When Does Active Learning Improve ADMET Prediction? A Multi-Endpoint Benchmark

Experimental ADMET profiling is essential to drug discovery but incredibly costly that only a fraction of candidate molecules can be measured, motivating active learning to prioritize informative compounds. However, evidence remains limited on how acquisition strategies and uncertainty estimators interact across different ADMET tasks. Thus, we present a protocol-matched benchmark of seven acquisition strategies and two uncertainty approaches (Deep Ensembles and MC Dropout) across eight ADMET regression endpoints. Our framework combines molecular prediction, uncertainty estimation, batch aquisition and empirically calibrated prediction intervals, with a strict chemotype split to evaluate generalization to held-out chemical groups. The results reveal that active-learning benefits were endpoint and strategy-dependent, with the broadest favorable pattern observed for solubility and diversity-aware approaches showing the most consistently favorable behavior across endpoints. Deep Ensembles achieved lower mean RMSE than MC Dropout in 45 of 56 matched comparisons, although this predictive advantage did not consistently translate into better acquisition. Post-hoc analyses associated this disconnect with weak uncertainty-error alignment and reduced chemotype coverage in selected batches. This benchmark offers a reproducible basis for choosing endpoint-specific acquisition strategies, assessing prediction quality alongside computational cost, and designing prospective studies with matched random controls.