Department of Computer Engineering

Drug-target interaction (DTI) prediction remains challenging due to the scarcity of annotated data in a vast input space. We treat DTI prediction as a binary classification problem with two inputs (drugs and proteins) and employ a state-of-the-art cross-attention mechanism to fuse these modalities. However, the data distribution typically differs between training and inference settings, making standard random splits overly optimistic. To simulate realistic conditions, we use dissimilar drug-protein pairs in training and test sets, introducing a domain shift. We then apply domain adaptation to learn a domain-invariant feature extractor, aligning source and target distributions alongside a primary classifier. In particular, we leverage advanced methods, including Maximum Mean Discrepancy (MMD) Loss—which, to the best of our knowledge, has not been used previously for DTI prediction—and adversarial training for robust feature extraction. Our multimodal learning with domain adaptation achieves performance on par with the state of the art on the widely used BindingDB dataset, demonstrating the effectiveness of our approach even under domain shifts.