Thinking on the Use of Artificial Intelligence in Drug Discovery

InfoMetricsFiguresRef.SI Journal of Medicinal ChemistryASAPArticle This publication is free to access through this site. Learn More CiteCitationCitation and abstractCitation and referencesMore citation options ShareShare onFacebookX (Twitter)WeChatLinkedInRedditEmailJump toExpandCollapse EditorialFebruary 24, 2025Thinking on the Use of Artificial Intelligence in Drug DiscoveryClick to copy article linkArticle link copied!Yuxi WangYuxi WangDepartment of Respiratory and Critical Care Medicine, Targeted Tracer Research and Development Laboratory, Frontiers Science Center for Disease-Related Molecular Network, West China Hospital of Sichuan University, Chengdu 610041, ChinaFrontiers Medical Center, Tianfu Jincheng Laboratory, Chengdu 610212, Sichuan, ChinaMore by Yuxi Wanghttps://orcid.org/0000-0002-4056-5829Zelin HuZelin HuDepartment of Respiratory and Critical Care Medicine, Targeted Tracer Research and Development Laboratory, Frontiers Science Center for Disease-Related Molecular Network, West China Hospital of Sichuan University, Chengdu 610041, ChinaFrontiers Medical Center, Tianfu Jincheng Laboratory, Chengdu 610212, Sichuan, ChinaMore by Zelin HuJunbiao ChangJunbiao ChangCollege of Chemistry, Pingyuan Laboratory, Zhengzhou University, Zhengzhou 450001, Henan, ChinaMore by Junbiao ChangBin Yu*Bin YuTianjian Laboratory of Advanced Biomedical Sciences, Institute of Advanced Biomedical Sciences, Zhengzhou University, Zhengzhou 450000, ChinaCollege of Chemistry, Pingyuan Laboratory, Zhengzhou University, Zhengzhou 450001, Henan, China*Bin Yu, Email: [email protected]More by Bin Yuhttps://orcid.org/0000-0002-7207-643XOpen PDFJournal of Medicinal ChemistryCite this: J. Med. Chem. 2025, XXXX, XXX, XXX-XXXClick to copy citationCitation copied!https://pubs.acs.org/doi/10.1021/acs.jmedchem.5c00373https://doi.org/10.1021/acs.jmedchem.5c00373Published February 24, 2025 Publication History Received 6 February 2025Published online 24 February 2025editorialPublished 2025 by American Chemical Society. This publication is available under these Terms of Use. Request reuse permissionsThis publication is licensed for personal use by The American Chemical Society. ACS PublicationsPublished 2025 by American Chemical SocietyAdvances in machine learning algorithms and big data processing capabilities have propelled artificial intelligence (AI) to the forefront, with its applications in drug discovery rapidly increasing. A report from McKinsey & Company indicates that the use of AI can help pharmaceutical companies save 30 to 50% in drug development costs and increase the pipeline speed by over 20%. (1) AI technology is becoming increasingly common from drug discovery to clinical trials. First, AI can quickly screen potential active compounds using machine learning and deep learning algorithms, guiding hit-to-lead optimization to improve therapeutic efficacy and reduce toxicity, significantly enhancing the efficiency of drug discovery. Second, AI-driven predictive models can enhance the success rate of experiments, reduce research and development (R&D) costs, and even change a drug's behavior in the body, including its pharmacokinetics (PK) and pharmacodynamics (PD), to optimize dosage and dosing regimens. By integrating expertise from various disciplines, it enables the derivation of novel conclusions, producing forward-looking results in traditionally empirical areas. Furthermore, AI has driven the development of personalized medicine, bringing significant transformations and opportunities for the future of healthcare.Drug discovery and development is the most advanced area of AI, with numerous breakthroughs already achieved (Figure 1A). For instance, Alex Zhavoronkov's team developed a deep generative model, generative tensorial reinforcement learning (GENTRL) for de novo small molecule design. (2) The model combines reinforcement learning, variational inference, and tensor decompositions for rapid generation and optimization of potentially pharmacologically active compounds. GENTRL uses three layers of self-organizing mappings (SOMs) to generate innovative DDR1 inhibitors, and these small molecules then undergo multiple screenings and optimizations, resulting in six candidate compounds in just 21 days, with four showing potent biological activity. Even more surprisingly, ISM001-055AI, a first-in-class small molecule inhibitor developed by generative AI technology, shows positive results in a phase IIa clinical trial for treating idiopathic pulmonary fibrosis (IPF). (3) In September 2024, Insilico Medicine announced the results of a Phase IIa clinical trial of