Quantitative structure–activity relationship models are commonly developed as bounded predictive artefacts, yet their evidential conditions may continue to change after deployment. Incoming compounds can occupy unfamiliar chemical space, assay procedures and biological contexts can alter endpoint meaning, labels can be corrected or reinterpreted, molecular representations and software environments can change, and prediction consequences can expand beyond the original context of use. Treating the deployed model as static therefore risks preserving an unchanged model identifier while the scientific object being predicted, the data-generating process, or the supported decision has changed. This Original Lifecycle-Governance Architecture Article proposes a continually learning QSAR lifecycle in which chemical-space state, endpoint contract, data and label state, model and representation version, and qualification state are governed as interacting controlled objects. The architecture separates surveillance from adaptation: drift indicators initiate investigation and change-impact assessment rather than automatic retraining. Updating options range from data correction and recalibration to retraining, representation replacement, endpoint remodelling, quarantine, or withdrawal. Revalidation is proposed as the generation of context-specific evidence regarding predictive behavior, applicability, calibration, and uncertainty, whereas requalification determines whether the complete changed system remains supportable for its intended use. Performance monitoring incorporates chemical-space coverage, predictive uncertainty, calibration, and abstention, while auditability requires traceable change records, human authorization, reproducible version bundles, and rollback capability. The contribution is an original conceptual integration rather than a validated operating standard. Its propositions require retrospective stress testing, prospective evaluation, reproducibility assessment, and implementation studies. No universal drift threshold, retraining interval, uncertainty method, requalification category, regulatory pathway, or deployment-readiness claim is established.
IJPRAS
