Quantitative structure–activity relationship models are commonly judged by how accurately they predict potency, yet medicinal-chemistry decisions also depend on whether a prediction is being made under conditions in which the model is likely to fail. Aggregate predictive performance can conceal chemically local errors, assay-dependent labels, representation-specific blind spots, and losses of validity under prospective distribution shift. This Methodological Perspective argues that these failure modes should become explicit prediction targets rather than remain post hoc explanations for inaccurate potency estimates. Evidence from molecular machine learning, activity-cliff analysis, assay-aware modeling, multimodal phenotypic prediction, applicability-domain studies, and uncertainty quantification indicates that model reliability is conditional on several separable factors: chemical support, local structure–activity continuity, measurement context, representational adequacy, and deployment regime. On that basis, we develop a proposed distinction between potency prediction and failure-risk prediction. The objective is not to replace conventional QSAR metrics, nor to claim that a universal failure classifier already exists, but to reorganize validation around decision-relevant questions: what could fail, why the current prediction is exposed to that failure, and which additional evidence would discriminate between competing explanations. Such an architecture would make uncertainty more actionable and prospective model evaluation more chemically interpretable.
IJPRAS
