ConfDTI: Structure Confidence-Guided Multimodal Fusion for Drug-Target Interaction Prediction

Protein structural information provides critical physical constraints for drug-target interaction (DTI) prediction. However, predicted protein structures inherently exhibit structural uncertainty, and treating them as deterministic inputs may introduce noise and amplify errors during feature aggregation and cross-modal interaction, thereby compromising model generalization. To overcome this issue, we explicitly incorporate structural uncertainty into DTI prediction and propose ConfDTI, a structural-confidence-guided multimodal framework. ConfDTI leverages the AlphaFold2-derived confidence score, predicted local distance difference test (pLDDT), as a residue-level indicator of structural reliability and integrates it into both representation learning and cross-modal interaction. Specifically, structural confidence modulates attention during protein encoding and regulates modality contributions during fusion, enabling adaptive utilization of structural information according to its reliability. By unifying confidence-aware encoding and confidence-modulated fusion within a coherent framework, ConfDTI effectively mitigates the impact of unreliable structural signals and promotes robust interaction modeling. Experimental results on the DrugBank, Davis, and KIBA data sets demonstrate that ConfDTI consistently outperforms representative baselines in both predictive performance and cold-start generalization. Further ablation studies and mechanistic analyses confirm that modeling structural confidence is the primary factor driving these improvements.