Finding Drug Candidate Hits With a Hundred Samples: Ultra-low Data Screening With Active Learning

Active learning (AL) can significantly accelerate drug discovery by iteratively selecting informative molecules, reducing experimental workload. However, existing AL studies typically assume access to large datasets, an unrealistic scenario for most academic labs. Here, we investigate AL strategies tailored specifically for small-scale molecular screening, using only 110 affinity evaluations approximated by docking scores from realistic compound libraries: the Developmental Therapeutics Program repository (DTP) and Enamine Discovery Diversity Set 10 (DDS-10). Among 20 tested combinations of molecular descriptors and machine learning models, we identify Continuous and Data-Driven Descriptors (CDDD) combined with a Multi-Layer Perceptron (MLP), augmented by the Pairwise Difference Regression (PADRE) data augmentation technique, as optimal. This combination achieves a 97% probability of discovering at least five top-1% hits from DTP using only 110 affinity evaluations, even under simulated experimental uncertainty. Similarly, the DDS-10 dataset achieves a 100% probability of discovering at least five top-1% hits from DDS-10 using only 110 affinity evaluations. Incorporating prior knowledge by enriching initial datasets with a single known hit molecule further increases this probability. Our findings underscore the feasibility and substantial potential of AL for small-scale drug discovery in resource-limited environments. Our results suggest that early in the AL search the algorithm benefits from accurately quantifying the binding strengths of very weak binders.