CODA-DTA: A context-guided model for predicting drug-target binding affinity in new compounds and targets

Accurate drug-target affinity (DTA) prediction is important for computational drug discovery, yet existing models often show limited generalization ability for new compounds or protein targets that were not included during model training. One important reason is that most DTA methods use the same prediction rule for all compound-target pairs, which is insufficient to capture the diverse and pair-specific nature of molecular interactions. To address this issue, this study proposes CODA-DTA (COntext-Driven Adaptive Drug-Target Affinity), a context-guided prediction model designed to adapt its prediction process to each specific compound-target pair. The proposed model combines a pretrained chemical sequence model, ChemBERTa-2, to represent compounds from SMILES strings, with a pretrained protein sequence model, ESM-2, to represent target amino acid sequences. It further uses a lightweight pair-guided matching module to adjust the compound-target comparison according to the current input pair. Experimental results on the Davis and KIBA datasets demonstrate that CODA-DTA achieves superior performance, particularly under cold-drug and cold-target settings. Component-removal experiments further verify the effectiveness of the proposed context-driven adaptive interaction mechanism and low-rank conditional parameterization. These results suggest that context-driven adaptive interaction learning is an effective strategy for robust cold-start DTA prediction and disease-oriented drug prioritization.