ISM001-055. The trial, which enrolled 71 IPF patients across 21 clinical research centers in China, demonstrated that ISM001-055 had a favorable safety profile at all dose levels and exhibited a dose-dependent trend of efficacy in forced vital capacity (FVC), an important measure of lung function in IPF patients. The improvement in lung function in just 12 weeks brings great hope to IPF patients. The success of ISM001-055 showcases the huge potential of AI in drug discovery and development, paving the way for larger-scale trials in the future. Fang's team recently developed a robotic system for high-throughput chemical synthesis, online characterization, and large-scale photocatalytic reaction condition screening (10,000 reaction conditions per day). The system is based on liquid-core waveguide, microfluidic liquid-handling, and artificial intelligence techniques to significantly reduce the time and effort required for complex organic synthesis. (4) They have also developed an AI-assisted method for predicting absorbance, analyzing factors that influence convective and molecular diffusion effects. When a candidate compound is selected at the chemistry, manufacturing, and control (CMC) stage, crystal structure prediction (CSP) and structure-based stability evaluation can be used with advanced experimental screening techniques to minimize the risk of polymorphic problems and accelerate decision-making for formulation studies. (5,6) During the development of Paxlovid, Pfizer's oral Covid-19 treatment, Pfizer and XtalPi elucidated the subtle differences in structure, dynamics, and stability of two enantiotropically related anhydrous polymorphs of the novel antiviral medicine nirmatrelvir by employing a series of highly orthogonal experimental and computational approaches. (7) This work ultimately determined Form 1 for formulation development as it is the stable form above the transition temperature of 17 °C, demonstrating the unprecedented speed in bringing Paxlovid to patients in record time amidst the pandemic.Figure 1Figure 1. (A) AI-driven drug discovery process, including de novo design, compound screening, interaction mulations, organic synthesis, and optimization of pharmacokinetic properties. (B) The AlphaFold model architecture predicts the initial protein structure, demonstrating near-experimental accuracy in protein structure prediction through recycling (PDB code: 6Y4F). Some elements were created with BioRender (https://biorender.com).High Resolution ImageDownload MS PowerPoint SlideAntibody drugs, along with small molecule drugs, are widely used in treating cancer, autoimmune diseases, and infectious diseases because of their high specificity for targets. (8) AI can be used for structure prediction, design optimization, humanization of antibody drugs, and predicting the optimal drug-to-antibody ratio (DAR) for antibody-drug conjugates (ADCs). This helps shorten development timelines and increase success rates. AI's superior generalization ability and high degree of automation make it highly effective in vaccine development, which usually takes over 10 years. For example, AI has played a crucial role in optimizing mRNA sequences. (9,10) The LinearDesign deep learning algorithm has greatly enhanced the stability, protein expression, and immunogenicity of mRNA vaccines by optimizing mRNA sequence structural stability and codon usage. The algorithm successfully solved the large-scale search problem in traditional mRNA design, finding the optimal mRNA sequence in less than 11 min. It showcases the potential of AI and computational linguistics technologies in mRNA design and was licensed to Moderna, a leading global RNA technology company in 2022 to enhance the development of mRNA vaccines and therapeutics. (10)AlphaFold is a recent breakthrough in protein structure prediction that uses a neural network-based machine learning method. It combines evolutionary data and geometric properties to accurately predict three-dimensional structures (Figure 1B). (9) After entering the amino acid sequence of a protein, Evoformer generates a multiple sequence alignment (MSA) and templates from relevant evolutionary databases to capture the relationships and covariation between sequences. The predicted structure is then fed back into Evoformer and the structure module multiple times for iterative optimization. Each recursive step, known as "Recycling," refines the structure further to enhance accuracy. Significant progress has been made in predicting protein structures, but this is only the beginning of exploration in life sciences. AI is currently mainly used at the molecular level, but future research will integrate diverse