SphereDiff-TS: Sphere Space Diffusion Modeling for Accurate 3D Transition State Geometry Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Conventional quantum chemistry methods are computationally expensive.
• Existing machine learning models struggle to incorporate physical constraints effectively.

Purpose of the Study:

• To develop a diffusion-based method for predicting 3D transition state (TS) structures.
• To address the limitations of computational cost and physical constraint incorporation in current models.

Main Methods:

• Introduced SphereDiff-TS, a diffusion model utilizing a spherical coordinate system.
• Incorporated flexible boundary and dynamic radius constraints.
• Evaluated model performance against true transition states.

Main Results:

• Achieved chemical accuracy in geometry prediction (median RMSD: 0.048 Å).
• Demonstrated high accuracy in energy prediction (median absolute error: 0.55 kcal/mol).
• Accurately reproduced reaction barrier heights with deviations below 1.5 kcal/mol.

Conclusions:

• SphereDiff-TS shows potential as a robust computational tool.
• The method can aid in exploring chemical reaction mechanisms.
• It supports computer-driven reaction design.