KA4GANC: A Kolmogorov–Arnold graph attention network approach for predicting gene regulations using single-cell RNA-sequencing data

Abstract

Gene regulatory networks (GRNs) have revealed the internal mechanism and complex relationship of gene expression regulation, and its research is of great significance for the in-depth understanding of life activities and accurate disease diagnosis. Although the existing methods can realize the expression analysis at the cell level with the promotion of single-cell RNA sequencing(scRNA-seq) technology, most focus on the interaction between local genes, and it is difficult to capture the overall organizational structure and long-term regulatory effects. In this study, we present a novel supervised method named KA4GANC, which integrates the Kolmogorov–Arnold Network (KAN) with a Graph Attention Network (GAT) to address the critical limitations in capturing global regulatory architecture from scRNA-seq data. KA4GANC’s novelty lies in two key components. First, it leverages a Fourier KAN for nonlinear feature transformation via adaptive, multi-scale Fourier basis functions, thereby generating highly expressive gene embeddings. Second, it replaces linear transformations in graph attention layers with a KAN-based convolution, enabling the model to learn complex nonlinear local interactions and effectively preserve neighborhood topology in the latent space. Benchmark evaluations on seven scRNA-seq datasets across three ground-truth network types demonstrate KA4GANC’s state-of-the-art performance, achieving average AUROC of 0.84 with 34.6% faster convergence.