Anchor Transfer Learning for Cross-Dataset Drug-Target Affinity Prediction

Drug target affinity models often score well on one benchmark but fail on proteins they have not seen before. We study anchor transfer: instead of scoring a protein and a drug on their own, the model is also shown an anchor, a second protein already known to bind the same drug, and judges the new protein against that known example. The anchor is retrieved by chemical similarity at test time, so nothing about the test set is needed. We add this to three kinds of model and compare each against its plain version, training on a hard split where test proteins are at most 30 percent similar to any training protein, then carrying each over to the Davis kinase set. Anchoring helps two of the three, and the pattern differs between them. For AnchorMLP the gain appears only under distribution shift: nothing measurable in distribution, but BindingDB-trained Davis RMSE falls from 1.92 to 1.26 over five seeds. ConciseAnchor gains in error and AUROC on both training sources and on three independent protein splits, though its concordance-index advantage is too small to establish. For AnchorDrugBAN we could not measure a difference over five seeds, and independent protein splits, though its concordance-index advantage is too small to establish. For AnchorDrugBAN we could not measure a difference over five seeds, and report that comparison as inconclusive. Because the anchor model is handed information the plain model never sees, we also score a parameter-free floor that predicts the retrieved anchor's own affinity and ignores the query protein. It separates the two architectures: for AnchorMLP the floor already outranks the plain model and supplies most of the ranking gain, while for ConciseAnchor the plain model clears the floor by a wide margin. The error reduction survives in both, because retrieval alone is badly calibrated. Permuting which drug gets which anchor collapses AUROC from 0.82 to 0.54, so the model reads the anchor relationship rather than exploiting the extra input. Finally, on activity cliffs where a drug binds one kinase far more strongly than a close homolog, the anchor model orders the pair correctly 81 percent of the time while its plain counterpart is at chance.report that comparison as inconclusive. Because the anchor model is handed information the plain model never sees, we also score a parameter-free floor that predicts the retrieved anchor's own affinity and ignores the query protein. It separates the two architectures: for AnchorMLP the floor already outranks the plain model and supplies most of the ranking gain, while for ConciseAnchor the plain model clears the floor by a wide margin. The error reduction survives in both, because retrieval alone is badly calibrated. Permuting which drug gets which anchor collapses AUROC from 0.82 to 0.54, so the model reads the anchor relationship rather than exploiting the extra input. Finally, on activity cliffs where a drug binds one kinase far more strongly than a close homolog, the anchor model orders the pair correctly 81 percent of the time while its plain counterpart is at chance.