Deciphering tissue architecture with StKAN: A multi-modal deep learning framework combining morphology and spatial transcriptomics.
TL;DR
StKAN is introduced, a novel framework integrating Kolmogorov-Arnold Network with variational autoencoder to effectively model spatially resolved gene expression with graph attention network and shows strong potential for downstream analyses, offering deeper insights into disease pathology and tumor invasion.
Abstract
Spatial transcriptomics facilitates tissue microenvironment analysis by retaining gene expression alongside spatial context, with spatial domain detection being crucial. Conventional clustering or graph-based approaches often fail to capture global spatial dependencies and low-dimensional features due to complex nonlinear patterns and intricate neighborhood structures, limiting both accuracy and generalizability. We introduce stKAN, a novel framework integrating Kolmogorov-Arnold Network with variational autoencoder to effectively model spatially resolved gene expression with graph attention network. StKAN fuses spatial information, gene expression, and optional morphological features, and applies contrastive learning to identify biologically coherent domains. Leveraging explicit function decomposition, it ensures flexible adaptation to diverse data scales. Evaluated on seven spatial transcriptomics datasets, stKAN outperforms existing methods in domain detection accuracy and robustness. It shows strong potential for downstream analyses, offering deeper insights into disease pathology and tumor invasion. By bridging deep learning and spatial context, stKAN advances spatial biology with enhanced generalizability.