Chemical genomics language model toward reliable and explainable compound-protein interaction exploration.

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence in medicine

Background:

• Accurate prediction of compound-protein interactions (CPIs) is vital for drug discovery.
• Existing deep learning (DL) models for CPI prediction often lack generalization, confidence quantification, and explainability.

Purpose of the Study:

• To develop a novel chemical genomics language model, ChemGLaM, for reliable and explainable CPI prediction.
• To address the limitations of current DL-based CPI models in generalization, confidence estimation, and interpretability.

Main Methods:

• ChemGLaM integrates independently pre-trained chemical and protein language models using a cross-attention mechanism.
• The model incorporates uncertainty estimation and attention visualization for enhanced interpretability.
• A comprehensive database of CPI predictions was constructed for human proteins and drugs.

Main Results:

• ChemGLaM achieves near state-of-the-art performance in predicting novel CPIs with low computational cost.
• The model demonstrates practical utility in virtual screening, enhancing success rates.
• Attention visualization provides molecular insights into compound-protein interactions.

Conclusions:

• ChemGLaM offers a unified framework for CPI prediction, improving generalization, confidence quantification, and explainability.
• The developed database and model serve as a valuable resource for AI-driven CPI exploration and drug discovery.
• The study highlights the practical impact of ChemGLaM in prioritizing drug candidates and identifying targets, as shown in an amyotrophic lateral sclerosis case study.