Kolmogorov-Arnold Networks (KANs) have recently emerged as a promising alternative to conventional multilayer perceptrons (MLPs) for scientific machine-learning applications. Their rapid adoption has stimulated several KAN-based methods in genomics and bioinformatics. In this study, we evaluate the utility of KANs for single-cell RNA-sequencing (scRNA-seq) analysis.
We benchmark a KAN architecture against a parameter-matched MLP for cell-type annotation across simulated, synthetic, and real scRNA-seq datasets. Hyperparameter tuning is performed independently for each model. We find that KANs yield consistently higher validation accuracy on datasets exhibiting continuous cellular trajectories, such as differentiation processes. We attribute this improvement to the implicit regularisation induced by the smooth B-spline basis functions underlying KAN layers. To assess interpretability, we conduct a case study on real scRNA-seq data and compare KAN-derived feature attributions with SHAP values computed for the MLP.
Beyond annotation tasks, we explore the use of KANs for gene regulatory network (GRN) inference from temporal scRNA-seq. The smooth and interpretable spline formulation of KANs makes them a candidate for modelling gene-gene relationships. We integrate a KAN module into the dynGENIE3 framework and also develop a convolution-style KAN variant designed to incorporate temporal structure. Using the BoolODE simulation platform, we show competitive or superior performance in GRN inference methods including GENIE3, SCODE, and SINCERITIES. Ongoing evaluation with real scRNA-seq data sets will also be discussed.