Active Learning for Molecular-Docking Algorithm Selection under Heterogeneous Runtime

Training an algorithm selector for structure-based drug discovery typically requires evaluating every docking method on every protein--ligand complex. We formulate this costly matrix construction as closed-loop active learning, where each query reveals a docking outcome and runtime. Using MolAS, we evaluate cell-wise active acquisition and compare five predicted-runtime transformations of Mean-UCB under a leakage-safe protocol. Across seven docking methods and five held-out folds, selective acquisition recovers decision-useful training data without exhaustive evaluation. However, no cost-aware formulation dominates universally: runtime transformations reshape algorithm allocation and can suppress expensive but valuable evidence, while matrix-quality metrics do not consistently predict held-out performance. These results establish a reproducible benchmark for reliable, resource-aware active learning in molecular docking.