high-quality data and advance to higher levels to discover new patterns and principles in biological systems. For example, the Peking University International Cancer Institute has partnered with XtalPi to investigate disease mechanisms and drug actions through multimodal data integration. (11) The Shu group at Peking University utilized high-throughput CRISPR technology for large-scale gene editing in cells, whereas XtalPi's independently developed cell research platform, X-Map, which allows for collection of large-scale, high-content imaging and transcriptome data following cell perturbation. The X-Profiler algorithm developed by XtalPi is proficient at extracting relevant information for specific downstream tasks. It effectively addresses issues like out-of-focus blurring in high-content imaging due to variations in well plate edge height, reduces data noise, and enhances the signal-to-noise ratio (SNR). Furthermore, the X-Profiler can adaptively adjust data quality control strategies based on task requirements, significantly boosting model performance. By integrating experimental methods with AI algorithms, this approach enables precise observations at the cellular level using real-world multidimensional data, establishing correlations between physiological changes and gene or drug regulation. This method offers higher throughput and lower costs compared to animal models, allowing for quicker generation of high-quality data for specific research systems and improving the efficiency and success rate of drug discovery and development.Although AI is well established in many aspects of drug discovery and development, the outdated traditional education system is no longer suitable for the rapid development of AI and urgently needs reform and optimization. To address the educational challenges posed by the rapid development of AI, it is essential to have a clear understanding of its specific applications. This will help us pinpoint which areas of teaching content require additional support. The traditional teaching methods have issues like outdated curricula, lack of practical experience, and limited interdisciplinary integration, focusing too much on foundational subjects and neglecting crucial skills in data science and AI. The lack of interdisciplinary integration and practical experience makes it difficult to meet the demands of the AI era. Given AI's increasing importance, teaching approaches should be updated. First and foremost, teachers should be trained with AI technologies to ensure they are prepared to teach AI courses. Second, interdisciplinary integration in education should be promoted to inspire insights from different disciplines. Computer science and data science should be integrated into the pharmaceutical curriculum to break down disciplinary silos. For example, the basic theory component of the drug molecular design course should include fundamental concepts of AI, machine learning, and deep learning, along with their applications in the pharmaceutical industry. This will cover skills such as data preprocessing, model training and evaluation, and feature engineering. The core course section should explore AI applications in drug target screening, molecular docking, virtual screening, molecular structure analysis, and AI modeling comprehensively. Cutting-edge AI technologies in drug discovery, such as reinforcement learning and Generative Adversarial Networks (GANs), should be introduced. GANs are a specialized type of deep learning with two neural networks─one for image generation and one for discrimination. They can efficiently create diverse, high-quality molecules, showing unique advantages and potential by learning existing data distribution to explore new chemical spaces. Third, new courses should be added and professors from various fields should collaborate to edit textbooks that combine pharmaceutical science with AI for the updated curriculum. International academic conferences and online courses are also valuable educational resources for students to learn AI. It is important to expand AI-related courses, including big data analysis, natural language processing, bioinformatics, etc., to enhance relevant skills. However, increasing reliance on AI learning unilaterally will only add more pressure on students, thus learning new courses should be personalized to avoid overwhelming students with the broad scope of AI content. Reform should focus on innovation, professionalism, and personalization to better address individual learning needs. Students should choose courses based on their professional needs, focusing on relevant skills after gaining foundational knowledge in the early stages of