Novel Tensor Decomposition-Based Approach for Cell-Type Deconvolution in Visium Datasets

This presentation outlines a novel computational approach for cell-type deconvolution in spatial transcriptomics (Visium), published in Mathematics (2025).
Conventional methods such as RCTD, SPOTlight, SpaCET, and cell2location often fail when reference single-cell RNA-seq data contains multiple minor cell types. In this study, we demonstrate how Tensor Decomposition (TD)-based unsupervised feature extraction (FE) overcomes these limitations.
Key Highlights:
• The Problem: Existing state-of-the-art (SOTA) tools often misidentify cell types or produce inaccurate spatial distributions when handling complex reference data.
• The Solution: A refined TD-based unsupervised learning method that integrates multiple Visium datasets to retrieve spatial gene expression profiles.
• Performance: The proposed method successfully identifies major cell types (Microglia, Neurons, Oligodendrocytes) consistent with biological references, whereas conventional methods fail to do so.
• Efficiency: While Bayesian methods like cell2location can take days to compute, our TD-based approach completes the analysis in just a few minutes.
Paper Reference: Taguchi, Y.-H.; Turki, T. Mathematics 2025, 13, 4028. https://doi.org/10.3390/math13244028

Transcript