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

Active learning for ADMET regression is endpoint- and strategy-dependent, with solubility and diversity-aware approaches showing the broadest consistently favorable gains across eight endpoints.

• Why it matters: Experimental ADMET profiling is costly, so only a fraction of molecules can be measured; limited evidence exists on how acquisition strategies and uncertainty estimators interact across tasks.
• What they did: A protocol-matched benchmark evaluated seven acquisition strategies with two uncertainty methods (Deep Ensembles, MC Dropout) across eight ADMET regression endpoints using a strict chemotype split and calibrated prediction intervals.
• The result: Deep Ensembles produced lower mean RMSE than MC Dropout in 45 of 56 matched comparisons, but this did not consistently improve acquisition, linked to weak uncertainty–error alignment and reduced chemotype coverage in selected batches.

AI summary of the abstract · may contain errors · not medical advice · read the paper.