Safety Incidents and Accident Precursors in Light of Safety Principles. Saf. Sci. 2017 , 91 , 154–167. [ Google Scholar ] [ CrossRef ]
• Thoroman, B.; Goode, N.; Salmon, P. System Thinking Applied to near Misses: A Review of Industry-Wide near Miss Reporting Systems. Theor. Issues Ergon. Sci. 2018 , 19 , 712–737. [ Google Scholar ] [ CrossRef ]
• Gnoni, M.G.; Tornese, F.; Guglielmi, A.; Pellicci, M.; Campo, G.; De Merich, D. Near Miss Management Systems in the Industrial Sector: A Literature Review. Saf. Sci. 2022 , 150 , 105704. [ Google Scholar ] [ CrossRef ]
• Bird, F. Management Guide to Loss Control ; Loss Control Publications: Houston, TX, USA, 1975. [ Google Scholar ]
• Zimmermann. Bauer International Norms and Identity ; Zimmermann: Sydney, NSW, Australia, 2006; pp. 5–21. [ Google Scholar ]
• Arslan, M.; Cruz, C.; Ginhac, D. Semantic Trajectory Insights for Worker Safety in Dynamic Environments. Autom. Constr. 2019 , 106 , 102854. [ Google Scholar ] [ CrossRef ]
• Arslan, M.; Cruz, C.; Ginhac, D. Visualizing Intrusions in Dynamic Building Environments for Worker Safety. Saf. Sci. 2019 , 120 , 428–446. [ Google Scholar ] [ CrossRef ]
• Zhou, C.; Chen, R.; Jiang, S.; Zhou, Y.; Ding, L.; Skibniewski, M.J.; Lin, X. Human Dynamics in Near-Miss Accidents Resulting from Unsafe Behavior of Construction Workers. Phys. A Stat. Mech. Its Appl. 2019 , 530 , 121495. [ Google Scholar ] [ CrossRef ]
• Chen, F.; Wang, C.; Wang, J.; Zhi, Y.; Wang, Z. Risk Assessment of Chemical Process Considering Dynamic Probability of near Misses Based on Bayesian Theory and Event Tree Analysis. J. Loss Prev. Process Ind. 2020 , 68 , 104280. [ Google Scholar ] [ CrossRef ]
• Wright, L.; Van Der Schaaf, T. Accident versus near Miss Causation: A Critical Review of the Literature, an Empirical Test in the UK Railway Domain, and Their Implications for Other Sectors. J. Hazard. Mater. 2004 , 111 , 105–110. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Saleh, J.H.; Saltmarsh, E.A.; Favar, F.M.; Loı¨c Brevault, L. Accident Precursors, near Misses, and Warning Signs: Critical Review and Formal Definitions within the Framework of Discrete Event Systems. Reliab. Eng. Syst. Saf. 2013 , 114 , 148–154. [ Google Scholar ] [ CrossRef ]
• Fred, A. Manuele Reviewing Heinrich. Am. Soc. Saf. Prof. 2011 , 56 , 52–61. [ Google Scholar ]
• Love, P.E.D.; Tenekedjiev, K. Understanding Near-Miss Count Data on Construction Sites Using Greedy D-Vine Copula Marginal Regression: A Comment. Reliab. Eng. Syst. Saf. 2022 , 217 , 108021. [ Google Scholar ] [ CrossRef ]
• Jan van Eck, N.; Waltman, L. VOSviewer Manual ; Universiteit Leiden: Leiden, The Netherlands, 2015. [ Google Scholar ]
• Scopus. Content Coverage Guide ; Elsevier: Amsterdam, The Netherlands, 2023; pp. 1–24. [ Google Scholar ]
• Lukic, D.; Littlejohn, A.; Margaryan, A. A Framework for Learning from Incidents in the Workplace. Saf. Sci. 2012 , 50 , 950–957. [ Google Scholar ] [ CrossRef ]
• Teizer, J.; Cheng, T. Proximity Hazard Indicator for Workers-on-Foot near Miss Interactions with Construction Equipment and Geo-Referenced Hazard Area. Autom. Constr. 2015 , 60 , 58–73. [ Google Scholar ] [ CrossRef ]
• Zong, L.; Fu, G. A Study on Designing No-Penalty Reporting System about Enterprise Staff’s near Miss. Adv. Mater. Res. 2011 , 255–260 , 3846–3851. [ Google Scholar ] [ CrossRef ]
• Golovina, O.; Teizer, J.; Pradhananga, N. Heat Map Generation for Predictive Safety Planning: Preventing Struck-by and near Miss Interactions between Workers-on-Foot and Construction Equipment. Autom. Constr. 2016 , 71 , 99–115. [ Google Scholar ] [ CrossRef ]
• Zou, P.X.W.; Lun, P.; Cipolla, D.; Mohamed, S. Cloud-Based Safety Information and Communication System in Infrastructure Construction. Saf. Sci. 2017 , 98 , 50–69. [ Google Scholar ] [ CrossRef ]
• Hinze, J.; Godfrey, R. An Evaluation of Safety Performance Measures for Construction Projects. J. Constr. Res. 2011 , 4 , 5–15. [ Google Scholar ] [ CrossRef ]
• Construction Inspection Software|IAuditor by SafetyCulture. Available online: https://safetyculture.com/construction/ (accessed on 25 August 2023).
• Incident Reporting Made Easy|Safety Compliance|Mobile EHS Solutions. Available online: https://www.safety-reports.com/lp/safety/incident/ (accessed on 25 August 2023).
• Wu, F.; Wu, T.; Yuce, M.R. An Internet-of-Things (IoT) Network System for Connected Safety and Health Monitoring Applications. Sensors 2019 , 19 , 21. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Fang, W.; Luo, H.; Xu, S.; Love, P.E.D.; Lu, Z.; Ye, C. Automated Text Classification of Near-Misses from Safety Reports: An Improved Deep Learning Approach. Adv. Eng. Inform. 2020 , 44 , 101060. [ Google Scholar ] [ CrossRef ]
• Gatti, U.C.; Lin, K.-Y.; Caldera, C.; Chiang, R. Exploring the Relationship between Chronic Sleep Deprivation and Safety on Construction Sites. In Proceedings of the Construction Research Congress 2014: Construction in a Global Network, Atlanta, GA, USA, 19–24 May 2014; pp. 1772–1781. [ Google Scholar ] [ CrossRef ]
• Hon, C.K.H.; Chan, A.P.C.; Yam, M.C.H. Relationships between Safety Climate and Safety Performance of Building Repair, Maintenance, Minor Alteration, and Addition (RMAA) Works. Saf. Sci. 2014 , 65 , 10–19. [ Google Scholar ] [ CrossRef ]
• Oni, O.; Olanrewaju, A.; Cheen, K.S. Accidents at construction sites and near-misses: A constant problem. Int. Struct. Eng. Constr. 2022 , 9 , 2022. [ Google Scholar ] [ CrossRef ]
• Wu, W.; Yang, H.; Chew, D.A.S.; Yang, S.-H.; Gibb, A.G.F.; Li, Q. Towards an Autonomous Real-Time Tracking System of near-Miss Accidents on Construction Sites. Autom. Constr. 2010 , 19 , 134–141. [ Google Scholar ] [ CrossRef ]
• Aria, S.S.; Yang, K.; Ahn, C.R.; Vuran, M.C. Near-Miss Accident Detection for Ironworkers Using Inertial Measurement Unit Sensors. In Proceedings of the International Symposium on Automation and Robotics in Construction, ISARC 2014, Sydney, Australia, 9–11 July 2014; Volume 31, pp. 854–859. [ Google Scholar ] [ CrossRef ]
• Hasanzadeh, S.; Garza, J.M. de la Productivity-Safety Model: Debunking the Myth of the Productivity-Safety Divide through a Mixed-Reality Residential Roofing Task. J. Constr. Eng. Manag. 2020 , 146 , 04020124. [ Google Scholar ] [ CrossRef ]
• Teizer, J. Magnetic Field Proximity Detection and Alert Technology for Safe Heavy Construction Equipment Operation. In Proceedings of the 32nd International Symposium on Automation and Robotics in Construction, Oulu, Finland, 15–18 June 2015. [ Google Scholar ] [ CrossRef ]
• Mohajeri, M.; Ardeshir, A.; Banki, M.T.; Malekitabar, H. Discovering Causality Patterns of Unsafe Behavior Leading to Fall Hazards on Construction Sites. Int. J. Constr. Manag. 2022 , 22 , 3034–3044. [ Google Scholar ] [ CrossRef ]
• Kisaezehra; Farooq, M.U.; Bhutto, M.A.; Kazi, A.K. Real-Time Safety Helmet Detection Using Yolov5 at Construction Sites. Intell. Autom. Soft Comput. 2023 , 36 , 911–927. [ Google Scholar ] [ CrossRef ]
• Li, C.; Ding, L. Falling Objects Detection for near Miss Incidents Identification on Construction Site. In Proceedings of the ASCE International Conference on Computing in Civil Engineering, Atlanta, GA, USA, 17–19 June 2019; pp. 138–145. [ Google Scholar ] [ CrossRef ]
• Jeelani, I.; Ramshankar, H.; Han, K.; Albert, A.; Asadi, K. Real-Time Hazard Proximity Detection—Localization of Workers Using Visual Data. In Proceedings of the ASCE International Conference on Computing in Civil Engineering, Atlanta, GA, USA, 17–19 June 2019; pp. 281–289. [ Google Scholar ] [ CrossRef ]
• Lim, T.-K.; Park, S.-M.; Lee, H.-C.; Lee, D.-E. Artificial Neural Network–Based Slip-Trip Classifier Using Smart Sensor for Construction Workplace. J. Constr. Eng. Manag. 2015 , 142 , 04015065. [ Google Scholar ] [ CrossRef ]
• Yang, K.; Jebelli, H.; Ahn, C.R.; Vuran, M.C. Threshold-Based Approach to Detect Near-Miss Falls of Iron Workers Using Inertial Measurement Units. In Proceedings of the 2015 International Workshop on Computing in Civil Engineering, Austin, TX, USA, 21–23 June 2015; 2015; 2015, pp. 148–155. [ Google Scholar ] [ CrossRef ]
• Yang, K.; Ahn, C.R.; Vuran, M.C.; Aria, S.S. Semi-Supervised near-Miss Fall Detection for Ironworkers with a Wearable Inertial Measurement Unit. Autom. Constr. 2016 , 68 , 194–202. [ Google Scholar ] [ CrossRef ]
• Raviv, G.; Shapira, A.; Fishbain, B. AHP-Based Analysis of the Risk Potential of Safety Incidents: Case Study of Cranes in the Construction Industry. Saf. Sci. 2017 , 91 , 298–309. [ Google Scholar ] [ CrossRef ]
• Saurin, T.A.; Formoso, C.T.; Reck, R.; Beck da Silva Etges, B.M.; Ribeiro JL, D. Findings from the Analysis of Incident-Reporting Systems of Construction Companies. J. Constr. Eng. Manag. 2015 , 141 , 05015007. [ Google Scholar ] [ CrossRef ]
• Williams, E.; Sherratt, F.; Norton, E. Exploring the Value in near Miss Reporting for Construction Safety. In Proceedings of the 37th Annual Conference, Virtual Event, 6–10 December 2021; pp. 319–328. [ Google Scholar ]
• Baker, H.; Smith, S.; Masterton, G.; Hewlett, B. Data-Led Learning: Using Natural Language Processing (NLP) and Machine Learning to Learn from Construction Site Safety Failures. In Proceedings of the 36th Annual ARCOM Conference, Online, 7–8 September 2020; pp. 356–365. [ Google Scholar ]
• Jin, R.; Wang, F.; Liu, D. Dynamic Probabilistic Analysis of Accidents in Construction Projects by Combining Precursor Data and Expert Judgments. Adv. Eng. Inform. 2020 , 44 , 101062. [ Google Scholar ] [ CrossRef ]
• Zhou, Z.; Li, C.; Mi, C.; Qian, L. Exploring the Potential Use of Near-Miss Information to Improve Construction Safety Performance. Sustainability 2019 , 11 , 1264. [ Google Scholar ] [ CrossRef ]
• Boateng, E.B.; Pillay, M.; Davis, P. Predicting the Level of Safety Performance Using an Artificial Neural Network. Adv. Intell. Syst. Comput. 2019 , 876 , 705–710. [ Google Scholar ] [ CrossRef ]
• Zhang, M.; Cao, T.; Zhao, X. Using Smartphones to Detect and Identify Construction Workers’ Near-Miss Falls Based on ANN. J. Constr. Eng. Manag. 2018 , 145 , 04018120. [ Google Scholar ] [ CrossRef ]
