TransScore: A Graph Model for Pose Scoring and Affinity Prediction Based on Transformer Convolution Network.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Molecular docking is crucial for identifying potential drug candidates by predicting protein-compound interactions.
• AI-based scoring functions enhance molecular docking but often lack extensibility and robustness, particularly in cold-start scenarios.
• Existing AI models typically focus on single prediction tasks, limiting their adaptability.

Purpose of the Study:

• To develop a novel deep learning-based graph model for robust protein-compound pose scoring and affinity prediction.
• To address the limitations of extensibility and cold-start performance in current AI-driven molecular docking scoring functions.
• To enhance the accuracy and precision of predicting binding affinities and their relative ordering.

Main Methods:

• A deep learning graph model utilizing a transformer convolution network was developed.
• The model, named TransScore, incorporates a self-attention mechanism to capture protein-compound pose characteristics.
• Gated residual algorithms were employed to enhance adaptability for diverse related tasks.

Main Results:

• TransScore demonstrated superior performance in pose scoring across both cold and warm scenarios, outperforming existing methods.
• The model exhibited robustness on imbalanced datasets and consistent improvements in affinity prediction for both warm and cold start scenarios.
• TransScore achieved high accuracy and precision in predicting binding affinities and their relative ordering.

Conclusions:

• TransScore offers a robust and extensible deep learning solution for molecular docking, significantly improving pose scoring and affinity prediction.
• The model's ability to handle cold-start scenarios and diverse tasks highlights its potential for accelerating drug discovery.
• Analysis of carbonic anhydrase II interactions suggests TransScore can elucidate protein-ligand binding mechanisms, aiding in rational drug design.