Multimodal Drug-Target Affinity Prediction Via FastKAN-Based Hierarchical Fusion of Sequence, Structure, and Tabular...

Drug-target binding affinity (DTBA) prediction is a cornerstone of AI-aided drug discovery. Although multimodal integration and explicit pocket modeling have recently boosted predictive accuracy, existing methods predominantly rely on feature concatenation or fixed-kernel cross-attention mechanisms to fuse heterogeneous modalities, struggling to capture the complex non-linear synergy between chemical structure and biological sequence data, while failing to dynamically quantify inter-modal importance, thereby limiting model interpretability and generalization. To bridge this gap, we propose TabKAN-DTA, a framework centered on a FastKAN-based adaptive gating mechanism that performs input-adaptive nonlinear transformations on heterogeneous modality feature spaces through learnable activation functions, enabling dynamic quantification of inter-modal importance. We further introduce a TabPFN-based tabular encoder to systematically supplement the physicochemical properties and global molecular descriptors absent from sequence and graph models. Comprehensive experiments on standard DTBA benchmarks demonstrate state-of-the-art performance, and ablation studies further reveal that the tabular modality contributes a disproportionately large share of the performance gain. By disentangling modality-specific contributions on a non-small cell lung cancer drug repurposing dataset, we validate that our framework not only improves predictive accuracy but also establishes a new paradigm for interpretable multimodal DTBA prediction. Code is available at https://github.com/trybestxk/TabKAN-DTA.