• Gadekar, H.; Bugalia, N. Automatic Classification of Construction Safety Reports Using Semi-Supervised YAKE-Guided LDA Approach. Adv. Eng. Inform. 2023 , 56 , 101929. [ Google Scholar ] [ CrossRef ]
• Zhu, Y.; Liao, H.; Huang, D. Using Text Mining and Multilevel Association Rules to Process and Analyze Incident Reports in China. Accid. Anal. Prev. 2023 , 191 , 107224. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Li, M.; Lin, Q.; Jin, H. Research on Near-Miss Incidents Monitoring and Early Warning System for Building Construction Sites Based on Blockchain Technology. J. Constr. Eng. Manag. 2023 , 149 , 04023124. [ Google Scholar ] [ CrossRef ]
• Chung, W.W.S.; Tariq, S.; Mohandes, S.R.; Zayed, T. IoT-Based Application for Construction Site Safety Monitoring. Int. J. Constr. Manag. 2020 , 23 , 58–74. [ Google Scholar ] [ CrossRef ]
• Liu, X.; Xu, F.; Zhang, Z.; Sun, K. Fall-Portent Detection for Construction Sites Based on Computer Vision and Machine Learning. Eng. Constr. Archit. Manag. 2023; ahead-of-print . [ Google Scholar ] [ CrossRef ]
• Abbasi, H.; Guerrieri, A.; Lee, J.; Yang, K. Mobile Device-Based Struck-By Hazard Recognition in Construction Using a High-Frequency Sound. Sensors 2022 , 22 , 3482. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Wang, F.; Li, H.; Dong, C. Understanding Near-Miss Count Data on Construction Sites Using Greedy D-Vine Copula Marginal Regression. Reliab. Eng. Syst. Saf. 2021 , 213 , 107687. [ Google Scholar ] [ CrossRef ]
• Bugalia, N.; Tarani, V.; Student, G.; Kedia, J.; Gadekar, H. Machine Learning-Based Automated Classification Of Worker-Reported Safety Reports In Construction. J. Inf. Technol. Constr. 2022 , 27 , 926–950. [ Google Scholar ] [ CrossRef ]
• Chen, S.; Xi, J.; Chen, Y.; Zhao, J. Association Mining of Near Misses in Hydropower Engineering Construction Based on Convolutional Neural Network Text Classification. Comput. Intell. Neurosci. 2022 , 2022 , 4851615. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Tang, S.; Golparvar-Fard, M.; Naphade, M.; Gopalakrishna, M.M. Video-Based Activity Forecasting for Construction Safety Monitoring Use Cases. In Proceedings of the ASCE International Conference on Computing in Civil Engineering, Atlanta, GA, USA, 17–19 June 2019; pp. 204–210. [ Google Scholar ] [ CrossRef ]
• Rashid, K.M.; Behzadan, A.H. Risk Behavior-Based Trajectory Prediction for Construction Site Safety Monitoring. J. Constr. Eng. Manag. 2018 , 144 , 04017106. [ Google Scholar ] [ CrossRef ]
• Shen, X.; Marks, E. Near-Miss Information Visualization Tool in BIM for Construction Safety. J. Constr. Eng. Manag. 2016 , 142 , 04015100. [ Google Scholar ] [ CrossRef ]
• Erusta, N.E.; Sertyesilisik, B. An Investigation into Improving Occupational Health and Safety Performance of Construction Projects through Usage of BIM for Lean Management. In Communications in Computer and Information Science (CCIS) ; Springer: Berlin/Heidelberg, Germany, 2020; Volume 1188, pp. 91–100. [ Google Scholar ] [ CrossRef ]
• Coffland, M.M.; Kim, A.; Sadatsafavi, H.; Uber, M.M. Improved Data Storage for Better Safety Analysis and Decision Making in Large Construction Management Firms. Available online: https://www.researchgate.net/publication/320474383_Improved_Data_Storage_for_Better_Safety_Analysis_and_Decision_Making_in_Large_Construction_Management_Firms (accessed on 12 June 2024).
• Zhou, Z.; Li, Q.; Wu, W. Developing a Versatile Subway Construction Incident Database for Safety Management. J. Constr. Eng. Manag. 2011 , 138 , 1169–1180. [ Google Scholar ] [ CrossRef ]
• Wu, W.; Yang, H.; Li, Q.; Chew, D. An Integrated Information Management Model for Proactive Prevention of Struck-by-Falling-Object Accidents on Construction Sites. Autom. Constr. 2013 , 34 , 67–74. [ Google Scholar ] [ CrossRef ]
• Hoła, B. Identification and Evaluation of Processes in a Construction Enterprise. Arch. Civ. Mech. Eng. 2015 , 15 , 419–426. [ Google Scholar ] [ CrossRef ]
• Zhou, C.; Ding, L.; Skibniewski, M.J.; Luo, H.; Jiang, S. Characterizing Time Series of Near-Miss Accidents in Metro Construction via Complex Network Theory. Saf. Sci. 2017 , 98 , 145–158. [ Google Scholar ] [ CrossRef ]
• Woźniak, Z.; Hoła, B. Time Series Analysis of Hazardous Events Based on Data Recorded in a Polish Construction Company. Arch. Civ. Eng. 2024; in process . [ Google Scholar ]
• Drozd, W. Characteristics of Construction Site in Terms of Occupational Safety. J. Civ. Eng. Environ. Archit. 2016 , 63 , 165–172. [ Google Scholar ]
• Meliá, J.L.; Mearns, K.; Silva, S.A.; Lima, M.L. Safety Climate Responses and the Perceived Risk of Accidents in the Construction Industry. Saf. Sci. 2008 , 46 , 949–958. [ Google Scholar ] [ CrossRef ]
• Bugalia, N.; Maemura, Y.; Ozawa, K. A System Dynamics Model for Near-Miss Reporting in Complex Systems. Saf. Sci. 2021 , 142 , 105368. [ Google Scholar ] [ CrossRef ]
• Gyi, D.E.; Gibb, A.G.F.; Haslam, R.A. The Quality of Accident and Health Data in the Construction Industry: Interviews with Senior Managers. Constr. Manag. Econ. 1999 , 17 , 197–204. [ Google Scholar ] [ CrossRef ]
• Menzies, J. Structural Safety: Learning and Warnings. Struct. Eng. 2002 , 80 , 15–16. [ Google Scholar ]
• Fullerton, C.E.; Allread, B.S.; Teizer, J. Pro-Active-Real-Time Personnel Warning System. In Proceedings of the Construction Research Congress 2009: Building a Sustainable Future, Seattle, WA, USA, 5–7 April 2009; pp. 31–40. [ Google Scholar ] [ CrossRef ]
• Marks, E.D.; Wetherford, J.E.; Teizer, J.; Yabuki, N. Potential of Leading Indicator Data Collection and Analysis for Proximity Detection and Alert Technology in Construction. In Proceedings of the 30th ISARC—International Symposium on Automation and Robotics in Construction Conference, Montreal, QC, Canada, 11–15 August 2013; pp. 1029–1036. [ Google Scholar ] [ CrossRef ]
• Martin, H.; Lewis, T.M. Pinpointing Safety Leadership Factors for Safe Construction Sites in Trinidad and Tobago. J. Constr. Eng. Manag. 2014 , 140 , 04013046. [ Google Scholar ] [ CrossRef ]
• Hobson, P.; Emery, D.; Brown, L.; Bashford, R.; Gill, J. People–Plant Interface Training: Targeting an Industry Fatal Risk. Proc. Inst. Civ. Eng. Civ. Eng. 2014 , 167 , 138–144. [ Google Scholar ] [ CrossRef ]
• Marks, E.; Mckay, B.; Awolusi, I. Using near Misses to Enhance Safety Performance in Construction. In Proceedings of the ASSE Professional Development Conference and Exposition, Dallas, TX, USA, 7–10 June 2015. [ Google Scholar ]
• Popp, J.D.; Scarborough, M.S. Investigations of near Miss Incidents—New Facility Construction and Commissioning Activities. IEEE Trans. Ind. Appl. 2016 , 53 , 615–621. [ Google Scholar ] [ CrossRef ]
• Nickel, P.; Lungfiel, A.; Trabold, R.J. Reconstruction of near Misses and Accidents for Analyses from Virtual Reality Usability Study. In Lecture Notes in Computer Science ; Springer: Berlin/Heidelberg, Germany, 2017; Volume 10700, pp. 182–191. [ Google Scholar ] [ CrossRef ]
• Gambatese, J.A.; Pestana, C.; Lee, H.W. Alignment between Lean Principles and Practices and Worker Safety Behavior. J. Constr. Eng. Manag. 2017 , 143 , 04016083. [ Google Scholar ] [ CrossRef ]
• Van Voorhis, S.; Korman, R. Reading Signs of Trouble. Eng. News-Rec. 2017 , 278 , 14–17. [ Google Scholar ]
• Doan, D.R. Investigation of a near-miss shock incident. IEEE Trans. Ind. Appl. 2016 , 52 , 560–561. [ Google Scholar ] [ CrossRef ]
• Oswald, D.; Sherratt, F.; Smith, S. Problems with safety observation reporting: A construction industry case study. Saf. Sci. 2018 , 107 , 35–45. [ Google Scholar ] [ CrossRef ]
• Raviv, G.; Shapira, A. Systematic approach to crane-related near-miss analysis in the construction industry. Int. J. Constr. Manag. 2018 , 18 , 310–320. [ Google Scholar ] [ CrossRef ]
• Whiteoak, J.; Appleby, J. Mate, that was bloody close! A case history of a nearmiss program in the Australian construction industry. J. Health Saf. Environ. 2019 , 35 , 31–43. [ Google Scholar ]
• Duryan, M.; Smyth, H.; Roberts, A.; Rowlinson, S.; Sherratt, F. Knowledge transfer for occupational health and safety: Cultivating health and safety learning culture in construction firms. Accid. Anal. Prev. 2020 , 139 , 105496. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Shaikh, A.Y.; Osei-Kyei, R.; Hardie, M. A critical analysis of safety performance indicators in construction. Int. J. Build. Pathol. Adapt. 2020 , 39 , 547–580. [ Google Scholar ] [ CrossRef ]
• Martin, H.; Mohan, N.; Ellis, L.; Dunne, S. Exploring the Role of PPE Knowledge, Attitude, and Correct Practices in Safety Outcomes on Construction Sites. J. Archit. Eng. 2021 , 27 , 05021011. [ Google Scholar ] [ CrossRef ]
• Qin, Z.; Wu, S. A simulation model of engineering construction near-miss event disclosure strategy based on evolutionary game theory. In Proceedings of the 2021 4th International Conference on Information Systems and Computer Aided Education, Dalian, China, 24–26 September 2021; pp. 2572–2577. [ Google Scholar ] [ CrossRef ]
• Alamoudi, M. The Integration of NOSACQ-50 with Importance-Performance Analysis Technique to Evaluate and Analyze Safety Climate Dimensions in the Construction Sector in Saudi Arabia. Buildings 2022 , 12 , 1855. [ Google Scholar ] [ CrossRef ]
• Herrmann, A.W. Development of CROSS in the United States. In Proceedings of the Forensic Engineering 2022: Elevating Forensic Engineering—Selected Papers from the 9th Congress on Forensic Engineering, Denver, Colorado, 4–7 November 2022; Volume 2, pp. 40–43. [ Google Scholar ] [ CrossRef ]
• Al Shaaili, M.; Al Alawi, M.; Ekyalimpa, R.; Al Mawli, B.; Al-Mamun, A.; Al Shahri, M. Near-miss accidents data analysis and knowledge dissemination in water construction projects in Oman. Heliyon 2023 , 9 , e21607. [ Google Scholar ] [ CrossRef ] [ PubMed ]
• Agnusdei, G.P.; Gnoni, M.G.; Tornese, F.; De Merich, D.; Guglielmi, A.; Pellicci, M. Application of Near-Miss Management Systems: An Exploratory Field Analysis in the Italian Industrial Sector. Safety 2023 , 9 , 47. [ Google Scholar ] [ CrossRef ]
• Duan, P.; Zhou, J. A science mapping approach-based review of near-miss research in construction. Eng. Constr. Archit. Manag. 2023 , 30 , 2582–2601. [ Google Scholar ] [ CrossRef ]

