GADIFF: a transferable graph attention diffusion model for generating molecular conformations.

Area of Science:

• Computational chemistry
• Machine learning
• Molecular modeling

Background:

• Diffusion generative models show promise in various research fields.
• Generating accurate and diverse molecular conformations is crucial for drug discovery and materials science.
• Existing models face challenges in capturing both local and global molecular information effectively.

Purpose of the Study:

• To propose GADIFF, a transferable graph attention diffusion model for molecular conformation generation.
• To enhance feature representation by integrating local and global molecular information.
• To develop a transferable model for noncovalent interaction (NCI) molecular systems.

Main Methods:

• GADIFF employs Graph Isomorphism Networks (GIN) to capture local subgraph information with diverse edge types.
• Multi-head Self-attention (MSA) is used as a noise attention mechanism to capture global molecular context.
• Dynamic noise weights calculated by MSA are utilized to improve molecular conformation noise prediction.

Main Results:

• GADIFF achieves competitive performance in generation diversity (COV-R, COV-P) and accuracy (MAT-R, MAT-P) on GEOM-QM9 and GEOM-Drugs datasets.
• The model shows a 3.75% improvement in average COV-R on the GEOM-Drugs dataset compared to the best baseline.
• The transferable GADIFF-NCI model successfully generates reasonable conformations for noncovalent interaction systems.

Conclusions:

• GADIFF offers improved molecular conformation generation by effectively integrating local and global features.
• The transferable nature of GADIFF, demonstrated by GADIFF-NCI, highlights its potential for studying multi-molecular systems.
• The proposed model provides a valuable tool for advancing molecular modeling and computational chemistry research.