TrustDTI: Trustworthy Drug–Target Interaction Prediction via Evidential Deep Learning and Selective Abstention
Kyungmin Jeon, Seongyun Park, Daesik Jeong · IEEE Access · 2026
Accurate prediction of Drug-Target Interactions (DTI) is a fundamental step in accelerating the drug discovery process and repositioning existing drugs. Although deep learning models have demonstrated significant potential in DTI prediction, they often suffer from overconfidence and produce unreliable predictions in out-of-distribution (OOD) or cold-start scenarios, posing substantial risks in critical biomedical applications. Furthermore, evaluating these models without rigorous control for data leakage—such as overlapping drug-target pairs between the training and testing sets—often leads to inflated performance metrics. To address these limitations we propose TrustDTI, which is a reliable DTI prediction framework built on an enhanced multi-modal architecture. TrustDTI extracts molecular features using a Hybrid Drug Encoder (combining pretrained KPGT and DrugGNN representations) and protein representations using a graph attention encoder that refines ESM-2 residue features over a residue contact graph derived from predicted 3D structures. A bidirectional co-attention block and gated channel fusion module capture complex inter-modal interactions. Crucially, TrustDTI introduces a Deep Evidential Head that models the output as a Dirichlet distribution, quantifying both the epistemic and aleatoric uncertainties. Coupled with a Selective Head, the model dynamically defers predictions on highly uncertain pairs. Evaluated under a rigorous cross-domain cold-start protocol on BindingDB and DAVIS with exact-match overlap removal, TrustDTI achieves the best Matthews Correlation Coefficient (MCC) across all four cold-start settings; on the in-domain unseen-drug and unseen-protein splits of BIOSNAP, it attains the highest Area Under the Receiver Operating Characteristic curve (AUROC) among all compared methods. More importantly, by abstaining from its most uncertain predictions the model substantially improves reliability, demonstrating that explicitly quantifying uncertainty yields trustworthy decisions in the high-stakes setting of drug discovery.