Trustworthy Compound-Protein Interaction Prediction with Interpretable and Conformalized Cross-Attention...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Deep learning accelerates virtual screening but lacks reliability guarantees.
• Compound-protein interaction (CPI) prediction is crucial but faces challenges with imbalanced and noisy data.
• Current models often function as "black boxes," hindering trust and experimental validation.

Purpose of the Study:

• To introduce ConfBiXtCPI, an integrated framework for accurate, interpretable, and uncertainty-quantified CPI prediction.
• To address data imbalance and enhance prediction reliability in drug discovery.
• To enable principled control over false discovery rates in virtual screening.

Main Methods:

• Developed a bidirectional cross-attention transformer for sequence-level molecular recognition.
• Incorporated Mondrian conformal prediction for guaranteed coverage across imbalanced datasets.
• Implemented a conformal selection procedure for controlled false discovery rates.

Main Results:

• Achieved state-of-the-art accuracy on multiple CPI prediction benchmarks.
• Demonstrated mechanistic interpretability via attention maps localizing to binding sites.
• Showcased uncertainty quantification supporting efficient active learning strategies.

Conclusions:

• ConfBiXtCPI offers a trustworthy and practical tool for drug discovery.
• The framework unifies accuracy, interpretability, and rigorous uncertainty quantification.
• Enables efficient experimental validation and accelerates the discovery of novel therapeutics.