Abstract
Ultra-large-scale structure-based virtual screening (SBVS) for identifying novel bioactive compounds poses significant computational challenges. These challenges arise from the size of available chemical libraries, which can contain billions of molecules that require exhaustive docking and scoring, placing prohibitive demands on CPU/GPU resources. Small- and mid-sized laboratories often lack access to the high-performance computing clusters or cloud resources necessary to process such workloads in a timely manner. Furthermore, managing and analyzing the resulting terabytes of docking data requires robust data-handling pipelines and expertise that are not universally accessible. Here, we present a data-driven drug development pipeline that leverages a subset of molecules from a database with a common scaffold, reducing the chemical search space by tens to hundreds of orders of magnitude. In this case, the common scaffold that is the key to allowing this reduction is the 2-phenylthiazole moiety, identified through NMR fragment screening. We started with a subset of over 400 000 drug-sized 2-phenylthiazole-containing molecules selected from the zinc database and trained a random forest regression model on about 1% of this data to predict binding scores for the entire library. For this purpose, we used a distribution-preserving sampling approach based on KMeans clustering and binning, and we evaluated its statistical fidelity using KS, Wasserstein, JS, and KL divergence metrics. Our approach preserved the distribution of docking scores, demonstrating the utility of data-driven strategies for scalable virtual screening and establishing a benchmark dataset for machine learning in drug discovery.