Drug–target interaction (DTI) prediction is a core task in computational drug discovery, where accurate and efficient modeling can substantially accelerate virtual screening. However, drugs and proteins exist as heterogeneous modalities. Many existing approaches emphasize increasingly complex unimodal encoders while modeling cross-modal interaction through relatively simple fusion operations, leading to high computational cost with largely implicit alignment. In this work, we propose X-ADDer, an architecture that explicitly prioritizes cross-modal alignment in DTI modeling. Instead of relying on deep Transformer encoders, X-ADDer employs lightweight graph-informed atom descriptors and convolutional protein filters, coupled with a bi-directional cross-attention mechanism to directly model atom–residue interactions. Evaluations on DrugBank and Davis benchmarks demonstrate that X-ADDer achieves performance comparable to heavy graph Transformer baselines, while consistently improving generalization to unseen drugs (e.g., F1: 0.517 vs. 0.456) and early-stage recall. In addition, X-ADDer reduces training time and GPU memory consumption relative to GraphormerDTI, suggesting that explicit alignment offers a favorable efficiency–performance tradeoff for practical large-scale drug screening.
Liu et al. (Wed,) studied this question.