Preferential Multi-Objective Bayesian Optimization for Drug Discovery

Abstract

Despite decades of advancements in automated ligand screening, large-scale docking remains resource-intensive and requires post-processing hit selection, a step where chemists manually select a few promising molecules based on their chemical intuition. This creates a major bottleneck in the virtual screening process for drug discovery, demanding experts to repeatedly balance complex trade-offs among drug properties across a vast pool of candidates. To improve the efficiency and reliability of this process, we propose a novel human-centered framework CheapVS that allows chemists to guide the ligand selection process through pairwise preference feedback. Our framework combines preferential multi-objective Bayesian optimization with an efficient diffusion docking model to capture human chemical intuition for improving hit identification. Specifically, on a library of 100K chemical candidates that target EGFR, a cancer-associated protein, CheapVS outperforms state-of-the-art docking methods in identifying drugs within a limited computational budget. Notably, our multi-objective algorithm can recover up to 16 out of 37 known drugs while scanning only 6\% of the library, showcasing its potential to advance drug discovery\footnote{Code and data for these experiments can be found at \url{https://anonymous.4open.science/r/vs-9A83}}.