Click here to enlarge figure

Share and Cite

Woźniak, Z.; Hoła, B. Analysing Near-Miss Incidents in Construction: A Systematic Literature Review. Appl. Sci. 2024 , 14 , 7260. https://doi.org/10.3390/app14167260

Woźniak Z, Hoła B. Analysing Near-Miss Incidents in Construction: A Systematic Literature Review. Applied Sciences . 2024; 14(16):7260. https://doi.org/10.3390/app14167260

Woźniak, Zuzanna, and Bożena Hoła. 2024. "Analysing Near-Miss Incidents in Construction: A Systematic Literature Review" Applied Sciences 14, no. 16: 7260. https://doi.org/10.3390/app14167260

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

Global AI adoption is outpacing risk understanding, warns MIT CSAIL

As organizations rush to implement artificial intelligence (AI), a new analysis of AI-related risks finds significant gaps in our understanding, highlighting an urgent need for a more comprehensive approach.

The adoption of AI is rapidly increasing; census data shows a significant (47%) rise in AI usage within US industries, jumping from 3.7% to 5.45% between September 2023 and February 2024. However, a comprehensive review from researchers at MIT CSAIL  and  MIT FutureTech has uncovered critical gaps in existing AI risk frameworks. Their analysis reveals that even the most thorough individual framework overlooks approximately 30% of the risks identified across all reviewed frameworks.

To help address this, they have collaborated with colleagues from the University of Queensland , Future of Life Institute , KU Leuven , and Harmony Intelligence , to release the first-ever AI Risk Repository : a comprehensive and accessible living database of 700+ risks posed by AI that will be expanded and updated to ensure that it remains current and relevant.

“Since the AI risk literature is scattered across peer-reviewed journals, preprints, and industry reports, and quite varied, I worry that decision-makers may unwittingly consult incomplete overviews, miss important concerns, and develop collective blind spots,” says Dr. Peter Slattery , an incoming postdoc at the MIT FutureTech Lab and current project lead.

After searching several academic databases, engaging experts, and retrieving more than 17,000 records, the researchers identified 43 existing AI risk classification frameworks. From these, they extracted more than 700 risks. They then used approaches that they developed from two existing frameworks to categorize each risk by cause (e.g., when or why it occurs), risk domain (e.g., “Misinformation”), and risk subdomain (e.g., “False or misleading information”). 

Examples of risks identified include “Unfair discrimination and misrepresentation”, “Fraud, scams, and targeted manipulation”, and “Overreliance and unsafe use.” More of the risks analyzed were attributed to AI systems (51%) than humans (34%) and presented as emerging after AI was deployed (65%) rather than during its development (10%).  The most frequently addressed risk domains included “AI system safety, failures, and limitations” (76% of documents); “Socioeconomic and environmental harms” (73%); “Discrimination and toxicity” (71%); “Privacy and security” (68%); and “Malicious actors and misuse” (68%). In contrast, “Human-computer interaction” (41%) and “Misinformation” (44%) received comparatively less attention. 

