A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery

A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery Feixiong Cheng, Hongxin Xiang, Li Zeng, Linlin Hou, Kenli Li, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3773235/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 08 Nov, 2024 Read the published version in Nature Communications → Version 1 posted You are reading this latest preprint version Abstract Precise capture of three-dimensional (3D) dynamic conformation during molecular representation learning is crucial for accurate prediction of drug targets and molecular properties. In this study, we propose a molecular video-based foundation model, named VideoMol, pretrained on 120 million frames of 2 million unlabeled drug-like and bioactive molecules with 3D conformations. VideoMol renders the molecular 3D conformation as a 60-frame dynamic video and designs three self-supervised learning strategies on molecular videos to capture diverse conformational changes. We demonstrate high performance of VideoMol in predicting molecular targets and properties across 44 benchmark drug discovery datasets. VideoMol achieves high accuracy in identifying antiviral molecules against SARS-CoV-2 across 11 high-throughput experimental datasets from the National Center for Advancing Translational Sciences and other diverse disease-specific drug targets. We further present high interpretability of VideoMol through observed key chemical substructures related to dynamic 3D conformational changes compared to traditional state-of-the-art deep learning approaches. In summary, VideoMol offers a powerful tool to expedite drug discovery and development. Biological sciences/Chemical biology/Cheminformatics Health sciences/Medical research/Drug development Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Full Text Additional Declarations There is NO Competing Interest. Supplementary Files SupplementaryTablesS1S18Cheng.zip SupplementaryTables_S1-S18 SupplementaryInformationCheng.pdf SupplemntaryVideoFig.5Cheng.zip SupplementaryVideo for main Figure 5 Cite Share Download PDF Status: Published Journal Publication published 08 Nov, 2024 Read the published version in Nature Communications → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3773235","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":267885452,"identity":"6e6a4195-4629-4a44-b9c7-c22760f78d7e","order_by":0,"name":"Feixiong Cheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAt0lEQVRIiWNgGAWjYDACCRBhY8PAwHyAJC1paQwMbAmkaTlMghb+2c3HHn5JOC/P38bA+LjiFzGW3DmWbiyTcNtwxjEGZsOzfURoMZDIMZOW/HE7wUC+gU2ysYcoLfnfpCUSziUYsDEQrSWHTfJDwgGIloYfRGiRuJFmJs2QkAz0C2OzYWMDEVr4ZyQ/k/yRYAcMMeaDDxv+EKEFBJh5wBRjAwNjG5FaGBE+INaWUTAKRsEoGFEAAKT8MmqgJc+hAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-1736-2847","institution":"Cleveland Clinic","correspondingAuthor":true,"prefix":"","firstName":"Feixiong","middleName":"","lastName":"Cheng","suffix":""},{"id":267885453,"identity":"028540b4-a60a-4f4c-9267-000e1b63ed06","order_by":1,"name":"Hongxin Xiang","email":"","orcid":"https://orcid.org/0000-0001-8345-8735","institution":"Hunan University","correspondingAuthor":false,"prefix":"","firstName":"Hongxin","middleName":"","lastName":"Xiang","suffix":""},{"id":267885454,"identity":"2d2d93e7-2b70-4693-8516-96e43946d660","order_by":2,"name":"Li Zeng","email":"","orcid":"","institution":"Hunan University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Zeng","suffix":""},{"id":267885455,"identity":"b2b30391-506b-4dba-ae92-31f56bf33ead","order_by":3,"name":"Linlin Hou","email":"","orcid":"","institution":"Hunan University","correspondingAuthor":false,"prefix":"","firstName":"Linlin","middleName":"","lastName":"Hou","suffix":""},{"id":267885456,"identity":"4f9d5b77-ebf3-4008-9093-7648ba8e1bff","order_by":4,"name":"Kenli Li","email":"","orcid":"","institution":"Hunan University","correspondingAuthor":false,"prefix":"","firstName":"Kenli","middleName":"","lastName":"Li","suffix":""},{"id":267885457,"identity":"d675617b-ef52-4ddd-9c04-da52b1a04ba5","order_by":5,"name":"Zhimin Fu","email":"","orcid":"","institution":"Northeast Ohio Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhimin","middleName":"","lastName":"Fu","suffix":""},{"id":267885458,"identity":"c40f4638-118d-40f2-81f8-8cc89298233b","order_by":6,"name":"Yunguang