AGDIFF: Attention-Enhanced Diffusion for Molecular Geometry Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Accurate molecular geometry prediction is vital for drug discovery and materials science.
• Existing fast methods lack accuracy, while accurate methods are computationally expensive.
• There is a need for efficient and accurate molecular structure prediction tools.

Purpose of the Study:

• Introduce AGDIFF, a novel machine learning framework for efficient and accurate molecular structure prediction.
• Enhance diffusion models for improved molecular geometry prediction.
• Advance computational chemistry, drug discovery, and materials design.

Main Methods:

• Utilized diffusion models for molecular structure prediction.
• Enhanced global, local, and edge encoders with attention mechanisms.
• Improved SchNet architecture, batch normalization, and feature expansion techniques.

Main Results:

• AGDIFF outperformed GeoDiff on GEOM-QM9 and GEOM-Drugs datasets.
• Achieved 93.08% mean COV-R and 0.1965 Å mean MAT-R on GEOM-QM9 (δ=0.5 Å).
• Attained 100.00% median COV-R and 0.8237 Å mean MAT-R on GEOM-Drugs (δ=1.25 Å).

Conclusions:

• AGDIFF demonstrates significant potential for advancing molecular modeling.
• Enables more efficient and accurate prediction of molecular geometries.
• Contributes to progress in computational chemistry, drug discovery, and materials design.