Abstract
BACKGROUND: Analysis of single-cell RNA sequencing (scRNA-seq) data has revolutionized our understanding of cellular heterogeneity, yet current approaches face challenges in efficiency, interpretability, and connecting molecular insights to therapeutic applications. Despite advances in deep learning methods, identifying cell-type-specific functional gene sets remains difficult. RESULTS: In this study, we present scKAN, an interpretable framework for scRNA-seq analysis with two primary goals: accurate cell-type annotation and the discovery of cell-type-specific marker genes and gene sets. The key innovation is using the learnable activation curves of the Kolmogorov-Arnold network to model gene-to-cell relationships. This approach provides a more direct way to visualize and interpret these specific interactions compared to the aggregated weighting schemes typical of attention mechanisms. This architecture achieves superior performance in cell-type annotation, with a 6.63% improvement in macro F1 score over state-of-the-art methods. Additionally, it enables the systematic identification of functionally coherent cell-type-specific gene sets. We demonstrate the framework's translational potential through a case study on pancreatic ductal adenocarcinoma, where gene signatures identified by scKAN led to a potential drug repurposing candidate, whose binding stability was supported by molecular dynamics simulations. CONCLUSIONS: Our work establishes scKAN as an efficient and interpretable framework that effectively bridges single-cell analysis with drug discovery. By combining lightweight architecture with the ability to uncover nuanced biological patterns, our approach offers an interpretable method for translating large-scale single-cell data into actionable therapeutic strategies. This approach provides a robust foundation for accelerating the identification of cell-type-specific targets in complex diseases.