KAN-GIN: Adaptive Nonlinear Molecular Representation Learning for Drug-Target Affinity Prediction

Background/Objectives: Accurate drug-target affinity (DTA) prediction supports virtual screening, lead optimization, and drug repurposing. This study investigates whether replacing the conventional post-aggregation multilayer perceptron within a Graph Isomorphism Network drug encoder with a Kolmogorov-Arnold Network (KAN) transformation improves DTA prediction. Methods: The resulting KAN-GIN framework combines molecular-graph representations with a convolutional protein-sequence encoder. It was evaluated on the DAVIS, KIBA, and METZ benchmarks against a matched GIN baseline and representative multiscale and graph-pretraining-based DTA architectures under a common split and evaluation protocol. Complementary analyses examined latent-space organization, gradient-based molecular attribution, and an enhancer of zeste homolog 2 (EZH2) candidate-ranking case study. Results: KAN-GIN consistently reduced prediction error relative to the matched GIN baseline and achieved competitive performance against the additional architectures, including the strongest overall results on METZ. The latent-space findings were dataset-dependent, while gradient-based attribution identified molecular substructures contributing to individual predictions. The EZH2 case study illustrated how KAN-GIN can support candidate ranking before structure-based evaluation. Conclusions: KAN-based post-aggregation transformations represent an emerging design option for graph-based DTA prediction. However, the latent-space and structure-based findings remain computational and should not be interpreted as evidence of mechanistic biological relevance without prospective experimental validation.