Molecular Predictions Should Remember Temperature, pH, Assay Format, and Biological Context Rather Than Treating Bioactivity as a Context-Free Label

Download PDF

Molecular machine-learning models commonly treat bioactivity as a property attached to a molecular structure, even though experimentally reported activity is produced under particular physicochemical, analytical, and biological conditions. Temperature can alter molecular conformational states and target behavior; pH can redistribute protonation states and modify transport or binding; assay technologies measure different pharmacological levels; and cellular systems can change or confound the response being modeled. This methodological framework argues that these conditions should remain computationally associated with bioactivity rather than disappearing during database harmonization and label construction. The analysis separates molecular identity from experimental context, distinguishes semantic standardization from experimental equivalence, and treats missing contextual metadata as potentially informative rather than automatically neutral. A proposed context-retentive representation is developed in which activity records preserve the measured endpoint together with physicochemical conditions, assay and detection characteristics, biological system, and explicit missingness status. The framework further distinguishes molecule-dependent effects from molecular-by-context interactions and separates generalization to new molecules from prediction under new assays, biological systems, or condition combinations. Its purpose is not to assert that every activity measurement is strongly context sensitive, nor that recording additional metadata guarantees improved prediction. Instead, it provides a structure for determining when contextual information should be represented, tested, withheld, or reported as uncertain. The approach requires prospective evaluation under deliberately shifted conditions and does not establish universal correction rules, mechanistic causality, clinical validity, or deployment readiness.