education.Use AI technology in real-world drug development scenarios to teach students how to solve practical problems. AI-driven practical training and hands-on experience must be included. For example, creating a virtual lab platform that integrates AI models and data analysis tools for simulating real-world scenarios like drug screening, molecular docking, and drug safety prediction. This platform should also have online collaboration features to allow students to work together in real time, mimicking a professional work environment that promotes interdisciplinary teamwork. Additionally, enhancing partnerships with pharmaceutical and AI technology companies to develop a platform that connects industry and academia is crucial. Project content should be tailored to meet the actual needs of industry partners to offer students practical solutions relevant to the field. By gaining a strong grasp of AI concepts, students will be better prepared for future roles in drug development.To be expected, AI technology has and will profoundly revolutionize the field of drug discovery. Educational policymakers and academic institutions globally should update their training programs to incorporate advanced AI technologies. Students need to learn latest AI techniques like deep learning and natural language processing to be better prepared for future challenges in drug development. Students without AI skills can be readily replaced by those who embrace AI advancement. However, we have to keep in mind that AI may not be a panacea due to its limitation in the quality of the data. High-quality, comprehensive experimental data sets are essential for generating reliable results, yet not all entities with an AI platform possess such data. The impact of AI and machine learning in drug discovery is directly dependent on the quality and depth of the training data used to develop the AI models. Regardless of how advanced the algorithms are, if the training data is biased or of poor quality, the AI predictions will be inaccurate, thereby limiting their potential to drive new drug discovery. And algorithmic biases can lead to inaccurate therapeutic effects of the model on certain populations, which is especially critical in the safety of patients' medication. Additionally, it is not only important to pay attention to the quality of the data, but also to ensure the protection of sensitive patient data. We need to establish strict data protection measures to ensure the security of the data during acquisition, storage, and processing, thereby safeguarding patient data privacy.Author InformationClick to copy section linkSection link copied!Corresponding AuthorBin Yu, Tianjian Laboratory of Advanced Biomedical Sciences, Institute of Advanced Biomedical Sciences, Zhengzhou University, Zhengzhou 450000, China; College of Chemistry, Pingyuan Laboratory, Zhengzhou University, Zhengzhou 450001, Henan, China, https://orcid.org/0000-0002-7207-643X, Email: [email protected]AuthorsYuxi Wang, Department of Respiratory and Critical Care Medicine, Targeted Tracer Research and Development Laboratory, Frontiers Science Center for Disease-Related Molecular Network, West China Hospital of Sichuan University, Chengdu 610041, China; Frontiers Medical Center, Tianfu Jincheng Laboratory, Chengdu 610212, Sichuan, China, https://orcid.org/0000-0002-4056-5829Zelin Hu, Department of Respiratory and Critical Care Medicine, Targeted Tracer Research and Development Laboratory, Frontiers Science Center for Disease-Related Molecular Network, West China Hospital of Sichuan University, Chengdu 610041, China; Frontiers Medical Center, Tianfu Jincheng Laboratory, Chengdu 610212, Sichuan, ChinaJunbiao Chang, College of Chemistry, Pingyuan Laboratory, Zhengzhou University, Zhengzhou 450001, Henan, ChinaAuthor ContributionsY.W. and Z.H. contributed equally to this work.NotesViews expressed in this editorial are those of the authors and not necessarily the views of the ACS.AcknowledgmentsClick to copy section linkSection link copied!This work was provided by the National Natural Science Foundation of China (Nos. 82473782, 22277110, and 82473761), Sichuan Science and Technology Program (No. 2024ZYD0130), Natural Science Foundation of Henan Province (No. 242301420005), Key Research Project for Basic Research in Henan Province Universities (No. 25ZX001). The authors acknowledge Professor Craig W. Lindsley (Editor-in-Chief of Journal of Medicinal Chemistry; Executive Director, Warren Center for Neuroscience Drug Discovery, Vanderbilt University) and Dr. Shuhao Wen (Chairman of XtalPi) for their insightful discussions.ReferencesClick to copy section linkSection link copied! This article references 11 other publications. 