Qiu","email":"","orcid":"https://orcid.org/0000-0002-5094-0936","institution":"Cleveland Clinic","correspondingAuthor":false,"prefix":"","firstName":"Yunguang","middleName":"","lastName":"Qiu","suffix":""},{"id":267885459,"identity":"ea5c2b83-4a4f-48d4-abba-a5f140c27324","order_by":7,"name":"Ruth Nussinov","email":"","orcid":"https://orcid.org/0000-0002-8115-6415","institution":"Frederick National Laboratory for Cancer Research (NIH/NCI)","correspondingAuthor":false,"prefix":"","firstName":"Ruth","middleName":"","lastName":"Nussinov","suffix":""},{"id":267885460,"identity":"827097b8-85c7-4f49-bc44-4c80b23670b4","order_by":8,"name":"Jianying Hu","email":"","orcid":"https://orcid.org/0000-0001-7753-886X","institution":"IBM Research","correspondingAuthor":false,"prefix":"","firstName":"Jianying","middleName":"","lastName":"Hu","suffix":""},{"id":267885461,"identity":"75d37f26-415c-45a6-be4b-7686c1fc28b4","order_by":9,"name":"Xiang-Xiang Zeng","email":"","orcid":"","institution":"Hunan University","correspondingAuthor":false,"prefix":"","firstName":"Xiang-Xiang","middleName":"","lastName":"Zeng","suffix":""},{"id":267885462,"identity":"d7c4a79d-6884-48d3-badc-aee90f74df9e","order_by":10,"name":"Michal Rosen-Zvi","email":"","orcid":"https://orcid.org/0000-0001-7616-9724","institution":"IBM Research","correspondingAuthor":false,"prefix":"","firstName":"Michal","middleName":"","lastName":"Rosen-Zvi","suffix":""}],"badges":[],"createdAt":"2023-12-18 18:25:44","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3773235/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3773235/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41467-024-53742-z","type":"published","date":"2024-11-08T05:00:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":49894139,"identity":"5f4aae80-93cc-407f-ac10-273abe269b7b","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":443905,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the VideoMol foundational model. a, Feature extraction of molecular videos. First, we render 2 million molecules with conformers in 3D spatial structure. We then rotate the rendered molecule around the x, y, z axes and generate snapshots for each frame of the molecule video. Finally, we feed the molecular frames into a video encoder to extract latent features. b-d, Three self-supervised tasks for pre-training video encoder. The directionaware pretraining (DAP) task is used to distinguish the relationship between pairs of molecular frames (such as the axis of rotation, the direction of rotation, and the angle of rotation) by using axis classifier (orange), rotation classifier (green) and angle classifier (blue). The video-aware pretraining (VAP) task is used to maximize intra-video similarity and minimize inter-video similarity. The chemical-aware pretraining (CAP) task is used to recognize information related to physicochemical structures in molecular videos by using chemical classifier (pink). e, The finetuning of VideoMol on downstream benchmarks (such as binding activity prediction and molecular property prediction). A multi-layer perceptron (lavender) is added after the pre-trained video encoder for fine-tuning on four types of downstream tasks (20 target prediction, 13 property prediction, 11 SARS-CoV-2 inhibitor prediction and 4 virtual screening and docking). We ensemble the results (logits) of each frame as the prediction result of molecular video (video logit).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/f97dcaf27f3f7ef8199dc0c6.png"},{"id":49894138,"identity":"39f76863-a63f-4b69-b7ec-d087bfb100ea","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":200281,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of the VideoMol framework on various drug discovery tasks. a, The ROC-AUC curves of ImageMol and VideoMol on 10 main types of biochemical kinases with balanced scaffold split. b, The RMSE and MAE performance of ImageMol and VideoMol on 10 GPCR with balanced scaffold split. c, The ROC-AUC curves of ImageMol and VideoMol on 7 molecular property prediction benchmarks with scaffold split. d, The RMSE (FreeSolv, ESOL, Lipo) and MAE (QM7, QM8, QM9) performance of SOTA methods and VideoMol with scaffold split. For ease of presentation, the values of FreesSolv and QM7 are scaled down by a factor of 2 and 100, respectively, and the values of QM8 and QM9 are scaled up by a factor of 50 and 100, respectively. e, The ROC-AUC performance of REDIAL-2020, ImageMol and VideoMol on 11 SARS-CoV-2 datasets with balanced scaffold split.