Some risk subdomains were discussed more frequently than others. For example, “Unfair discrimination and misrepresentation” (63%), “Compromise of privacy” (61%), and “Lack of capability or robustness” (59%), were mentioned in more than 50% of documents. Others, like “AI welfare and rights” (2%), “Pollution of information ecosystem and loss of consensus reality” (12%), and “Competitive dynamics” (12%), were mentioned by less than 15% of documents. 

On average, frameworks mentioned just 34% of the 23 risk subdomains identified, with nearly a quarter covering less than 20%. No document or overview mentioned all 23 risk subdomains, and the most comprehensive ( Gabriel et al., 2024 ) covered only 70%. 

The work addresses the urgent need to help decision-makers in government, research, and industry understand and prioritize the risks associated with AI and work together to address them. “Many AI governance initiatives are emerging across the world focused on addressing key risks from AI,” says collaborator Risto Uuk , EU Research Lead at the Future of Life Institute. “These institutions need a more comprehensive and complete understanding of the risk landscape.” 

Researchers and risk evaluation professionals are also impeded by the fragmentation of current literature . "It is hard to find specific studies of risk in some niche domains where AI is used, such as weapons and military decision support systems,” explains Taniel Yusef , a Research Affiliate, at The Centre for the Study of Existential Risk, at the University of Cambridge who was not involved in the research. “Without referring to these studies, it can be difficult to speak about technical aspects of AI risk to non-technical experts. This repository helps us do that.”

"There's a significant need for a comprehensive database of risks from advanced AI which safety evaluators like Harmony Intelligence can use to identify and catch risks systematically,” argues collaborator Soroush Pour , CEO & Co-founder of AI safety evaluations and red teaming company Harmony Intelligence. “Otherwise, it’s unclear what risks we should be looking for, or what tests need to be done. It becomes much more likely that we miss something by simply not being aware of it”.

AI’s Risky Business 

The researchers built on two frameworks ( Yampolskiy 2016 & Weidinger et al., 2022 ) in categorizing the risks they extracted.   Based on these approaches, they group the risks in two ways.

First by causal factors: 

• Entity: Human, AI, and Other; 
• Intentionality: Intentional, Unintentional, and Other; and 
• Timing: Pre-deployment; Post-deployment, and Other. 

Second, by seven AI risk domains: 

• Discrimination & toxicity, 
• Privacy & security, 
• Misinformation, 
• Malicious actors & misuse, 
• Human-computer interaction, 
• Socioeconomic & environmental, and 
• AI system safety, failures, & limitations. 

These are further divided into 23 subdomains (full descriptions here ):

• 1.1. Unfair discrimination and misrepresentation
• 1.2. Exposure to toxic content
• 1.3. Unequal performance across groups
• 2.1. Compromise of privacy by leaking or correctly inferring sensitive information
• 2.2. AI system security vulnerabilities and attacks
• 3.1. False or misleading information
• 3.2. Pollution of the information ecosystem and loss of consensus reality
• 4.1. Disinformation, surveillance, and influence at scale
• 4.2. Cyberattacks, weapon development or use, and mass harm
• 4.3. Fraud, scams, and targeted manipulation
• 5.1. Overreliance and unsafe use
• 5.2. Loss of human agency and autonomy
• 6.1. Power centralization and unfair distribution of benefits
• 6.2. Increased inequality and decline in employment quality
• 6.3. Economic and cultural devaluation of human effort
• 6.4. Competitive dynamics
• 6.5. Governance failure
• 6.6. Environmental harm
• 7.1. AI pursuing its own goals in conflict with human goals or values
• 7.2. AI possessing dangerous capabilities
• 7.3. Lack of capability or robustness
• 7.4. Lack of transparency or interpretability
• 7.5. AI welfare and rights

"The AI Risk Repository is, to our knowledge, the first attempt to rigorously curate, analyze, and extract AI risk frameworks into a publicly accessible, comprehensive, extensible, and categorized risk database. It is part of a larger effort to understand how we are responding to AI risks and to identify if there are gaps in our current approaches," says Dr. Neil Thompson , head of the MIT FutureTech Lab and one of the lead researchers on the project. "We are starting with a comprehensive checklist, to help us understand the breadth of potential risks. We plan to use this to identify shortcomings in organizational responses. For instance, if everyone focuses on one type of risk while overlooking others of similar importance, that's something we should notice and address."

The next phase will involve experts evaluating and prioritizing the risks within the repository, then using it to analyze public documents from influential AI developers and large companies. The analysis will examine if organizations respond to risks from AI -- and do so in proportion to experts’ concerns -- and compare risk management approaches across different industries and sectors.

The repository is freely available online to download , copy , and use. Feedback and suggestions can be submitted here .

Neil Thompson

Research areas, impact areas, press contact, rachel gordon, related news.

From recurrent networks to GPT-4: Measuring algorithmic progress in language models

Rethinking AI's impact: MIT CSAIL study reveals economic limits to job automation

• Open access
• Published: 13 August 2024

Application of artificial intelligence in dental crown prosthesis: a scoping review

• Hyun-Jun Kong 1 &
• Yu-Lee Kim 2  

BMC Oral Health volume  24 , Article number:  937 ( 2024 ) Cite this article

165 Accesses

Metrics details

In recent years, artificial intelligence (AI) has made remarkable advancements and achieved significant accomplishments across the entire field of dentistry. Notably, efforts to apply AI in prosthodontics are continually progressing. This scoping review aims to present the applications and performance of AI in dental crown prostheses and related topics.

We conducted a literature search of PubMed, Scopus, Web of Science, Google Scholar, and IEEE Xplore databases from January 2010 to January 2024. The included articles addressed the application of AI in various aspects of dental crown treatment, including fabrication, assessment, and prognosis.

The initial electronic literature search yielded 393 records, which were reduced to 315 after eliminating duplicate references. The application of inclusion criteria led to analysis of 12 eligible publications in the qualitative review. The AI-based applications included in this review were related to detection of dental crown finish line, evaluation of AI-based color matching, evaluation of crown preparation, evaluation of dental crown designed by AI, identification of a dental crown in an intraoral photo, and prediction of debonding probability.

Conclusions

AI has the potential to increase efficiency in processes such as fabricating and evaluating dental crowns, with a high level of accuracy reported in most of the analyzed studies. However, a significant number of studies focused on designing crowns using AI-based software, and these studies had a small number of patients and did not always present their algorithms. Standardized protocols for reporting and evaluating AI studies are needed to increase the evidence and effectiveness.

Peer Review reports

Artificial intelligence (AI) refers to the capability of computers to perform tasks that normally require human intelligence, such as learning, reasoning, and problem-solving. In medicine, AI is used to identify and categorize pathologies in radiographs and photographs, predict clinical events, and simulate drug interactions with targets [ 1 , 2 , 3 ] Similar to its transformative impact in medicine, AI research has surged in dentistry, demonstrating its potential in a variety of areas including diagnosis, prevention, and treatment [ 4 , 5 , 6 ]. In dentistry, AI is applied in several diagnostic tasks: identifying dental caries from radiographs [ 7 , 8 ], assessing the complexity of endodontic cases [ 9 , 10 ], automating cephalometric landmark localization [ 11 ], and classifying dental implant systems [ 12 , 13 , 14 ]. The majority of AI applications is based on machine learning, a methodology where mathematical models are trained to recognize statistical patterns within datasets to perform predictions. A subset of machine learning known as deep learning employs multi-layered neural networks with intricate architecture, allowing deep learning algorithms that often surpass other machine learning strategies in discerning patterns within vast and varied datasets [ 15 ]. This is particularly advantageous in the field of dentistry, where datasets often encompass images, proteomic information, and clinical data [ 16 ].

Prosthodontics stands at the intersection of artistic expression and scientific principles within the field of dentistry [ 17 ]. It constitutes the art and science involved in the diagnosis, strategic planning, rehabilitation, and preservation of the functional, comfortable, aesthetically pleasing, and healthy aspects of the oral structures in patients. Its primary objective is the replacement of absent teeth and related structures through the integration of artificial substitutes. The application of AI holds considerable promise across various therapeutic modalities within this domain [ 18 ].

The implementation of AI is significantly impacting the creation of dental crown prostheses, a fundamental component of prosthodontic dentistry. The dental crown prosthesis plays an integral role in reinstating both the structural integrity and functionality of teeth affected by damage or decay, offering patients solutions that are both visually appealing and durable. The integration of AI in this field holds the potential to transform the processes involved in designing, manufacturing, and placing dental crowns.

Despite increasing publications on deep learning in dental crown prostheses, there is uncertainty about the prevalent tasks, preferred methods, and performance variations. This scoping review aims to evaluate studies on deep learning applications in dental crown prostheses, addressing these uncertainties

The workflow of this study is shown in Fig. 1 . We followed the reporting recommendations specified in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Scoping Reviews (PRISMA-ScR) guidelines [ 19 ] to convey the findings of this investigation. The specific review question was the applications and performance of AI in dental crown prostheses.

The workflow of the study consists of three steps: database search, screening, and literature review

Our literature exploration was based on the PICO (problem/patient/population, intervention/indicator, comparison, and outcome) elements [ 20 ].

Population: Images and other data types for dental crown prostheses used in prosthodontic rehabilitation.

Intervention/Comparison: AI models for diagnosis, prognosis assessment, and treatment procedures compared to reference standards.

