Bayesian Active Learning for Optimization and Uncertainty Quantification in Protein Docking.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning

Background:

• Protein docking is crucial for understanding molecular interactions but faces challenges in optimization and uncertainty quantification.
• Existing methods lack rigorous uncertainty quantification (UQ) for noisy, high-dimensional optimization problems.

Purpose of the Study:

• To introduce a novel Bayesian active learning (BAL) algorithm for optimization and UQ in flexible protein docking.
• To address the need for reliable confidence estimates in predicting native protein structures.

Main Methods:

• Developed BAL algorithm modeling posterior distribution of global optimum with iterative active sampling and posterior estimation.
• Utilized complex normal modes for high-dimensional conformational space and funnel-like energy models for quality estimation.
• Applied BAL to protein-docking benchmark and CAPRI datasets, including homology docking.

Main Results:

• BAL significantly outperforms rigid docking and particle swarm optimization refinements, achieving top-3 near-native predictions for one-third of targets.
• UQ provides tight quality intervals (25% of actual interface RMSD) with 85% confidence.
• BAL demonstrates strong performance in predicting near-native structures (AUROC 0.93, PR AUC > 0.60).

Conclusions:

• BAL is the first UQ solution for protein docking, offering a rigorous framework and empirical validation.
• The algorithm enhances prediction accuracy and provides reliable quality assessment for protein complex structures.
• BAL improves the ranking of predicted protein complex structures.