Molecular Models Should Adapt at Test Time When Drug-Like Chemical Space Shifts Beyond the Distributions Seen During Training

Download PDF

Molecular machine-learning models are generally optimized before deployment and subsequently treated as fixed predictors, although prospective drug discovery continuously presents compounds, property regimes, assay contexts, and target domains that may differ from those represented during training. This article develops an original computational framework for molecular test-time adaptation under such distribution shift. The analysis distinguishes structural or scaffold displacement, property-support shift, target- or assay-domain change, and mixtures of these conditions from ordinary in-distribution prediction error. It then separates information legitimately available during inference—molecular structure, alternative encodings, neighborhood support, predictive uncertainty, test-batch composition, and retained source information—from unavailable experimental bioactivity labels. The proposed framework treats adaptation as a conditional decision rather than a default response to novelty. Unlabeled self-supervised, consistency-based, entropy-related, transductive, or support-assisted objectives may be activated only when shift evidence, information sufficiency, and reliability criteria justify updating. A parallel risk channel addresses pseudo-label reinforcement, unstable sequential updates, and deterioration of previously reliable predictions, while uncertainty and applicability information support retention, restricted adaptation, rollback, or abstention. The framework further separates label-free test-time adaptation from later active recovery once experimental measurements become available. Its central implication is that prospective molecular prediction should incorporate controlled model plasticity without treating every unfamiliar molecule as an invitation to retrain. The framework remains methodological: molecular test-time adaptation requires prospective validation across chemically, biologically, and operationally realistic shifts before performance, discovery utility, or deployment readiness can be inferred.