Outcome: Any type of performance measurement.

Literature search

An electronic literature search was conducted in January 2024 across five databases: PubMed, Scopus, Web of Science, Google Scholar, and IEEE Xplore. We focused on literature published after 2010 to capture the most relevant and recent advancements in deep learning applications in dentistry. Manual searches and searches for gray literature were not conducted.

Search terms were developed and combined by two reviewers (H.J.K. and Y.L.K.) with reference to previous studies [ 16 , 18 ]. Each category consists of a combination of Medical Subject Headings (MeSH) Terms and related dental terms with conjunctions: (“dental crown” OR “crown preparation” OR “fixed prosthesis” OR “prosthodontic” OR “dental prosthesis”) and (“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network”).

Eligibility criteria

The following inclusion criteria were employed in the selection of articles:

Articles related to AI applications in dental crown prosthesis

Articles composed in English and released between January 2010 and January 2024

Our exclusion criteria were:

Articles that used AI for conditions not related to dental crown prosthesis

Articles that did not report performance metrics such as accuracy

Articles without the full text available

Review articles and letters to the editor

Two reviewers (H.J.K. and Y.L.K.) assessed the titles and abstracts of articles based on pre-established inclusion and exclusion criteria. When titles and abstracts were deemed insufficient in providing necessary information, a thorough analysis of the entire text was conducted. Both reviewers diligently examined articles that were potentially relevant and collaboratively selected papers for further analysis. Any discrepancies in opinions were resolved through discussion.

Selection of sources

The flowchart in Fig. 2 outlines the article selection process adhering to PRISMA-ScR for this scoping review. Initially, a search of electronic literature produced 393 records, which was decreased to 315 after eliminating duplicate references. Following the examination of titles and abstracts, 26 studies underwent a more detailed review, after which 14 records were excluded for not meeting the inclusion criteria. Finally, 12 eligible publications were included in the qualitative review (Table 1 ). The evaluators unanimously agreed on the literature selection and categorization of publications.

PRISMA-ScR flowchart of study selection

Characteristics of the included studies

The AI-based applications included in this review were related to detection of dental crown finish line, evaluation of AI-based color matching, evaluation of crown preparation, evaluation of dental crown designed by AI, identification of dental crowns in intraoral photos, and prediction of debonding probability (Fig. 3 ). While the search was initially planned to encompass articles published between January 2010 and January 2024, the findings indicate a surge in the popularity of AI applications starting in 2019.

Number of included articles by publication year and purpose of artificial intelligence application

The dataset size exhibited a wide range owing to the diversity in inputs and outcomes across study types (Table 2 ). Dataset size for included studies ranged from 12 to 8640. Studies using AI software included a small number of data (mean = 24.25). Since these studies do not learn new models, they do not require large datasets like existing deep learning studies. The most common types of data used were scanned digital images of the jaw and casts ( n =10), and two studies used intraoral photographs.

AI architecture

Three types of AI architecture were used in this body of literature. AI software was used most often (5 times). CNN was used four times, and GAN was used three times. Among AI software, CNN and GAN algorithms were used in two studies. However, in the remaining three documents, the name of the software was provided, although the specific algorithm was not disclosed.

Excluding studies where the algorithm could not be confirmed, in designs using AI, the program used only 3D-GAN or used it in combination with CNN. In other types of studies, CNNs, which are widely used in image classification and object detection, were used. In two studies using a combination of CNN and GAN, CNN was used to extract the preparation tooth and set the margin line, and then GAN was used to create the outer surface.

Outcome metrics

With a scoping review, there is considerable heterogeneity in data forms and methodologies, leading to diverse outcome metrics. In studies evaluating AI design, RMS was used as an evaluation indicator in six of the studies. In these studies, mean deviation to evaluate positional accuracy and cusp angle to evaluate crown shape were used as evaluation indicators. Working time was also included in two studies to compare with existing programs or traditional methods. When an object detection algorithm was used, Intersection over union (IoU) was used as an evaluation index. Accuracy, precision, recall, and F-score were used in one prediction study.

According to this scoping review, an increasing number of studies have employed AI for various tasks related to dental crown prostheses. To align with the rapid evolution of digital technologies, the investigation focused on the most recent 14 years. The application of AI in the fabrication and evaluation of dental crowns can contribute to increased efficiency for prosthodontists, raising expectations for improved productivity. The application of AI to the following fields related to dental crown prosthesis was reviewed.

Dental crown prosthesis designed by AI

The production of crown prostheses has been separated into traditional wax-up methods and digital approaches using CAD/CAM systems. Recently, there has been a widespread adoption of CAD/CAM methodologies, particularly in conjunction with the popularization of zirconia usage. Notably, these methods report high levels of accuracy and success rates [ 33 ]. In the CAD/CAM process, there have been efforts to enhance the speed and precision of prosthesis design through the application of AI [ 34 ].

Seven studies evaluated the feasibility of AI models to design dental crown prostheses [21,22,25,28,29,31,32. Four studies used AI software [ 25 , 28 , 31 , 32 ], while three studies [ 21 , 22 , 29 ] employed 3D-GAN algorithms for model training. In six studies, integration of AI into the fabrication process demonstrated higher accuracy compared with traditional CAD or manually designed methods. One study [ 28 ] reported that knowledge-based AI, compared to human-designed CAD software, exhibited a higher occlusal profile discrepancy.

Two studies [ 25 , 31 ] reported high time efficiency of AI-based programs. While traditional CAD work does not involve adding and modifying wax, it requires human thinking. Both wax-up and CAD processes are heavily influenced by the accumulated experience of dental technicians [ 35 , 36 ]. However, AI-based operations minimize human intervention using automated calculations. Therefore, designing with AI significantly reduces the time spent on the design process, especially in cases with extensive restoration requirements.

Among the seven studies, four used commercially produced AI software, with two of them lacking specific mention of the program's algorithm [ 25 , 28 ]. The absence of detailed information on the algorithms and metrics used for training the models in these studies acts as a limitation, emphasizing the need for transparency in the application of AI in dentistry, especially with commercially available software. Additionally, studies employing software had datasets ranging from 12 to 30 cases, posing a limitation in detecting statistical significance due to the potentially insufficient sample size. Despite these limitations, AI for dental crown design has the potential to significantly increase production efficiency by saving time. Considering the performance demonstrated in aspects such as morphology, internal fit, and occlusion, there is a promising outlook for the future utilization of AI in dentistry.

Detection of dental crown finish line

Choi et al. [ 23 ] compared the accuracy of hybrid software combined with deep learning with existing traditional CAD software in detecting the crown finish line. An accurate marginal fit is crucial for preventing microgaps. This, in turn, lowers the risk of caries and ensures that the restoration retains its function [ 37 ]. Recently, finish lines have been extracted and processed manually using CAD programs, but this is a repetitive and time-consuming process [ 38 ]. In Choi et al., as a result of evaluation using Hausdorff distance and chamfer distance, the hybrid system showed statistically more accurate results. This implies that a hybrid approach, integrating both deep learning and computer-aided design methods, may allow robust and precise extraction of finish lines with minimal adjustments required.

Evaluation of crown preparation

One of the most fundamental aspects in dental prosthodontics education is understanding and practicing the principles of tooth preparation [ 39 ]. However, evaluating students' tooth preparation outcomes in dental education can lack consistency due to factors such as subjective grading scales and insufficient inter-rater agreement. This difficulty hinders the provision of ongoing and reliable feedback [ 40 , 41 ].

Han et al. [ 27 ] assessed the viability of software-based automated evaluation (SAE) with AI to evaluate abutment tooth preparation for single crowns. This was done through a comparison with a human-based digitally-assisted evaluation (DAE), which showed perfect intra-rater agreement and almost perfect inter-rater agreement with SAE. The findings of this study substantiate the credibility of SAE within prosthodontics education and propose its potential clinical utility for evaluating tooth preparation

Evaluation of AI-based color matching

A crucial aspect of the dental technician's role is replicating the natural color of teeth in dental prostheses. An experienced dental technician possesses the ability to precisely assess the authentic color. However, this proves to be a challenging task for a less-experienced dental technician [ 42 ].

Ueki et al. [ 26 ] extracted 62 images of patient teeth, which were annotated by experienced dental technicians. They then used a neural network to estimate the true color. The accuracy of the first candidate's output was six of 22 (27%), considerably lower than the desired level. However, the outputs for the second and third candidates encompassed 12 (55%) and 15 (68%) of the total 22 images, respectively. This affirmed accurate classification of certain colors.

One notable limitation of this study is the relatively small size of the image dataset used. To more accurately assess the potential of AI in shade selection, a substantial amount of training data is required.

Identification of dental crown in intraoral photo

In clinical situations, it is crucial for dentists to gather intraoral information about patients—a process that demands time and effort. Additionally, the effectiveness of this procedure relies on the dentist's knowledge and experience. Consequently, there is a demand for an automated system that can rapidly assess the intraoral situation.

Takahashi et al. [ 24 ] used a deep learning object detection method to recognize dental prostheses and restorations. In their study, ‘You Only Look Once version 3’ (YOLOv3) was used for object detection because it has shown high performance in other dental deep learning studies [ 43 , 44 , 45 ].

