Clear Sky Science · en

HarveST uses a heterogeneous graph learning framework to reveal spatial transcriptomics patterns

2026-03-27 · Back to index

Seeing Where Genes Act Inside Tissues

Our bodies are built from neighborhoods of cells that each do different jobs, from thinking in the brain to fighting tumors in cancer. A growing technology called spatial transcriptomics lets scientists map which genes are active in different spots across a tissue slice. But turning these raw maps into clear, meaningful regions and lists of important genes is still difficult. This study introduces a new computer method, HarveST, that helps researchers see tissue structure and its key genes more clearly, even in complex organs and tumors.

Why Mapping Cell Neighborhoods Is Hard

In tissues, nearby cells often work together as functional communities, such as layers in the brain or zones around a tumor. Spatial transcriptomics measures gene activity at thousands of small spots across a tissue, along with their exact locations. Existing computer tools try to group these spots into spatial domains, regions that share both similar gene activity and physical proximity. Many methods either ignore spatial location or assume that neighboring spots must belong to the same region. This can blur important boundaries, miss rare cell pockets, and fail to link each region to the genes that actually define it.

A New Way to Connect Space and Genes

HarveST tackles these problems by turning each dataset into a rich network that links tissue spots and genes at the same time

Figure 1. How a unified network of tissue spots and genes reveals clear regions inside organs and tumors.

. In this network, spots connect to nearby spots, spots connect to the genes they express, and genes connect to each other when they tend to turn on together. This “heterogeneous graph” lets the method consider physical layout, gene activity, and gene-gene relationships in a single model. HarveST first uses a self-supervised learning step, similar to an autoencoder, to compress this complex network into a cleaner, low-noise representation. It then uses a two-stage strategy: it roughly groups spots, identifies the most confident assignments, and uses these as guides to refine the final regions, sharpening borders and protecting small but important structures.

Finding the Right Genes for Each Region

Beyond drawing maps of tissue regions, HarveST aims to find the marker genes that make each region unique. Instead of testing each gene in isolation, the method sends out a “random walk” from spots inside a chosen region and lets this signal spread through the network

Figure 2. How linking tissue spots and gene interactions sharpens tissue regions and highlights region-specific genes.

. Genes that end up with high scores are those that sit close to that region in both space and the gene network. This approach naturally favors genes that fit the overall tissue context, rather than just showing a numerical difference. The authors then use standard tools to check which biological pathways these genes belong to, linking each spatial region to functions such as nerve signaling, immune activity, or cancer invasion.

What HarveST Reveals in Brains and Tumors

To test HarveST, the team applied it to human brain, breast cancer, pancreatic cancer, and mouse olfactory bulb samples measured on several spatial transcriptomics platforms. In human prefrontal cortex, HarveST more accurately recovered the classic six cortical layers and white matter than fourteen competing methods, scoring higher on a broad set of clustering and boundary metrics. In high-resolution mouse olfactory bulb data, it stood out by separating thin, closely packed layers that other tools blended together. In breast and pancreatic cancers, HarveST better distinguished tumor edges, normal regions, and invasive cancer zones, and linked them to genes already known to mark malignancy, immune involvement, and tissue remodeling, as well as highlighting additional, plausible candidates.

Consistent Maps Across Tissue Slices

Real tissues are three-dimensional, but spatial transcriptomics typically measures thin two-dimensional slices. HarveST can jointly analyze several consecutive slices by building a shared network that respects each slice’s internal layout while also connecting similar regions across slices. In paired human brain sections, this joint analysis improved alignment between slices and produced continuous cortical layers that extended smoothly from one section to the next. This suggests HarveST can help reconstruct more faithful three-dimensional views of tissue architecture from standard experiments.

What This Means for Future Biomedical Research

In summary, HarveST offers a unified way to go from raw spatial gene maps to clear tissue regions and their key genes, while respecting both spatial structure and molecular relationships. For non-specialists, this means scientists gain sharper “weather maps” of gene activity inside organs and tumors, and more trustworthy gene lists that explain why one region differs from another. As spatial technologies advance, approaches like HarveST may become central tools for understanding how healthy tissues are organized, how diseases like cancer reshape that organization, and where new diagnostic markers or treatment targets might be found.

Citation: Feng, J., Yu, T. & Zhang, Y. HarveST uses a heterogeneous graph learning framework to reveal spatial transcriptomics patterns. Commun Biol 9, 681 (2026). https://doi.org/10.1038/s42003-026-09841-2

Keywords: spatial transcriptomics, tissue architecture, graph learning, cancer microenvironment, marker genes