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/4cdb05f1deb970980ee409f3.png"},{"id":49894140,"identity":"6beab371-0323-4034-9117-f780b1a9cc87","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":690800,"visible":true,"origin":"","legend":"\u003cp\u003eThe virtual screening on four common targets (BACE1, COX-1, COX-2 and EP4). a, The ROC-AUC curves of ImageMol and VideoMol on validation set (the first row) and test set (the second row). b, The t-SNE visualization of latent features extracted by VideoMol. c, The drug discovery on BACE1, COX-1, COX-2 and EP4. Green and red points represent inhibitors and non-inhibitors from the ChEMBL dataset, respectively. Green points indicate known inhibitors. The decision boundary is drawn by training an SVM using ChEMBL dataset, where blue to red indicates that the probability of belonging to the inhibitor gradually increases. d, Virtual screening on known inhibitors of BACE1, COX-1, COX2 and EP4. The x-axis and y-axis represent the number of the drug and the predicted probability that the drug is an inhibitor, respectively. The orange and blue backgrounds represent inhibitor area and non-inhibitor area, respectively. The green and blue triangles represent inhibitors predicted by ImageMol and VideoMol respectively. The green and blue circles represent non-inhibitors predicted by ImageMol and VideoMol respectively. The green and blue numbers represent the precision of ImageMol and VideoMol respectively.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/ef125339412cc24043ea1cf3.png"},{"id":49894142,"identity":"9d4e2f6b-3b44-4c32-bda1-e40de0d0a920","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1465478,"visible":true,"origin":"","legend":"\u003cp\u003eThe docking analysis of the 4IVS crystal structure of Homo sapiens beta-secretase 1 (BACE1). a, The 4IVS crystal structure of BACE1. The gray tetragon represents the area of the docking pocket. The grid score is calculated by Dock6.10 (the smaller, the better). b, Top 20 drugs predicted by VideoMol and ImageMol to be active against the BACE1 target. Green and blue represent ImageMol’s drugs and VideoMol’s drugs respectively. Minus/plus signs indicate that the predicted drug is not/is supported by existing experimental data from publish literatures. c, The docking results of the Top 20 drugs predicted by VideoMol. The x-axis and y-axis represent the index and grid scores of the drug respectively. Light blue and light orange areas indicate worse and better grid scores than the 4IVS (grid score=-52.47), respectively. d, Docking examples of 6 drugs (Gonadorelin, Angiotensin II, Edotreotide gallium Ga-68, Cyanocobalamin, Isavuconazonium and Elbasvir) with the best grid scores. The numerical value in the bracket represents the grid score.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/61f1f61841cc55d04f672dc9.png"},{"id":49894801,"identity":"e5f9d74f-48d9-4ea2-ab9d-b53b0ab06980","added_by":"auto","created_at":"2024-01-19 21:34:11","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1090584,"visible":true,"origin":"","legend":"\u003cp\u003eFeature distribution and biological interpretation of VideoMol. a, Visualization of each frame in 100 molecular videos (60 frames for each video). Representations are extracted by VideoMol and dimensionally reduced by t-SNE. Different colors represent frames in different cluster videos. DB index is anmetric to evaluate the clustering quality, and the larger the value, the better the clustering performance. b, Similarity distribution of intra-video and inter-video. Similarity is computed using a pair of frames from intra-video or inter-video. The content in brackets indicates the average similarity of the distribution. c, t-SNE visualization of VideoMol fingerprints. Different colors represent different cluster labels (this cluster label is obtained in the chemical-aware pretrainin task). d-f, Grad-CAM visualization of VideoMol on molecular frames. Warmer/cooler colors indicate that the region is more/less important for the model’s inferences. The light blue indicates the lowest importance (a value of 0). We use 0.6 as the threshold for visualization, that is, set the importance lower than 0.6 to 0. In d, each row represents a molecular video. In e, Pairs of molecular frames represent frames where structure is missing and frames where structure appears, respectively. In f, Each panel represents examples of key structures related to BACE-1 inhibitory activities from frames of different molecules.