Interdisciplinary Sciences: Computational Life Sciences Health Informatics

A Hierarchical KAN-Augmented Graph Convolutional Model for Inferring Cell-Type-Specific Gene Regulations in Single-Cell Omics

Tang, Binhua, Mao, Mengyao, Wang, Jingyu, Zhou, Fengfeng
Springer Nature · 2026-11-20 · DOI 10.1007/s12539-026-00890-z

Abstract

Accurate inference of gene regulatory networks (GRNs) from single-cell RNA-seq data remains challenging due to the complexity of transcription factor (TF)–target interactions and the difficulty in capturing both local and global regulatory dependencies. Here, we present GeneKAN, a novel deep learning framework that integrates multiscale graph convolution with Kolmogorov-Arnold Networks (KANs) for robust GRN inference. GeneKAN employs a kernel-based graph convolution module to hierarchically encode gene representations, while KANs model high-order, interpretable TF-target relationships through learnable spline activations. Benchmarking against eleven existing models, and seven scRNA-seq datasets from BEELINE, including diverse cell types, species, and ground-truth networks (ChIP-seq, STRING, LOF/GOF), demonstrates that our proposed GeneKAN achieves state-of-the-art performance in both AUROC and AUPRC. Ablation studies of the proposed model confirm the necessity of each component. Biological validation on human hepatocyte and lung cancer datasets reveals that GeneKAN not only recovers known regulatory interactions with high precision but also identifies functionally relevant hub genes and oncogenic pathways. GeneKAN thus offers a powerful, interpretable, and scalable solution for GRN reconstruction under real-world conditions of data sparsity and noise, with significant potential for advancing systems biology and precision medicine. Graphical Abstract Overall architecture of GeneKAN. ( A ) Data collection. The model uses single-cell gene expression profiles as input, along with a prior knowledge-based ground-truth GRN that includes annotated regulatory interactions (activation or repression). ( B ) Gene embedding learning module. Multiscale gene representations are learned using Graph Convolution with Kernel Methods (GCKM) equipped with varying kernel widths; wider kernels capture long-range global dependencies, whereas narrower kernels preserved fine-grained local interactions. Additionally, a multi-head attention mechanism is integrated to enhance the quality and discriminability of the resulting low-dimensional gene embeddings. ( C ) Gene interaction prediction module. The embeddings of TFs and their candidate target genes are concatenated and fed into the KAN module, which explicitly models complex, high-order nonlinear regulatory relationships. Finally, it outputs a probability score for each TF-target pair, indicating the likelihood of regulatory interactions.

Read at publisher →