Abstract
Ab initio protein docking represents a major challenge for optimizing a noisy and costly "black box"-like function in a high-dimensional space. Despite progress in this field, there is a lack of rigorous uncertainty quantification (UQ). To fill the gap, we introduce a novel algorithm, Bayesian active learning (BAL), for optimization and UQ of such black-box functions with applications to flexible protein docking. BAL directly models the posterior distribution of the global optimum (i.e., native structures) with active sampling and posterior estimation iteratively feeding each other. Furthermore, it uses complex normal modes to span a homogeneous, Euclidean conformation space suitable for high-dimensional optimization and constructs funnel-like energy models for quality estimation of encounter complexes. Over a protein-docking benchmark set and a CAPRI set including homology docking, we establish that BAL significantly improves against starting points from rigid docking and refinements by particle swarm optimization, providing a top-3 near-native prediction for one third targets. Quality assessment empowered with UQ leads to tight quality intervals with half range around 25% of the actual interface root-mean-square deviation and confidence level at 85%. BAL's estimated probability of a prediction being near-native achieves binary classification AUROC at 0.93 and area under the precision recall curve over 0.60 (compared to 0.50 and 0.14, respectively, by chance), which also improves ranking predictions. This study represents the first UQ solution for protein docking, with rigorous theoretical frameworks and comprehensive empirical assessments.