A Molecular Video-derived Foundation Model Streamlines Scientific Drug Discovery

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.