When Biological Evidence Misleads: Reliability-Aware Transfer for Cross-Scale Drug Response Prediction

Knowledge-guided drug-response models often inject external biological evidence as though its availability implied its usefulness. This assumption is fragile when evidence is sparse, unevenly covered across compounds, or shifted between training and test drugs. We study cross-scale drug-response prediction with a frozen BindingDB drug–protein encoder, pathway-level evidence construction, and a cell-line response model in which the external channel can be suppressed independently of the response prediction. The evaluation is built around matched no-evidence counterfactuals and two distribution-shift protocols: held-out-drug and Bemis–Murcko scaffold-disjoint cross-validation. Across these settings, unconditional fusion can create large drug-specific regret, whereas a training-support density rule reduces the most severe observed harm without using test-time response labels. The rule improves supported-drug macro-AUC in the held-out-drug experiment and preserves a safer worst-drug profile under scaffold-disjoint replication; a provenance-weighted variant is slightly better on the latter mean and regret, so neither mechanism is a uniform winner. Fixed-checkpoint corruption tests show that the density rule is sensitive to out-of-support evidence but remains vulnerable when partial corruption leaves coverage nominally in range. The learned gate confidence is not a calibrated utility score in this data regime, and a SelectiveNet-style coverage constraint reintroduces harmful transfer on an out-of-support drug. These results support a bounded mechanism-level conclusion: evidence availability, support under the training distribution, and realized predictive utility are distinct quantities that should be evaluated separately under drug-level shift.