Understanding where genes are active within tissues is crucial for advancing biology and medicine. Spatial transcriptomics (ST) is a powerful tool that allows scientists to see gene expression patterns in their spatial context. However, a major challenge with many ST technologies is resolution—each measurement often includes a mix of multiple cell types, making it difficult to determine which genes belong to which cells.

Researchers at Tsinghua University heve developed LETSmix, a new computational method designed to address this challenge. LETSmix improves the accuracy of cell-type deconvolution by incorporating spatial information and learning-based techniques to refine data interpretation.

How LETSmix Works

LETSmix stands out by integrating several key factors:

• Layer Annotations & Spatial Correlations: It takes into account the organization of tissue layers, which helps distinguish different cell types.
• Gene Expression & Image Features: By analyzing similarities in gene expression and extracting texture features from tissue images, LETSmix enhances cell-type identification.
• Spatial Coordinates: It considers the precise locations of gene expression spots to refine estimates of cell composition.
• Mixup-Augmented Domain Adaptation: LETSmix addresses differences between spatial transcriptomics and reference single-cell RNA sequencing (scRNA-seq) datasets, making its predictions more reliable across different ST platforms.Importance

Schematic overview of the LETSmix workflow

LETSmix is designed to perform cell-type deconvolution for spatial transcriptomics based on the labeled reference scRNA-seq data. A An adjacency matrix reflecting internal correlations between spots in ST was constructed leveraging information from layer annotations, spot gene expression, histological image texture features, and spot spatial coordinates. B Pseudo-ST data with known cell type compositions were synthesized by randomly selecting cells from the reference scRNA-seq dataset. The real-ST data were refined by the LETS filter constructed from A and then underwent a mixup procedure for data augmentation. C The network structure comprises a shared feature extractor and two classifiers: the source classifier estimates cell type proportions, and the domain classifier identifies spots as real or pseudo. The two branches are trained in an adversarial manner, aiming to eliminate domain shifts in the extracted features. After training the model, the spatially refined real-ST data were directly fed into the feature extractor and the source classifier for cell-type deconvolution

The researchers tested LETSmix on various tissue types and ST technologies, demonstrating that it outperforms existing methods in accurately estimating cell-type proportions and spatial patterns. This advancement has significant implications for research in fields like cancer biology, neuroscience, and regenerative medicine. By improving the resolution of spatial transcriptomics, LETSmix brings us closer to understanding the complex cellular landscapes of tissues and diseases.

As spatial transcriptomics continues to evolve, tools like LETSmix will be essential for unlocking new discoveries in biology and precision medicine.

Availability – https://github.com/ZhanYangen/LETSmix