Quantitative structure–activity relationship (QSAR) models are often interpreted as if accurate prediction were evidence that a molecular feature participates mechanistically in producing activity. That inference is not generally warranted. Predictive models can exploit correlations arising from scaffold structure, assay context, sampling, representation choice, redundancy, or hidden common causes without identifying the variables through which a structural intervention changes a biological outcome. This Conceptual Perspective separates molecular association, model explanation, mechanistic mediation, and causal effect as distinct inferential objects. Molecular descriptors are treated initially as proxies whose causal status must be established rather than assumed. Dataset confounding, activity cliffs, representation-sensitive explanations, counterfactual edits, and designed perturbations are organized into a proposed causal-QSAR reasoning framework. Stronger causal language is reserved for cases in which structural interventions, mechanistic variables, competing explanations, causal structure, and validation evidence align. The framework is nonempirical and does not claim prospective validation; its purpose is to prevent predictive accuracy or explanation from being mistaken for mechanism.
IJPRAS
