MLI-DTA: An interpretable multimodal framework with multi-level interactions for drug-target affinity prediction

Drug-target affinity prediction provides a computational basis for virtual screening and drug repositioning optimization by estimating the binding strength between compounds and target proteins. Existing deep learning methods have evolved from early SMILES/amino acid sequence modeling to graph neural networks and multi-modal fusion. However, there are still three challenges: one-dimensional sequence models struggle to explicitly represent atomic connections and residue interactions; some graph models incorporate molecular or protein contact graphs, but often fail to properly align the semantics of drugs with those of proteins; some multi-modal approaches simply stack multiple source features together, without organizing the interaction processes based on different levels of binding information. To address these issues, this paper proposes MLI-DTA, a hierarchical multi-source feature fusion network based on multi-level attention. MLI-DTA encodes both sequence and graph modalities of drugs and targets. At a finer granularity level, the Bilinear Attention Network models high-level pairings between drug fragments and protein sequence fragments. At a medium granularity level, the Global Graph Cross-Attention mechanism aligns the topological structure of drugs with the contact graph of proteins. Additionally, the Multi-head Collaborative Attention mechanism integrates graph interaction vectors with sequence-level global representations. At a coarser granularity level, the Gate Fusion mechanism dynamically combines global features from both sequences and graphs. Multiple layers of information are combined to create an affinity predictor, enabling the model to utilize both sequence interactions, graph topological semantics, and global context simultaneously. Experiments conducted in the Davis and KIBA benchmarks demonstrate that MLI-DTA performs exceptionally well.