A satisfactory level of performance is typically associated with an IoU exceeding 0.7 [ 46 , 47 ]. In the present investigation, the IoU was 0.76. Consequently, the proficiency of this learning system was high. In assessing the accuracy of the object detection model, the mAP is employed, with values above 0.7 considered favorable in previous research [ 48 ]. The mAP achieved in the present study was 0.80, supporting the learning system's commendable performance from a mAP perspective.

Irrespective of the overall count of objects across all images, there was a tendency for higher average precision (AP) scores in cases of metallic-colored prostheses, while tooth-colored prostheses exhibited a tendency toward lower AP scores. These findings suggest that the identification was influenced by the color distinctions between the natural teeth and prostheses.

Prediction of debonding probability

CAD/CAM composite resin (CR) crowns cemented to dentin often exhibit a propensity for debonding within one year, and the reported debonding rate for CAD/CAM CR crowns cemented on implant abutments stands at 80% within one year [ 49 ]. Inadequate preparation has been identified to contribute to debonding [ 50 , 51 ].

Yamaguchi et al. [ 30 ] aimed to predict the debonding probability of CAD/CAM CR crowns using scanned images of prepared models employing convolutional neural networks. The reported prediction accuracy was 98.5%. Despite the good performance, this study acknowledges a limitation in explaining the primary factor contributing to debonding, stating that it was difficult to pinpoint the main cause.

Limitations

This study has several limitations. First, as a scoping review, it provides an overview of the current application of AI in dental crown prosthetics and assesses the need for further research, but it does not establish the reliability of existing knowledge as a systematic review would. Second, the reviewed literature was classified into six categories, but half of the articles are included in the category of "evaluation of dental crowns designed by AI," with only one article in each of the other categories. Lastly, some of the articles included in the review lack transparency as the algorithms used are not clearly specified.

In addition to the articles reviewed in this study, there are an increasing number of studies integrating next-generation technologies, such as robotics, into dental crown prosthesis [ 52 , 53 ]. The advancement of these technologies, including AI, will bring continuous progress to the entire field of prosthodontics.

The number of studies applying AI to dental crown prostheses is gradually increasing. According to the results of this review, AI has the potential to increase efficiency in processes such as fabricating and evaluating dental crowns, with a high level of accuracy reported in most of the included studies. However, a significant number of studies focused on designing crowns using AI-based software, and these studies have limitations in that the study size was small and the algorithms used were sometimes not disclosed. The small number of AI-related studies means that AI application to dental crowns is needed in more diverse aspects. Additionally, research involving a large number of patients and data in various clinical situations is needed. AI study, by its nature, uses a variety of methods and evaluation metrics. Therefore, standardized protocols for reporting and evaluating AI studies are needed to increase the evidence and effectiveness.

Availability of data and materials

All data generated or analysed during this study are included in this published article.

Abbreviations

• Artificial intelligence

Computer-aided design

Computer-aided design/computer-aided manufacturing

Convolutional neural network

Crown composite resin crown

Digitally-assisted evaluation

Deep convolutional generative adversarial network

Two-stage deep generative adversarial network

Generative adversarial network

Software-based automated evaluation

Intersection over union

Root mean square error

Briganti G, Le Moine O. Artificial Intelligence in Medicine: Today and Tomorrow. Front Med (Lausanne). 2020;7:27.

Article   PubMed   Google Scholar  

Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44–56.

Article   CAS   PubMed   Google Scholar  

Paul D, Sanap G, Shenoy S, Kalyane D, Kalia K, Tekade RK. Artificial intelligence in drug discovery and development. Drug Discov Today. 2021;26:80–93.

Schwendicke F, Samek W, Krois J. Artificial intelligence in dentistry: chances and challenges. J Dent Res. 2020;99:769–74.

Ahmed N, Abbasi MS, Zuberi F, Qamar W, Halim MSB, Maqsood A, et al. Artificial intelligence techniques: analysis, application, and outcome in dentistry-a systematic review. Biomed Res Int. 2021;9751564.  https://doi.org/10.1155/2021/9751564 .

Shan T, Tay FR, Gu L. Application of artificial intelligence in dentistry. J Dent Res. 2021;100:232–44.

Lee JH, Kim DH, Jeong SN, Choi SH. Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm. J Dent. 2018;77:106–11.

Zhang X, Liang Y, Li W, Liu C, Gu D, Sun W, et al. Development and evaluation of deep learning for screening dental caries from oral photographs. Oral Dis. 2022;28:173–81.

Aminoshariae A, Kulild J, Nagendrababu V. Artificial intelligence in endodontics: current applications and future directions. J Endod. 2021;47:1352–7.

Ahmed ZH, Almuharib AM, Abdulkarim AA, Alhassoon AH, Alanazi AF, Alhaqbani MA, et al. Artificial intelligence and its application in endodontics: a review. J Contemp Dent Pract. 2023;24:912–7.

Nishimoto S, Sotsuka Y, Kawai K, Ishise H, Kakibuchi M. Personal computer-based cephalometric landmark detection with deep learning, using cephalograms on the internet. J Craniofac Surg. 2019;30:91–5.

Sukegawa S, Yoshii K, Hara T, Yamashita K, Nakano K, Yamamoto N, et al. Deep neural networks for dental implant system classification. Biomolecules. 2020;10(7):984.

Article   CAS   PubMed   PubMed Central   Google Scholar  

Kong HJ, Eom SH, Yoo JY, Lee JH. Identification of 130 dental implant types using ensemble deep learning. Int J Oral Maxillofac Implants. 2023;38:150–6.

Alqutaibi AY, Algabri RS, Elawady D, Ibrahim WI. Advancements in artificial intelligence algorithms for dental implant identification: A systematic review with meta-analysis. J Prosthet Dent. 2023;S0022–3913(23):00783–7.

Google Scholar  

Castiglioni I, Rundo L, Codari M, Di Leo G, Salvatore C, Interlenghi M, et al. AI applications to medical images: from machine learning to deep learning. Phys Med. 2021;83:9–24.

Arsiwala-Scheppach LT, Chaurasia A, Müller A, Krois J, Schwendicke F. Machine learning in dentistry: a scoping review. J Clin Med. 2023;12:937.

Article   PubMed   PubMed Central   Google Scholar  

Blatz MB, Chiche G, Bahat O, Roblee R, Coachman C, Heymann HO. Evolution of aesthetic dentistry. J Dent Res. 2019;98:1294–304.

Revilla-León M, Gómez-Polo M, Vyas S, Barmak AB, Gallucci GO, Att W, et al. Artificial intelligence models for tooth-supported fixed and removable prosthodontics: a systematic review. J Prosthet Dent. 2023;129:276–92.

Tricco AC, Lillie E, Zarin W, et al. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med. 2018;169:467–73.

Eriksen MB, Frandsen TF. The impact of patient, intervention, comparison, outcome (PICO) as a search strategy tool on literature search quality: a systematic review. J Med Libr Assoc. 2018;106:420–31.

Chau RCW, Hsung RT, McGrath C, Pow EHN, Lam WYH. Accuracy of artificial intelligence-designed single-molar dental prostheses: a feasibility study. J Prosthet Dent. 2023;S0022–3913(22):00764–8.

Tian S, Wang M, Dai N, Ma H, Li L, Fiorenza L, et al. Dcpr-gan: dental crown prosthesis restoration using two-stage generative adversarial networks. IEEE J Biomed Health Inform. 2022;26:151–60.

Choi J, Ahn J, Park JM. Deep learning-based automated detection of the dental crown finish line: an accuracy study. J Prosthet Dent. 2023;S0022–3913(23):00769–2.

Takahashi T, Nozaki K, Gonda T, Mameno T, Ikebe K. Deep learning-based detection of dental prostheses and restorations. Sci Rep. 2021;11:1960.

Liu CM, Lin WC, Lee SY. Evaluation of the efficiency, trueness, and clinical application of novel artificial intelligence design for dental crown prostheses. Dent Mater. 2024;40:19–27.

Ueki K, Wakamatsu H, Hagiwara Y. Evaluation of dental prosthesis colors using a neural network. 2020 IEEE 5th International Conference on Signal and Image Processing, Nanjing, China. 2020;210–214. https://ieeexplore.ieee.org/document/9339381 . Accessed 15 Feb 2024.

Han S, Yi Y, Revilla-León M, Yilmaz B, Yoon HI. Feasibility of software-based assessment for automated evaluation of tooth preparation for dental crown by using a computational geometric algorithm. Sci Rep. 2023;13:11847.

Chen Y, Lee JKY, Kwong G, Pow EHN, Tsoi JKH. Morphology and fracture behavior of lithium disilicate dental crowns designed by human and knowledge-based AI. J Mech Behav Biomed Mater. 2022;131:105256.

Ding H, Cui Z, Maghami E, Chen Y, Matinlinna JP, Pow EHN, et al. Morphology and mechanical performance of dental crown designed by 3D-DCGAN. Dent Mater. 2023;39:320–32.

Yamaguchi S, Lee C, Karaer O, Ban S, Mine A, Imazato S. Predicting the debonding of cad/cam composite resin crowns with ai. J Dent Res. 2019;98:1234–8.

Cho JH, Yi Y, Choi J, Ahn J, Yoon HI, Yilmaz B. Time efficiency, occlusal morphology, and internal fit of anatomic contour crowns designed by dental software powered by generative adversarial network: a comparative study. J Dent. 2023;138:104739.

