SURF: A Self‐Supervised Deep Learning Method for Reference‐Free Deconvolution in Spatial Transcriptomics

Abstract

Spatial transcriptomics has revolutionized tissue biology by enabling spatially resolved gene expression profiling. Nonetheless, current spot-level spatial transcriptomic technologies consolidate signals from multiple cells, complicating cellular-level analysis. Moreover, matched single-cell references required by reference-based deconvolution methods are frequently unavailable. To overcome these limitations, we present SURF, a reference-free deconvolution tool that integrates high-dimensional gene data analysis with self-supervised deep learning to effectively model nonlinear gene interactions and leverage spot relationships. Benchmarking on both synthetic and real datasets shows that SURF consistently outperforms existing reference-free methods and exceeds reference-based approaches when appropriate references are absent. Applications across datasets with varying resolutions, species, spatial patterns, and tissue states demonstrate SURF's robust capacity to precisely represent tissue microenvironments. Importantly, SURF successfully identifies clinically significant epithelial-to-mesenchymal transition states within tumor regions in a dataset of human colorectal liver metastasis, highlighting its utility in uncovering critical biological mechanisms relevant to disease progression.