Bulk RNA sequencing has been widely used to measure gene activity in tissues, but it has an important limitation. Because it analyzes all cells together, it averages gene expression across many different cell types, making it difficult to determine where individual cells are located or how they interact within a tissue.

Researchers Harbin Medical University have developed a new artificial intelligence framework called HistoMap that reconstructs single-cell spatial information from bulk RNA sequencing data. The approach combines deep learning with tissue images to create a detailed map of where different cell types are located within a sample.

HistoMap uses a two-step strategy. First, it employs a deep learning model called a β-variational autoencoder to estimate the gene expression profiles of individual cells from bulk RNA sequencing data. It then uses a Histological Vision Transformer to position these reconstructed cells within tissue sections by combining gene expression references with information from standard H&E stained pathology images.

Overview of the HistoMap framework for single-cell spatial deconvolution of bulk transcriptomes

(A) Deconvolution of bulk transcriptome data; (B) Image-guided spatial mapping.

The researchers evaluated HistoMap across multiple human tissue datasets and found that it accurately reconstructed spatial cell distributions. The framework achieved strong agreement with reference datasets, demonstrating that it can recover biologically meaningful spatial organization from bulk RNA sequencing data.

To demonstrate its biological value, the investigators applied HistoMap to 14 colorectal cancer samples. The analysis identified a previously unrecognized immune barrier formed by SPP1-positive macrophages at the invasive edge of tumors. These immune cells appeared to organize with surrounding fibroblasts into a physical barrier that limited the movement of cytotoxic T cells into the tumor core.

This finding may help explain why some colorectal cancers respond poorly to immune checkpoint inhibitors. If immune cells cannot effectively enter the tumor, therapies designed to activate them may have limited benefit. The researchers also identified the SPP1 fibroblast signaling axis as a potential therapeutic target that could improve responses to immunotherapy.

As RNA sequencing datasets continue to grow, computational methods that integrate transcriptomics with tissue imaging are becoming increasingly important. HistoMap demonstrates that artificial intelligence can recover valuable spatial information from existing bulk RNA sequencing datasets, providing new opportunities to study cancer biology and identify targets for future therapies.

Availability – The source code and implementation of the HistoMap framework are publicly accessible via GitHub: https://github.com/stat-hj/HistoMap.git