1Viswa, C. A.; Bleys, J.; Leydon, E.; Shah, B.; Zurkiya, D. Generative AI in the pharmaceutical industry: Moving from hype to reality; McKinsey & Company, 2024.Google ScholarThere is no corresponding record for this reference.2Zhavoronkov, A. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat. Biotechnol. 2019, 37, 1038– 1040, DOI: 10.1038/s41587-019-0224-x Google Scholar2Deep learning enables rapid identification of potent DDR1 kinase inhibitorsZhavoronkov, Alex; Ivanenkov, Yan A.; Aliper, Alex; Veselov, Mark S.; Aladinskiy, Vladimir A.; Aladinskaya, Anastasiya V.; Terentiev, Victor A.; Polykovskiy, Daniil A.; Kuznetsov, Maksim D.; Asadulaev, Arip; Volkov, Yury; Zholus, Artem; Shayakhmetov, Rim R.; Zhebrak, Alexander; Minaeva, Lidiya I.; Zagribelnyy, Bogdan A.; Lee, Lennart H.; Soll, Richard; Madge, David; Xing, Li; Guo, Tao; Aspuru-Guzik, AlanNature Biotechnology (2019), 37 (9), 1038-1040CODEN: NABIF9; ISSN:1087-0156. (Nature Research) We have developed a deep generative model, generative tensorial reinforcement learning (GENTRL), for de novo small-mol. design. GENTRL optimizes synthetic feasibility, novelty, and biol. activity. We used GENTRL to discover potent inhibitors of discoidin domain receptor 1 (DDR1), a kinase target implicated in fibrosis and other diseases, in 21 days. Four compds. were active in biochem. assays, and two were validated in cell-based assays. One lead candidate was tested and demonstrated favorable pharmacokinetics in mice. >> More from SciFinder ®https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhs12gurnM&md5=b15262b61b9172ab2bc37e534a70f0103Ren, F. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nat. Biotechnol. 2025, 43, 63, DOI: 10.1038/s41587-024-02143-0 Google ScholarThere is no corresponding record for this reference.4Lu, J. M. Roboticized AI-assisted microfluidic photocatalytic synthesis and screening up to 10,000 reactions per day. Nat. Commun. 2024, 15, 8826, DOI: 10.1038/s41467-024-53204-6 Google ScholarThere is no corresponding record for this reference.5Zhang, P. Harnessing Cloud Architecture for Crystal Structure Prediction Calculations. Cryst. Growth Des. 2018, 18, 6891– 6900, DOI: 10.1021/acs.cgd.8b01098 Google Scholar5Harnessing Cloud Architecture for Crystal Structure Prediction CalculationsZhang, Peiyu; Wood, Geoffrey P. F.; Ma, Jian; Yang, Mingjun; Liu, Yang; Sun, Guangxu; Jiang, Yide A.; Hancock, Bruno C.; Wen, ShuhaoCrystal Growth & Design (2018), 18 (11), 6891-6900CODEN: CGDEFU; ISSN:1528-7483. (American Chemical Society) Accurate and rapid crystal structure predictions have the potential to transform the development of new materials, particularly in fields with highly complex mol. structures (such as in drug development). A novel cloud-computing crystal structure prediction (CSP) platform with the capability of scheduling hundreds of thousands CPU cores and integrating cutting-edge computational chem. algorithms is presented. This new cloud-computing based CSP platform was applied to 3 cryst. drug substances of increasing complexity. The lattice energies of the exptl. crystal structures are all within 4.0 kJ/mol of the lowest energy predicted structures. From the results of this work, the algorithm improvement and the mass computational power of cloud computing can reduce the whole CSP process to just 1-3 wk for Z' = 1 systems and <5 wk for significantly more complex systems. Also, it is possible to simultaneously perform calcns. for multiple mols. if desired. Of these improvements, CSP calcns. can potentially be applied in conjunction with state-of-the-art exptl. screening techniques to reduce the risk of finding new solid forms after product launch provided that a sufficient no. of stoichiometries and space groups are explored. >> More from SciFinder ®https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhvVCru7bK&md5=c0cf7c65fbb34d2b2c9021986fd7161b6Yang, M. Prediction of the Relative Free Energies of Drug Polymorphs above Zero Kelvin. Cryst. Growth Des. 2020, 20, 5211– 5224, DOI: 10.1021/acs.cgd.0c00422 Google Scholar6Prediction of the Relative Free Energies of Drug Polymorphs above Zero KelvinYang, Mingjun; Dybeck, Eric; Sun, Guangxu; Peng, Chunwang; Samas, Brian; Burger, Virginia M.; Zeng, Qun; Jin, Yingdi; Bellucci, Michael A.; Liu, Yang; Zhang, Peiyu; Ma, Jian; Jiang, Yide Alan; Hancock, Bruno C.; Wen, Shuhao; Wood, Geoffrey P. F.Crystal Growth & Design (2020), 20 (8), 5211-5224CODEN: CGDEFU; ISSN:1528-7483. (American Chemical Society) Crystal structure prediction (CSP) calcns. can reduce risk and improve efficiency during drug development. Traditionally, CSP calcns. use lattice energies computed