Mechanism or Memorisation? Stress-Testing Relational Generalisation in Drug–Target Interaction Prediction

Drug–target interaction (DTI) models are often evaluated on held-out interactions for which the drug and target have each appeared elsewhere during training. Such warm-start evaluation establishes that the exact edge is unseen, but it does not establish whether the relationship is supported by analogous training interactions. We introduce relational support, a training-relative evaluation approach that distinguishes familiarity with the individual drug and target from familiarity with their relationship. Using 1,584,215 positive interactions derived from BindingDB, we construct a positive-edge holdout containing 80,094 warm/warm interactions with valid molecular and protein similarity representations. We quantify marginal drug and target familiarity, same-endpoint support, and cross-endpoint relational support using chirality-aware molecular fingerprints and full-length protein sequence identity. We then evaluate three DTI representation families—MoLFormer+ESM-2, ChemBERTa+ESM-2, and Morgan fingerprints+ESM-2—under an identical supervised prediction head that never receives relational-support quantities as input. Across all three representation families, matched average precision was higher for high- than low-relational-support positives: Core-4 $\mathrm{SRG}_X$ ranged from +0.0064 to +0.0099, with all model-level 95% bootstrap confidence intervals above zero. Higher $R_X$ was also associated with lower adjusted positive-pair loss ($\beta_X = -0.0863$ to $-0.1933$). High-confidence false negatives were rare (7–22 of 8,010 top-confidence positives per seed) but sharply localised, with mean $R_X$ 0.363–0.369 lower than for high-confidence correct predictions. Thus, near-ceiling warm-start ranking can coexist with a small, systematically structured error tail in relationally unsupported regions. In practical DTI screening, relationship-level support may help identify apparently familiar drug–target pairs that remain weakly supported by training data and may warrant additional validation before experimental prioritisation.