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

Abstract

Drug target affinity models often score well on one benchmark but fail on proteins they have not seen before. We study anchor transfer, a small change to how the prediction is made. Instead of scoring a protein and a drug on their own, the model is also shown an anchor, a second protein that is already known to bind the same drug. The model then judges the new protein against this known example. The anchor is found by chemical similarity at test time, so no extra information about the test set is needed. We add this idea to three kinds of model and compare each one against its plain version: AnchorMLP, AnchorDrugBAN, and ConciseAnchor. Every model is trained and tested on a hard split, where test proteins are at most 30 percent similar to any training protein, and is then carried over to the Davis kinase set. In every case the anchor version transfers better. The plain ESM model drops to chance on Davis (AUROC 0.43) while AnchorMLP reaches 0.68. AnchorDrugBAN improves from 0.57 to 0.67. ConciseAnchor is the best at 0.84, and it is the only model that gets better when given a stronger anchor, reaching 0.93 with an ideal one. A final test shows where the method stops working. It helps across distant protein families but cannot separate very close proteins that bind the same drug with very different strength.