through d. functional theory. While this approach is often successful in predicting the low energy structures, it neglects the crucial role of thermal effects on polymorph stabilities. In the present study, we develop a robust and efficient protocol for predicting the relative stability of polymorphs at different temps. The protocol is executed on a highly parallel cloud computing infrastructure to produce results at time scales useful for drug development timelines. We demonstrate this protocol on mol. XXIII from the sixth crystal structure prediction blind test. Our results predict that Form D is the most stable exptl. obsd. polymorph at ambient temp. and Form C is the most stable at low temp. consistent with expts. also conducted in the present study. We have implemented a modified version of the Einstein Crystal Method to handle polymorph landscapes of a complex org. mols. The method computes the free energy differences between crystal polymorphs at finite temps. using alchem. transformations and is supplemented with d. functional theory calcns. of lattice energies at Zero Kelvin. Using this method, we are able to predict the known polymorph stability as a function of temp. of this complex mol. The protocol has been fully implemented on cloud services to take advantage of modern compute technol. >> More from SciFinder ®https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhtlSis77F&md5=4e1cc073ff6a93d7973acb164d8467527Sadeghi, M. S. Tale of Two Polymorphs: Investigating the Structural Differences and Dynamic Relationship between Nirmatrelvir Solid Forms (Paxlovid). Mol. Pharmaceutics 2024, 21, 3800– 3814, DOI: 10.1021/acs.molpharmaceut.3c01074 Google ScholarThere is no corresponding record for this reference.8Carter, P. J.; Rajpal, A. Designing antibodies as therapeutics. Cell 2022, 185, 2789– 2805, DOI: 10.1016/j.cell.2022.05.029 Google Scholar8Designing antibodies as therapeuticsCarter, Paul J.; Rajpal, ArvindCell (Cambridge, MA, United States) (2022), 185 (15), 2789-2805CODEN: CELLB5; ISSN:0092-8674. (Cell Press) A review. Antibody therapeutics are a large and rapidly expanding drug class providing major health benefits. We provide a snapshot of current antibody therapeutics including their formats, common targets, therapeutic areas, and routes of administration. Our focus is on selected emerging directions in antibody design where progress may provide a broad benefit. These topics include enhancing antibodies for cancer, antibody delivery to organs such as the brain, gastrointestinal tract, and lungs, plus antibody developability challenges including immunogenicity risk assessment and mitigation and s.c. delivery. Machine learning has the potential, albeit as yet largely unrealized, for a transformative future impact on antibody discovery and engineering. >> More from SciFinder ®https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhvV2lurfF&md5=fdf859df5545bcc52110651eae9340989Jumper, J. Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583– 589, DOI: 10.1038/s41586-021-03819-2 Google Scholar9Highly accurate protein structure prediction with AlphaFoldJumper, John; Evans, Richard; Pritzel, Alexander; Green, Tim; Figurnov, Michael; Ronneberger, Olaf; Tunyasuvunakool, Kathryn; Bates, Russ; Zidek, Augustin; Potapenko, Anna; Bridgland, Alex; Meyer, Clemens; Kohl, Simon A. A.; Ballard, Andrew J.; Cowie, Andrew; Romera-Paredes, Bernardino; Nikolov, Stanislav; Jain, Rishub; Adler, Jonas; Back, Trevor; Petersen, Stig; Reiman, David; Clancy, Ellen; Zielinski, Michal; Steinegger, Martin; Pacholska, Michalina; Berghammer, Tamas; Bodenstein, Sebastian; Silver, David; Vinyals, Oriol; Senior, Andrew W.; Kavukcuoglu, Koray; Kohli, Pushmeet; Hassabis, DemisNature (London, United Kingdom) (2021), 596 (7873), 583-589CODEN: NATUAS; ISSN:0028-0836. (Nature Portfolio) Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous exptl. effort, the structures of around 100,000 unique proteins have been detd., but this represents a small fraction of the billions of known protein sequences. Structural coverage is bottlenecked by the months to years of painstaking effort required to det. a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'-has been an important open research problem for more than 50 years. Despite recent progress, existing methods fall far short of at. accuracy, esp. when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with at. accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Crit. Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with exptl. structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates phys. and biol. knowledge about protein structure, leveraging multi-sequence alignments,