Active learning for predicting drug permeability across the blood-brain barrier

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11111903

Predicting blood-brain barrier (BBB) permeability of drugs is a critical task in drug discovery for central nervous system (CNS) disorders. Active learning (AL) offers a promising approach to reducing labeling costs by selecting the most informative samples for training. This study investigates the performance of several AL sampling strategies - random, uncertainty, and dissimilarity - alongside two novel methods: explore-intensify and round-robin cycle switching. We evaluate these strategies using XGBoost models trained on ECFP fingerprints. Experiments are conducted under two data-splitting strategies: label-stratified and scaffold-based splits. Our results show that AL methods match the performance of passive learning while using only 10-65% of the labeled data. In particular, our second novel strategy achieves superior performance. The results demonstrate the effectiveness of dynamic AL approaches in accelerating molecular property prediction with fewer labeled samples.