Learning Across Bioassays Without Pretending Same Biology via Context-Aware and Assay-Specific Prediction

Download PDF

Bioactivity modeling often gains statistical power by pooling assays, sharing representations, or transferring parameters across related prediction tasks. These strategies are useful, but they can quietly replace a biological question with a convenience assumption that measurements obtained under different assay conditions represent the same response variable up to noise. This methodological article argues that assay context should instead be treated as part of the prediction target whenever experimental conditions, measurement type, biological system, or provenance can alter the mapping from molecular structure to observed activity. The analysis distinguishes molecular representation from assay representation, related-task transfer from label pooling, and assay-specific prediction from universal endpoint modeling. It proposes a context-aware transfer framework in which molecular features are combined with explicit or learned assay-context information, cross-assay sharing is conditioned on estimated task relevance, and assay-specific outputs are preserved where measurements are not biologically interchangeable. The framework also treats conflicting labels, negative transfer, dataset shift, and context novelty as first-class evaluation problems rather than residual error. Its intended contribution is methodological: to organize existing evidence into a disciplined strategy for learning across bioassays without erasing meaningful heterogeneity. The proposal does not establish a universally optimal architecture, mechanistic interpretation, or prospective performance advantage. Its value therefore depends on future tests that separate chemical novelty from context novelty, compare selective against indiscriminate transfer, and assess calibration under assay families not represented during training.