Contribution of protein embeddings to drug–target interaction modeling

Ascertain the specific contribution of protein sequence embeddings (e.g., ESM-1b) to predictive performance in drug–target interaction models beyond ligand embeddings, and characterize the conditions under which incorporating protein embeddings materially affects model accuracy.

Background

Drug–target interaction (DTI) models often combine ligand and protein representations, but empirical studies suggest that ligand features can dominate predictive performance. The extent to which protein embeddings add incremental value remains uncertain.

In this work, ligand embeddings from single- and multi-view molecular models are paired with ESM protein embeddings on the Davis dataset; while this tests the ligand models, the precise value added by protein embeddings is explicitly raised as an open question.