An equivariant pretrained transformer for unified 3D molecular representation learning

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Pretraining on unlabeled 3D molecules enhances performance in scientific applications.
• Existing models often lack cross-domain knowledge integration.
• Leveraging diverse molecular data can improve generalization.

Purpose of the Study:

• Introduce a novel all-atom foundation model for 3D molecules.
• Enable pretraining on multi-domain molecular datasets.
• Enhance generalization across various scientific tasks.

Main Methods:

• Developed an E(3)-equivariant transformer architecture.
• Incorporated atom-level and graph-level feature learning.
• Pretrained the model on diverse 3D molecular datasets.

Main Results:

• Achieved significant improvements in ligand binding affinity prediction.
• Demonstrated competitive performance in predicting protein and small molecule properties.
• Successfully identified and validated potential antiviral compounds against SARS-CoV-2 main protease.

Conclusions:

• The Equivariant Pretrained Transformer effectively leverages cross-domain knowledge for 3D molecules.
• The model shows broad applicability in molecular property prediction and drug discovery.
• This foundation model aids in identifying promising antiviral candidates.