HistAgent Uses a Standard Stained Tissue Section to Suggest Which Genes to Test in a Specific Region
HistAgent Uses a Standard Stained Tissue Section to Suggest Which Genes to Test in a Specific Region
On August 18, Research Square published a preprint describing HistAgent. The authors trained the model on 2,23 million pairs, each consisting of an image of a region from an H&E-stained section, which is tissue stained using a standard microscopy method, and a measured gene activity profile from the same location. For each region, HistAgent produces a ranked list of 50 genes. The researchers compared these lists with spatial transcriptomics measurements from 135 human and mouse slides that were not used for training. Spatial transcriptomics measures gene activity while retaining each measurement's coordinates within the section.
H&E staining shows cell morphology and tissue architecture, while spatial transcriptomics measures gene activity. HistAgent examines a small region and its surroundings, then ranks the genes most characteristic of that location. This gives the researcher a specific question to test: which molecular features of this region should be examined directly.
The authors chose to rank genes rather than assign continuous numerical scores because their calculations indicated that rankings are less sensitive to data scale, normalization, and differences between measurement batches. To construct a numerical map of gene activity across a section, a separate procedure converts these ranks into estimates using an independent set of previously measured profiles.
For the 50 genes whose activity varied most strongly between locations, the mean correlation between predicted and measured activity maps across 135 slides was 0,440 for HistAgent, 0,157 for STPath, and 0,064 for OmiCLIP. Across five types of analysis, ranging from identifying tissue regions to estimating cell composition, HistAgent outperformed STPath and OmiCLIP in 88 of 90 comparisons covering different combinations of organs and tasks. These two baseline models also predict molecular properties from tissue sections.
The word Agent in the name refers to the next stage. For each region, the program assembles an evidence record containing the gene list, estimated cell composition, biological processes, data from six neighboring locations, and data provenance. When the agent answers a question about the region, it identifies the record fields that support its conclusion.
In a renal cell carcinoma section, the authors searched for a tertiary lymphoid structure, an organized cluster of immune cells. HistAgent identified an area using a 25-gene signature. The evidence record then connected immune response genes with specific cell types and the surrounding connective tissue. The resulting hypothesis was compared with the measured spatial profile and expert annotation.
A standard tissue section can therefore help researchers select a region and formulate a hypothesis about gene activity that can then be tested by measuring activity directly in the tissue.