TY  - JOUR
T1  - QSAR without False Confidence: A Model-Evidence Dossier for Applicability Domains, Calibration, Dataset Shift, and Decision-Relevant Uncertainty
A1  - João Silva
A1  - Pedro Costa
A1  - Ana Beatriz
JF  - International Journal of Pharmaceutical Research and Allied Sciences
JO  - Int J Pharm Res Allied Sci
SN  - 2277-3657
Y1  - 2025
VL  - 14
IS  - 1
DO  - 10.51847/C1XMwpDWx2
SP  - 90
EP  - 100
N2  - Quantitative structure activity relationship models support molecular prioritization, property estimation, virtual screening, and evidence preparation, yet apparently strong performance can create false confidence when detached from data provenance, endpoint meaning, chemical space coverage, applicability domains, calibration, dataset shift, and decision consequences. This article proposes the QSAR Model-Evidence Dossier, or Q-MED, as an original nonempirical architecture for organizing those dimensions around a declared intended use. Q-MED links an intended use contract, data and endpoint ledger, representation and modeling record, chemical space and applicability profile, calibration and uncertainty record, shift and challenge record, version and requalification log, and human accountability record through claim evidence traceability. A prediction is not decision eligible merely because an aggregate metric is favorable. Eligibility is conditioned on endpoint compatibility, relevant chemical support, evaluated uncertainty behavior, absence of unresolved shift, an identified model version, and accountable authorization. External validation and challenge sets test bounded claims across novel scaffolds, activity cliffs, sparse regions, altered assays, and other distributional changes. Monitoring, change control, and drift triggered requalification extend evidence assessment across the lifecycle. Defined human roles separate development, validation, scientific interpretation, decision authorization, monitoring, and retirement. Q-MED is a proposed conceptual synthesis, not a validated quality system, clinical recommendation, regulatory pathway, or deployment framework. Its value requires prospective evaluation of traceability, calibration, failure detection, selective prediction, external generalization, reviewer consistency, governance burden, and decision effects. It cannot compensate for invalid endpoints, biased data, inappropriate representations, unsupported causal interpretation, or inadequate experimental confirmation.
UR  - https://ijpras.com/article/qsar-without-false-confidence-a-model-evidence-dossier-for-applicability-domains-calibration-data-dwtiasd4vyklxut
ER  -