Cho JH, Çakmak G, Yi Y, Yoon HI, Yilmaz B, Schimmel M. Tooth morphology, internal fit, occlusion and proximal contacts of dental crowns designed by deep learning-based dental software: a comparative study. J Dent. 2024;141:104830.

Leitão CIMB, Fernandes GVO, Azevedo LPP, Araújo FM, Donato H, Correia ARM. Clinical performance of monolithic CAD/CAM tooth-supported zirconia restorations: systematic review and meta-analysis. J Prosthodont Res. 2022;66:374–84.

Bernauer SA, Zitzmann NU, Joda T. The use and performance of artificial intelligence in prosthodontics: a systematic review. Sensors (Basel). 2021;21:6628.

Teng TY, Wu JH, Lee CY. Acceptance and experience of digital dental technology, burnout, job satisfaction, and turnover intention for Taiwanese dental technicians. BMC Oral Health. 2022;22:342.

Litzenburger AP, Hickel R, Richter MJ, Mehl AC, Probst FA. Fully automatic CAD design of the occlusal morphology of partial crowns compared to dental technicians’ design. Clin Oral Invest. 2013;17:491–6.

Article   Google Scholar  

Larson TD. The clinical significance of marginal fit. Northwest Dent J. 2012;91:22.

Mai HN, Han JS, Kim HS, Park YS, Park JM, Lee DH. Reliability of automatic finish line detection for tooth preparation in dental computer aided software. J Prosthodont Res. 2023;67:138–43.

Liu L, Zhou R, Yuan S, Sun Z, Lu X, Li J, et al. Simulation training for ceramic crown preparation in the dental setting using a virtual educational system. Eur J Dent Educ. 2020;24:199–206.

Kateeb ET, Kamal MS, Kadamani AM, Abu Hantash RO, Abu Arqoub MM. Utilising an innovative digital software to grade pre-clinical crown preparation exercise. Eur J Dent Educ. 2017;21:220–7.

Feil PH, Gatti JJ. Validation of a motor skills performance theory with applications for dental education. J Dent Educ. 1993;57:628–33.

Hardan L, Bourgi R, Cuevas-Suárez CE, Lukomska-Szymanska M, Monjarás-Ávila AJ, Zarow M, et al. Novel trends in dental color match using different shade selection methods: a systematic review and meta-analysis. Materials (Basel). 2022;15:468.

Kong HJ, Yoo JY, Lee JH, Eom SH, Kim JH. Performance evaluation of deep learning models for the classification and identification of dental implants. J Prosthet Dent. 2023;S0022–3913(23):00467–5.

Almalki YE, Din AI, Ramzan M, Irfan M, Aamir KM, Almalki A, et al. Deep learning models for classification of dental diseases using orthopantomography x-ray opg images. Sensors (Basel). 2022;22:7370.

Yilmaz S, Tasyurek M, Amuk M, Celik M, Canger EM. Developing deep learning methods for classification of teeth in dental panoramic radiography. Oral Surg Oral Med Oral Pathol Oral Radiol. 2023;S2212–4403(23):00116–5.

Tao R, Gavves E and Smeulders AWM. Siamese instance search for tracking. The IEEE Conference on Computer Vision and Pattern Recognition. 2016:1420–9. https://www.computer.org/csdl/proceedings-article/cvpr/2016/8851b420/12OmNscOUb0 . Accessed 15 Feb 2024.

Behpour S, Kitani KM and Ziebart BD. Adversarially optimizing intersection over union for object localization tasks. CoRR. arXiv :1710.07735 2017. https://www.researchgate.net/publication/320582581_Adversarially_Optimizing_Intersection_over_Union_for_Object_Localization_Tasks . Accessed 15 Feb 2024.

Zhao ZQ, Zheng P, Xu ST, Wu X. Object detection with deep learning: a review. IEEE Trans Neural Netw Learn Syst. 2019. https://doi.org/10.48550/arXiv.1807.05511 .

Schepke U, Meijer HJ, Vermeulen KM, Raghoebar GM, Cune MS. Clinical bonding of resin nano ceramic restorations to zirconia abutments: a case series within a randomized clinical trial. Clin Implant Dent Relat Res. 2016;18:984–92.

Rosentritt M, Preis V, Behr M, Krifka S. In-vitro performance of cad/cam crowns with insufficient preparation design. J Mech Behav Biomed Mater. 2019;90:269–74.

Yang Y, Yang Z, Zhou J, Chen L, Tan J. Effect of tooth preparation design on marginal adaptation of composite resin CAD-CAM onlays. J Prosthet Dent. 2020;124:88–93.

Alqutaibi AY, Hamadallah HH, Alturki KN, Aljuhani FM, Aloufi AM, Alghauli MA. Practical applications of robots in prosthodontics for tooth preparation and denture tooth arrangement: a scoping review. J Prosthet Dent. 2024;S0022–3913(24):00120–3.

Alqutaibi AY, Hamadallah HH, Abu Zaid B, Aloufi AM, Tarawah RA. Applications of robots in implant dentistry: a scoping review. J Prosthet Dent. 2023;S0022–3913(23):00770–9.

Download references

Acknowledgements

Not applicable.

This study was supported by Wonkwang University in 2024.

Author information

Authors and affiliations.

Department of Prosthodontics and Wonkwang Dental Research Institute, School of Dentistry, Wonkwang University, Iksan, Republic of Korea

Hyun-Jun Kong

Department of Prosthodontics, School of Dentistry, Wonkwang University, Iksan, Republic of Korea

You can also search for this author in PubMed   Google Scholar

Contributions

HJK conceived the study. All authors developed the review protocol. HJK and YLK did the literature search and assessed the quality of included studies. All authors read, commented critically and approved the final manuscript.

Corresponding author

Correspondence to Hyun-Jun Kong .

Ethics declarations

Ethics approval and consent to participate, consent for publication, competing interests.

The authors declare no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ .

Reprints and permissions

About this article

Cite this article.

Kong, HJ., Kim, YL. Application of artificial intelligence in dental crown prosthesis: a scoping review. BMC Oral Health 24 , 937 (2024). https://doi.org/10.1186/s12903-024-04657-0

Download citation

Received : 29 May 2024

Accepted : 23 July 2024

Published : 13 August 2024

DOI : https://doi.org/10.1186/s12903-024-04657-0

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Dental crown
• Dental prosthesis
• Deep learning

BMC Oral Health

ISSN: 1472-6831

• Submission enquiries: [email protected]
• General enquiries: [email protected]

MIT Technology Review

• Newsletters

Here’s how people are actually using AI

Something peculiar and slightly unexpected has happened: people have started forming relationships with AI systems.

• Melissa Heikkilä archive page

This story is from The Algorithm, our weekly newsletter on AI. To get it in your inbox first, sign up here .

When the generative AI boom started with ChatGPT in late 2022, we were sold a vision of superintelligent AI tools that know everything, can replace the boring bits of work, and supercharge productivity and economic gains. 

Two years on, most of those productivity gains haven’t materialized. And we’ve seen something peculiar and slightly unexpected happen: People have started forming relationships with AI systems. We talk to them, say please and thank you, and have started to invite AIs into our lives as friends, lovers, mentors, therapists, and teachers. 

We’re seeing a giant, real-world experiment unfold, and it’s still uncertain what impact these AI companions will have either on us individually or on society as a whole, argue Robert Mahari, a joint JD-PhD candidate at the MIT Media Lab and Harvard Law School, and Pat Pataranutaporn, a researcher at the MIT Media Lab. They say we need to prepare for “addictive intelligence”, or AI companions that have dark patterns built into them to get us hooked. You can read their piece here . They look at how smart regulation can help us prevent some of the risks associated with AI chatbots that get deep inside our heads. 

The idea that we’ll form bonds with AI companions is no longer just hypothetical. Chatbots with even more emotive voices, such as OpenAI’s GPT-4o , are likely to reel us in even deeper. During safety testing , OpenAI observed that users would use language that indicated they had formed connections with AI models, such as “This is our last day together.” The company itself admits that emotional reliance is one risk that might be heightened by its new voice-enabled chatbot. 

There’s already evidence that we’re connecting on a deeper level with AI even when it’s just confined to text exchanges. Mahari was part of a group of researchers that analyzed a million ChatGPT interaction logs and found that the second most popular use of AI was sexual role-playing. Aside from that, the overwhelmingly most popular use case for the chatbot was creative composition. People also liked to use it for brainstorming and planning, asking for explanations and general information about stuff.  

These sorts of creative and fun tasks are excellent ways to use AI chatbots. AI language models work by predicting the next likely word in a sentence. They are confident liars and often present falsehoods as facts, make stuff up, or hallucinate. This matters less when making stuff up is kind of the entire point. In June, my colleague Rhiannon Williams wrote about how comedians found AI language models to be useful for generating a first “vomit draft” of their material; they then add their own human ingenuity to make it funny.

But these use cases aren’t necessarily productive in the financial sense. I’m pretty sure smutbots weren’t what investors had in mind when they poured billions of dollars into AI companies, and, combined with the fact we still don't have a killer app for AI,it's no wonder that Wall Street is feeling a lot less bullish about it recently.

The use cases that would be “productive,” and have thus been the most hyped, have seen less success in AI adoption. Hallucination starts to become a problem in some of these use cases, such as code generation, news and online searches , where it matters a lot to get things right. Some of the most embarrassing failures of chatbots have happened when people have started trusting AI chatbots too much, or considered them sources of factual information. Earlier this year, for example, Google’s AI overview feature, which summarizes online search results, suggested that people eat rocks and add glue on pizza. 

