Multi-Task Learning of Blood–Brain Barrier Permeability and hERG Cardiotoxicity: A Random- and Scaffold-Split Evaluation on Experimental Data

Background/Objectives: Blood–brain barrier (BBB) permeability and cardiac hERG liability share physicochemical determinants, motivating the hypothesis that multi-task learning could improve BBB prediction. Methods: We tested this hypothesis using B3DB logBB regression (n = 1058) and a curated hERG dataset (n = 8960). A shared-encoder multi-task neural network was compared with a matched single-task network and Bayesian-optimized gradient boosting under ten random splits, 30 Murcko scaffold splits, and held-out classification sets. Results: Multi-task learning did not improve random-split R2 (0.485 ± 0.016 vs. 0.482 ± 0.015). Under scaffold splitting, its mean R2 was higher (−0.054 vs. −0.817) only because its failures were less severe (the same number of partitions had negative R2); medians were nearly identical (0.394 vs. 0.398). Neither loss weighting, network tuning, nor removal of 39.6% of test compounds sharing a parent structure with hERG training data changed this. Gradient boosting was more robust (scaffold R2 = 0.388 ± 0.188; better in 28–29 of 30 partitions), although not significantly so under a dependence-corrected test. Conclusions: Network failures occurred only when a scaffold group containing 18 atypical compounds (e.g., xenon) entered the test set; after removing 20 such compounds, multi-task R2 (0.409) matched gradient boosting (0.420). On 3031 held-out parent structures, AUCs were similar (0.764–0.772).