MMPKAN-DTA: A Multimodal Framework for Drug-Target Affinity Prediction Using Pretrained Models and Kolmogorov-Arnold Networks

Drug-target affinity (DTA) prediction is crucial for advancing drug discovery and development, facilitating candidate prioritization and reducing experimental costs. However, accurate affinity prediction remains challenging, and most deep learning models often fail to generalize in scenarios involving novel drugs or targets (known as cold-start scenarios). This limitation is due to simplified feature representations, limited exploitation of complementary information, and insufficient capacity to model the complex nonlinear relationships in heterogeneous inputs. To overcome these limitations, we introduce a multimodal deep learning framework, MMPKAN-DTA, that combines pretrained features with GNN-derived features to learn representations for both drugs and proteins. The model incorporates a multimodal feature fusion block and Kolmogorov-Arnold Networks (KANs), which enhance representational expressivity by exploiting complementarity and nonlinear approximation capacity, thereby enabling a more effective and efficient model for drug-target affinity predictions. Comprehensive experiments on three benchmark datasets demonstrate that MMPKAN-DTA outperforms baseline methods and exhibits strong generalization ability. Moreover, a case study on the epidermal growth factor receptor (EGFR) and tyrosine-protein kinase ABL1 highlights the effectiveness of MMPKAN-DTA in predicting affinities for unseen drug-target pairs, supporting its applicability as a promising tool for drug discovery and development.