And that’s the problem with AI hype. It sets our expectations way too high, and leaves us disappointed and disillusioned when the quite literally incredible promises don’t happen. It also tricks us into thinking AI is a technology that is even mature enough to bring about instant changes. In reality, it might be years until we see its true benefit.

Now read the rest of The Algorithm

Deeper learning, ai “godfather” yoshua bengio has joined a uk project to prevent ai catastrophes.

Yoshua Bengio, a Turing Award winner who is considered one of the godfathers of modern AI, is throwing his weight behind a project funded by the UK government to embed safety mechanisms into AI systems. The project, called Safeguarded AI, aims to build an AI system that can check whether other AI systems deployed in critical areas are safe. Bengio is joining the program as scientific director and will provide critical input and advice. 

What are they trying to do: Safeguarded AI’s goal is to build AI systems that can offer quantitative guarantees, such as risk scores, about their effect on the real world. The project aims to build AI safety mechanisms by combining scientific world models, which are essentially simulations of the world, with mathematical proofs. These proofs would include explanations of the AI’s work, and humans would be tasked with verifying whether the AI model’s safety checks are correct. Read more from me here .

Bits and Bytes

Google deepmind trained a robot to beat humans at table tennis.

Researchers managed to get a robot  wielding a 3D-printed paddle to win 13 of 29 games against human opponents of varying abilities in full games of competitive table tennis. The research represents a small step toward creating robots that can perform useful tasks skillfully and safely in real environments like homes and warehouses, which is a long-standing goal of the robotics community. ( MIT Technology Review )

Are we in an AI bubble? Here’s why it’s complex.

There’s been a lot of debate recently, and even some alarm, about whether AI is ever going to live up to its potential, especially thanks to tech stocks’ recent nosedive. This nuanced piece explains why although the sector faces significant challenges, it’s far too soon to write off AI’s transformative potential. ( Platformer ) 

How Microsoft spread its bets beyond OpenAI

Microsoft and OpenAI have one of the most successful partnerships in AI. But following OpenAI’s boardroom drama last year, the tech giant and its CEO, Satya Nadella, have been working on a strategy that will make Microsoft more independent of Sam Altman’s startup. Microsoft has diversified its investments and partnerships in generative AI, built its own smaller, cheaper models, and hired aggressively to develop its consumer AI efforts. ( Financial Times ) 

Humane’s daily returns are outpacing sales

Artificial intelligence

How generative AI could reinvent what it means to play

AI-powered NPCs that don’t need a script could make games—and other worlds—deeply immersive.

• Niall Firth archive page

It was able to draw on vast amounts of data to refine its playing style and adjust its tactics as matches progressed.

• Rhiannon Williams archive page

We need to prepare for ‘addictive intelligence’

The allure of AI companions is hard to resist. Here’s how innovation in regulation can help protect people.

• Robert Mahari archive page
• Pat Pataranutaporn archive page

Synthesia’s hyperrealistic deepfakes will soon have full bodies

With bodies that move and hands that wave, deepfakes just got a whole lot more realistic.

Stay connected

Get the latest updates from mit technology review.

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at [email protected] with a list of newsletters you’d like to receive.

What are the risks from Artificial Intelligence?

A comprehensive living database of over 700 ai risks categorized by their cause and risk domain., what is the ai risk repository.

The AI Risk Repository has three parts:

• The AI Risk Database captures 700+ risks extracted from 43 existing frameworks, with quotes and page numbers.
• The Causal Taxonomy of AI Risks classifies how, when, and why these risks occur.
• The Domain Taxonomy of AI Risks classifies these risks into seven domains (e.g., “Misinformation”) and 23 subdomains (e.g., “False or misleading information”).

How can I use the Repository?

The AI Risk Repository provides:

• An accessible overview of the AI risk landscape.
• A regularly updated source of information about new risks and research.
• A common frame of reference for researchers, developers, businesses, evaluators, auditors, policymakers, and regulators.
• A resource to help develop research, curricula, audits, and policy.
• An easy way to find relevant risks and research.

AI Risk Database

The AI Risk Database links each risk to the source information (paper title, authors), supporting evidence (quotes, page numbers), and to our Causal and Domain Taxonomies. You can copy it on Google Sheets , or OneDrive . Watch our explainer video below.

Search below if you want to explore the risks extracted into our database. This search looks for exact text matches in one field: "Description". It returns information for four fields: "QuickRef", "Risk category", "Risk subcategory", and "Description". For example, try searching for "privacy" to see all risk descriptions which mention this term.

Causal Taxonomy of AI Risks

The Causal Taxonomy of AI risks classifies how, when, and why an AI risk occurs. You can explore the taxonomy (to three levels of depth) in the interactive figure below. Read our preprint for more detail.

Search below if you want to explore how we group risks by cause in our database. This search looks for exact text matches in three fields: "Entity", "Intention" and "Timing". It returns information for seven fields: "QuickRef", "Risk category", "Risk subcategory", "Description", "Entity", "Intent", and "Timing". For instance, try searching for "Pre-deployment" to see all risks of this category.

Domain Taxonomy of AI Risks

The Domain Taxonomy of AI Risks classifies risks from AI into seven domains and 23 subdomains. You can explore the taxonomy (to four levels of depth) in the interactive figure below. Read our preprint for more detail.

Search below if you want to explore how we group risks by domain. This search looks for exact text matches in two fields: "Domain" and "Subdomain". It returns information for six fields: "QuickRef", "Risk category", "Risk subcategory", "Description", "Domain" and "Subdomain". For instance, try searching for "Misinformation" to see all risks categorized in this domain.

How to use the AI Risk Repository

• Our Database is free to copy and use.
• The Causal and Domain Taxonomies can be used separately to filter this database to identify specific risks, for instance, risks occurring pre-deployment or post-deployment or related to Misinformation .
• The Causal and Domain Taxonomies can be used together to understand how each causal factor (i.e., entity , intention and timing ) relate to each risk domain. For example, to identify the intentional and unintentional variations of Discrimination & toxicity .
• ‍ Offer feedback or suggest missing resources risks here , or email pslat[at]mit.edu.

We provide examples of use cases for some key audiences below.

Frequently Asked Questions

Acknowledgments

Feedback and useful input: Anka Reuel, Michael Aird, Greg Sadler, Matthjis Maas, Shahar Avin, Taniel Yusef, Elizabeth Cooper, Dane Sherburn, Noemi Dreksler, Uma Kalkar, CSER, GovAI, Nathan Sherburn, Andrew Lucas, Jacinto Estima, Kevin Klyman, Bernd W. Wirtz, Andrew Critch, Lambert Hogenhout, Zhexin Zhang, Ian Eisenberg, Stuart Russel, and Samuel Salzer .

How to engage

Read our preprint, and copy and use our database, follow mit futuretech on social media for updates.

• SUGGESTED TOPICS
• The Magazine
• Newsletters
• Managing Yourself
• Managing Teams
• Work-life Balance
• The Big Idea
• Data & Visuals
• Reading Lists
• Case Selections
• HBR Learning
• Topic Feeds
• Account Settings
• Email Preferences

Embracing Gen AI at Work

• H. James Wilson
• Paul R. Daugherty

The skills you need to succeed in the era of large language models

Today artificial intelligence can be harnessed by nearly anyone, using commands in everyday language instead of code. Soon it will transform more than 40% of all work activity, according to the authors’ research. In this new era of collaboration between humans and machines, the ability to leverage AI effectively will be critical to your professional success.

This article describes the three kinds of “fusion skills” you need to get the best results from gen AI. Intelligent interrogation involves instructing large language models to perform in ways that generate better outcomes—by, say, breaking processes down into steps or visualizing multiple potential paths to a solution. Judgment integration is about incorporating expert and ethical human discernment to make AI’s output more trustworthy, reliable, and accurate. It entails augmenting a model’s training sources with authoritative knowledge bases when necessary, keeping biases out of prompts, ensuring the privacy of any data used by the models, and scrutinizing suspect output. With reciprocal apprenticing, you tailor gen AI to your company’s specific business context by including rich organizational data and know-how into the commands you give it. As you become better at doing that, you yourself learn how to train the AI to tackle more-sophisticated challenges.

The AI revolution is already here. Learning these three skills will prepare you to thrive in it.

Generative artificial intelligence is expected to radically transform all kinds of jobs over the next few years. No longer the exclusive purview of technologists, AI can now be put to work by nearly anyone, using commands in everyday language instead of code. According to our research, most business functions and more than 40% of all U.S. work activity can be augmented, automated, or reinvented with gen AI. The changes are expected to have the largest impact on the legal, banking, insurance, and capital-market sectors—followed by retail, travel, health, and energy.

• H. James Wilson is the global managing director of technology research and thought leadership at Accenture Research. He is the coauthor, with Paul R. Daugherty, of Human + Machine: Reimagining Work in the Age of AI, New and Expanded Edition (HBR Press, 2024). hjameswilson
• Paul R. Daugherty is Accenture’s chief technology and innovation officer. He is the coauthor, with H. James Wilson, of Human + Machine: Reimagining Work in the Age of AI, New and Expanded Edition (HBR Press, 2024). pauldaugh

Partner Center