DUW-MGCA: A Dynamic Uncertainty-Weighted Multi-Granularity Coattention Framework for Protein-Ligand Interaction...

Area of Science:

• Computational chemistry
• Machine learning in science
• Molecular dynamics simulations

Background:

• Bridging the accuracy gap between high-fidelity quantum mechanics (QM) and low-cost molecular dynamics (MD) is crucial for molecular simulations.
• Existing methods struggle to effectively integrate multimodal data from QM and MD.

Purpose of the Study:

• To develop a novel framework, dynamic uncertainty-weighted multigranularity coattention (DUW-MGCA), for joint modeling of QM/MD data.
• To enhance the physical realism of molecular energy landscapes.
• To achieve high accuracy and computational efficiency in molecular simulations.

Main Methods:

• Utilizing an SE(3) equivariant graph neural network to process QM/MD multimodal information.
• Implementing dynamic uncertainty weighting with heteroscedastic uncertainty and soft gating for adaptive modality adjustment.
• Incorporating weak physical regularization constraints to improve energy landscape realism.

Main Results:

• DUW-MGCA significantly outperforms existing baselines on the HiQBind-MISATO dataset.
• Achieved excellent performance metrics: RMSE of 0.979 kcal/mol, MAE of 0.823 kcal/mol, and Pearson R of 0.931.
• Demonstrated superior calibration with low expected calibration error (0.475) and negative log-likelihood (1.398).

Conclusions:

• The DUW-MGCA framework effectively bridges the fidelity gap between QM and MD simulations.
• The method offers a balance between high accuracy and acceptable computational cost, with inference times of approximately 1 minute per complex.
• This approach advances the integration of simulation data for more reliable molecular modeling.