TFRegNCI: Interpretable Noncovalent Interaction Correction Multimodal Based on Transformer Encoder Fusion.

Area of Science:

• Computational chemistry
• Machine learning
• Deep learning

Background:

• Interpretability is crucial for end-to-end learning models in scientific applications.
• Accurate prediction of noncovalent interactions (NCIs) is essential for understanding molecular behavior.
• Existing models may lack transparency or efficiency in NCI prediction.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate NCI prediction.
• To enhance model efficiency by using 2D electron density inputs.
• To validate the model's interpretability and feature extraction capabilities.

Main Methods:

• Proposed TFRegNCI model combining RegNet and Vision Transformer for feature extraction.
• Utilized a transformer encoder for feature fusion, outperforming multilayer perceptrons.
• Implemented Gradient-weighted Regression Activation Mapping (Grad-RAM) for feature visualization.

Main Results:

• Achieved high prediction accuracy (MAE ~0.1 kcal/mol) compared to CCSD(T) benchmark.
• 2D electron density inputs reduced computation time by 30% without sacrificing accuracy.
• Grad-RAM visualizations confirmed effective feature extraction and alignment with NCI isosurfaces.

Conclusions:

• TFRegNCI offers a promising, interpretable, and efficient tool for NCI prediction.
• The model's feature visualizations provide insights into electronic theory underlying NCIs.
• The approach demonstrates the potential of deep learning for advancing computational chemistry.