spOT-NMF: Optimal Transport-Based Matrix Factorization for Accurate Deconvolution of Spatial Transcriptomics

Abstract

Spatial transcriptomics technologies advance our understanding of complex biology by directly profiling cellular organization within tissues. However, accurate deconvolution of cell types and functional states remains challenging as most current computational methods either rely on high-quality matched single-cell reference profiles (often lacking for many tissues or disease states), struggle across spatial resolutions and under variable sequencing depths, and face scalability bottlenecks in large datasets. To address these challenges, we developed an optimal transport-based non-negative matrix factorization method (spOT-NMF) that leverages the Wasserstein distance to disentangle mixed gene expression signals in a reference-free manner. Benchmarking against well-established unsupervised deconvolution approaches demonstrates top performance of spOT-NMF in simulated and real spatial transcriptomics data spanning sub-cellular to multi-cellular resolutions, across multiple platforms, in two-species admixture scenarios such as xenografts, and in human cancer. We provide spOT-NMF as a freely available package for spatial data analysis, supporting GPU acceleration for large-scale analyses.