Frozen sequence encoders with gradient boosting match the state of the art on drug–target affinity domain generalization, and a sequence defect in the Davis kinase benchmark

We report two results on public kinase drug–target interaction (DTI) benchmarks. First, a deliberately simple model — frozen pretrained encoders for molecules and proteins, with gradient-boosted trees on top and no end-to-end training — reaches PCC 0.591 ± 0.004 on the Therapeutics Data Commons DTI-DG benchmark (BindingDB_Patent), evaluated with TDC’s own split and evaluator over five seeds. This is statistically indistinguishable from the best published entry (0.588 ± 0.002; difference 0.003 against a combined standard error of 0.0045, 0.67 SE) and separated from the second entry (0.538 ± 0.008) by 5.9 SE. The result reproduced exactly on an independent run. Second, we document a defect in the widely used Davis kinase dataset that affects any work using its mutant kinases. Davis lists 442 protein names but contains only 379 distinct sequences: all 45 mutation-labelled entries are assigned the wild-type sequence of their gene. ABL1 and ABL1(E255K) have bit-identical ESM-2 embeddings (L2 = 0.0000, cosine = 1.000000). The affinity labels differ while the model input does not, so these pairs are irreducible contradictions in training — 9.63% of all Davis rows. Reconstructing the mutant sequences from the names raises the distinct-sequence count to 423, reduces contradictory rows to 3.97%, and improves concordance index on mutant proteins by +0.0173 (5/5 seeds, t = 23.16) under an otherwise identical protocol.