MolGlueBench reveals heterogeneous transfer of context gains across molecular-glue DC50 domains

Retrospective molecular benchmarks can overstate transfer when chemical and experimental contexts recur across training and test data. We developed MolGlueBench to separate internal scaffold generalization from transfer across fixed molecular-glue DC50 domains. The benchmark contains 1,560 point measurements for 1,137 compounds and 667 Bemis–Murcko scaffolds from four databases, retaining biological, assay, provenance, interaction and missingness context. Under repeated scaffold-disjoint validation, chemistry-plus-context ExtraTrees achieved a Spearman correlation of 0.621 (95% scaffold-bootstrap interval, 0.541–0.689) and a root mean squared error (RMSE) of 0.792 (0.721–0.859) units on the negative base-10 logarithm of molar DC50 (pDC50) scale. Context increased internal Spearman correlation by 0.055 (0.023–0.095), with p = 0.059 in a conditional randomization diagnostic. After matching observations, domain weights and scaffold resamples, the context contrast changed from + 0.037 internally to − 0.086 across four held sources and from + 0.036 to − 0.066 across eight held exact-target-label domains. Effects were heterogeneous: omitting TPDdb reduced source-axis attenuation from − 0.123 to − 0.002, whereas GSPT1 and CDK2 gave the largest adverse target-domain contrasts. Strict scaffold removal and frozen alternative learners likewise showed no consistent transfer advantage. Held-domain macro RMSE values corresponded descriptively to approximately 8–10-fold DC50 errors. MolGlueBench therefore shows that a context increment measured internally need not transfer consistently across the evaluated fixed domains. The benchmark supports retrospective model comparison, not performance claims for future databases, protein-family zero-shot prediction, prospective candidate prioritization or calibrated absolute DC50 prediction. Scientific Contribution MolGlueBench provides a domain-aware evaluation framework that separates internal scaffold-disjoint generalization from transfer across fixed database-source and exact-target-label domains for molecular-glue DC50 regression. Unlike prior degrader benchmarks centred on within-collection performance or individual holdout schemes, it estimates matched chemistry-versus-context increments using aligned observations, domain weights and scaffold-cluster resampling, revealing heterogeneous transfer that aggregate internal metrics obscure. Public reconstruction and CPU analysis workflows make the framework auditable and reusable for testing how molecular representations and validation regimes shape potency-model claims.