Skip to main content
Premium Trial:

Request an Annual Quote

New Method Generates High-Resolution Transcriptome-Wide Spatial Gene Expression Maps

NEW YORK — A method presented in a new paper can predict spatial gene expression from histology pictures, according to its developers.

Most spatial transcriptomic approaches have to balance resolution with the ability to multiplex, but a KTH Royal Institute of Technology-led team has developed a new model to address these constraints. Their approach combines in situ RNA capturing, or ISC, with high-resolution histology images to yield highly resolved transcriptome-wide expression maps. Their benchmarking analyses, reported in Nature Biotechnology on Monday, indicated that the method is robust and sensitive.

"The expression maps resolve micrometer-scale expression signatures that are difficult to isolate in raw ISC data," senior author Jonas Maaskola from KTH and Stockholm University and colleagues wrote in their paper. "Identifying and describing such signatures is vital for characterizing small anatomical features and developing effective treatments for disease states."

As the approach can further predict spatial gene expression from histology pictures with reference ISC images, the researchers said it may also enable image-based in silico spatial transcriptomics.

This super-resolved spatial transcriptomic approach combines ISC data with high-resolution histological image data — typically ones with hematoxylin and eosin stains. In particular, the approach treats spatial gene expression and the histological image data as the effects of a latent tissue state that is then modeled over a number of spatial resolutions. This, the researchers noted, captures both global and local anatomical features. The image data is then mapped to the latent state by a recognition neural network.

The researchers first tested their approach using synthetic datasets to find that it had high accuracy. They likewise found their approach had 95 percent directional correspondence with the ground truth for a dataset of a dozen tissue sections from a mouse olfactory bulb and that their approach matched the reference in situ hybridization data from the Allen Mouse Brain Atlas. These findings suggested to the researchers that their method could separate ISC data into its high-resolution subcomponents.

Using both mouse olfactory bulb and human squamous cell carcinoma datasets, the researchers found that their approach could predict expression from histology images. In particular, they noted that it reproduced ground truth expression patterns and that the accuracy was in line with in-sample performance.

Further, in analyses using human small intestine and other datasets, the researchers found their approach to be robust and sensitive. Occluding part of the image data, they noted, leads to a small decline in the correlation between predicted and ground-truth expression, though visible regions are only slightly affected.

The researchers additionally applied their approach to examine gene expression in the mouse olfactory bulb and in human breast cancer. Within the mouse samples, they noted a number of upregulated and downregulated genes, and that 40 of the 100 most upregulated genes they noted were among the markers recently tied to the mouse olfactory bulb by a single-cell RNA sequencing analysis.

At the same time, they uncovered a number of tumor-related genes that were upregulated within a ductal carcinoma in situ lesion. This included CD74, a marker of metastatic tumor growth in breast cancer, that was expressed near the tumor edge, a finding the researchers said could have treatment implications.

The researchers added that their approach could allow for in silico spatial transcriptomics. "We envision future work to enable ISST on a larger scale, addressing the need to train tissue-specific models," Maaskola and colleagues wrote. "Such models need to be able to flag out-of-distribution samples and train on databases spanning a wide range of anatomical conditions."