A molecular video-derived foundation model for scientific drug discovery

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate molecular representation is crucial for predicting drug targets and properties.
• Existing methods face challenges in capturing complex molecular information.

Purpose of the Study:

• To introduce VideoMol, a molecular video-based foundation model for enhanced molecular representation.
• To evaluate VideoMol's performance in drug target prediction and property assessment.

Main Methods:

• Pretraining VideoMol on 120 million frames from 2 million unlabeled molecules.
• Rendering molecules as 60-frame videos.
• Employing three self-supervised learning strategies on molecular videos.

Main Results:

• Achieved high performance across 43 drug discovery benchmark datasets.
• Demonstrated high accuracy in identifying antiviral molecules against BACE1 and EP4 targets.
• Showcased superior binding affinity prediction compared to molecular docking.
• Illustrated model interpretability using chemical substructures.

Conclusions:

• VideoMol offers a powerful new approach for molecular representation learning.
• The model effectively aids in identifying potential drug candidates and understanding molecular interactions.
• VideoMol advances the field of AI-driven drug discovery.