Why Can Cross-End Point Transfer Fail in Computational Toxicology? Diagnosing Decision-Structure Incompatibility across 13 End Points under Scaffold-Held-Out Evaluation

Abstract Cross-end point information sharing may improve toxicity prediction when labeled data are limited, but biologically related end points are not necessarily predictively compatible. We evaluated transfer strategies across 13 toxicity end points using scaffold-disjoint repeated cross-validation with strict leakage control. The benchmark compared data-level pooling with higher-level approaches, including stacking, multitask learning, model-agnostic meta-learning, and frozen-encoder feature augmentation (FEFA). Direct cross-prediction between different end points was generally weak, with a mean area under the receiver operating characteristic curve (ROC-AUC) of 0.525 across 156 ordered end point pairs, indicating limited standalone portability of end point-specific decision rules. Global data merging reduced macro-average ROC-AUC by 0.0287 relative to end point-specific baselines, whereas several higher-level approaches produced modest gains. The strongest macro-average improvements were observed for stacking (+0.0101 ROC-AUC) and FEFA (+0.0070), although effects varied substantially among end points. Additional decision-structure and representation analyses were consistent with transfer performance being strategy- and target-dependent rather than determined by biological relatedness alone. These findings support leakage-controlled, end point-specific evaluation when using cross-end point learning for chemical toxicity assessment and caution against indiscriminate pooling of heterogeneous toxicity data.