GraESM-FuseDTA: adaptive gated multimodal fusion of graph neural networks and protein language models for robust drug-target affinity prediction.

Accurate drug-target affinity prediction is essential for virtual screening, lead compound prioritization, and drug repositioning, especially under cold-start scenarios involving unseen compounds, unseen targets, or unseen drug-target combinations. However, existing methods still face limitations in transferable molecular representation, contextual protein encoding, statistically reliable performance comparison, and biologically interpretable candidate prioritization. In this study, we propose GraESM-FuseDTA, a graph-enhanced protein language model framework for drug-target affinity prediction. In the drug branch, compound SMILES are converted into molecular graphs and encoded by a residual Graph Isomorphism Network with Edge features, enabling the model to capture atom-bond topology and chemically informative substructures. In the protein branch, a frozen ESM-2 protein language model is used to extract contextual sequence representations without fine-tuning the large protein encoder. To model drug-target compatibility more explicitly, GraESM-FuseDTA employs a gated interaction-aware fusion module that preserves drug features, protein features, element-wise interaction features, difference features, and sample-specific gated mixture features. Experiments on the Davis and KIBA datasets show that GraESM-FuseDTA achieves competitive overall performance and consistent advantages in ranking-oriented and variance-explanation metrics across warm start, drug cold start, target cold start, and strict pair cold start settings. Fold-wise statistical significance tests further support the reliability of most top-ranked metric improvements. Interpretability analyses show that the model identifies functionally relevant target sequence regions within the ESM-visible input segment, adaptively adjusts modality contributions across evaluation scenarios, and forms more organized affinity-related latent representations after gated fusion. In addition, a Yamanishi-based case study with SwissDock molecular docking demonstrates that top-ranked predictions can form structurally plausible binding patterns with favorable calculated affinities. These results suggest that GraESM-FuseDTA provides an effective, interpretable, and practically relevant framework for drug-target affinity prediction.