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/eefe87c090da8b66770cf742.png"},{"id":68606124,"identity":"348f5207-f1a5-4cfd-b208-2b6fbe7eb9bc","added_by":"auto","created_at":"2024-11-09 08:05:58","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1634813,"visible":true,"origin":"","legend":"","description":"","filename":"MainTextCheng.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1_covered_0f4237a4-8714-4a9a-9988-6c22ab3779bd.pdf"},{"id":49894141,"identity":"b0a11d54-3741-4325-b42c-e18f199d69dc","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":895634,"visible":true,"origin":"","legend":"SupplementaryTables_S1-S18","description":"","filename":"SupplementaryTablesS1S18Cheng.zip","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/fecd8310caad89db2707eedb.zip"},{"id":49894144,"identity":"eeefa748-5f9c-4523-9be6-abec24a2f5b3","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1148125,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cbr\u003e\u003c/p\u003e","description":"","filename":"SupplementaryInformationCheng.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/405afef7e699f4039dd093fc.pdf"},{"id":49894145,"identity":"f744bd88-e370-4d84-b867-eb9d7821f532","added_by":"auto","created_at":"2024-01-19 21:26:11","extension":"zip","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":4343520,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementaryVideo for main Figure 5\u003c/p\u003e","description":"","filename":"SupplemntaryVideoFig.5Cheng.zip","url":"https://assets-eu.researchsquare.com/files/rs-3773235/v1/7c7309099679e543f3c9f299.zip"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-3773235/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3773235/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Precise capture of three-dimensional (3D) dynamic conformation during molecular representation learning is crucial for accurate prediction of drug targets and molecular properties. In this study, we propose a molecular video-based foundation model, named VideoMol, pretrained on 120 million frames of 2 million unlabeled drug-like and bioactive molecules with 3D conformations. VideoMol renders the molecular 3D conformation as a 60-frame dynamic video and designs three self-supervised learning strategies on molecular videos to capture diverse conformational changes. We demonstrate high performance of VideoMol in predicting molecular targets and properties across 44 benchmark drug discovery datasets. VideoMol achieves high accuracy in identifying antiviral molecules against SARS-CoV-2 across 11 high-throughput experimental datasets from the National Center for Advancing Translational Sciences and other diverse disease-specific drug targets. We further present high interpretability of VideoMol through observed key chemical substructures related to dynamic 3D conformational changes compared to traditional state-of-the-art deep learning approaches. In summary, VideoMol offers a powerful tool to expedite drug discovery and development.","manuscriptTitle":"A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-19 21:26:06","doi":"10.21203/rs.3.rs-3773235/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"ad3f5c77-38bb-4bc9-b568-2566c61f1bf3","owner":[],"postedDate":"January 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":28229269,"name":"Biological sciences/Chemical biology/Cheminformatics"},{"id":28229270,"name":"Health sciences/Medical research/Drug development"}],"tags":[],"updatedAt":"2024-11-09T08:05:49+00:00","versionOfRecord":{"articleIdentity":"rs-3773235","link":"https://doi.org/10.1038/s41467-024-53742-z","journal":{"identity":"nature-communications","isVorOnly":false,"title":"Nature Communications"},"publishedOn":"2024-11-08 05:00:00","publishedOnDateReadable":"November 8th, 2024"},"versionCreatedAt":"2024-01-19 21:26:06","video":"","vorDoi":"10.1038/s41467-024-53742-z","vorDoiUrl":"https://doi.org/10.1038/s41467-024-53742-z","workflowStages":[]},"version":"v1","identity":"rs-3773235","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3773235","identity":"rs-3773235","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}