NEW YORK – A new computational method promises to advance spatial transcriptomics by overcoming shortfalls in RNA sequencing gene recovery or spatial resolution, leading to a deeper understanding of tumor microenvironments and, ultimately, better cancer treatments.
This method, called CytoSPACE — for Cyto Spatial Positioning Analysis via Constrained Expression alignment — goes beyond earlier techniques by mapping individual cells to a reference single-cell RNA-seq atlas.
In a paper published last week in Nature Biotechnology, researchers from the Stanford University Institute for Stem Cell Biology and Regenerative Medicine described CytoSPACE as an "optimization method for mapping individual cells from a single-cell RNA sequencing atlas to spatial expression profiles."
The authors found that CytoSPACE beat a dozen earlier computational methods in terms of noise tolerance and accuracy to support alignment at single-cell resolution across with multiple tissue types.
The method seeks to improve spatial transcriptomics by filling in previously missing pieces in single-cell sequencing alignment. Much of that is due to current assay design, according to corresponding author Aaron Newman, a computational biologist specializing in cancer at the Stanford University Institute for Stem Cell Biology and Regenerative Medicine.
Newman and colleagues said that 10x Genomics' Visium and other transcriptomics platforms are limited to measuring spatial gene expression from 10 cells or more at a time. Newman noted that while Visium does profile the entire transcriptome, it does not achieve single-cell resolution.
Vizgen's Merscope can produce single-cell resolution, but it is limited to a panel of about 500 preselected genes and thus does not produce a complete transcriptome, according to Newman. NanoString Technologies includes more than 1,000 genes in its relatively new CosMx spatial multiomics imaging platform, but that still covers just a fraction of the transcriptome.
"There is a need to increase spatial resolution for the low-resolution platforms and increase gene recovery for the single-cell, low-gene-recovery platforms," Newman said.
CytoSPACE aligns single-cell and spatial transcriptomes with a technique called global optimization. Newman said that earlier, low-resolution transcriptomics computational methods are based on deconvolution of RNA to identify various cell types.
"That's useful because that tells you the cell types that are present," he explained. "What it doesn't do is tell you what genes they express in each spot."
Newman and colleagues at Stanford benchmarked CytoSPACE against 12 earlier spatial transcriptomic prediction methods, including two notable recent ones: Tangram and CellTrek. According to the Nature Biotechnology paper, CytoSPACE achieved single-cell precision better than every other method across many cell types and at varying spatial resolutions with both simulated data and real melanoma, breast cancer, and colon cancer tumor cells.
While the Nature Biotechnology paper does not have a visual comparison between CytoSPACE, Tangram, and CellTrek, an article published by Stanford includes images showing the difference in resolution from mapping mouse kidney cells. Newman said that previous method 1 is Tangram and previous method 2 is CellTrek.
Even earlier techniques work with probabilities that a specific cell is present in a certain spot of the transcriptome, according to Newman. "The problem with that sort of approach is it's not clear how to use those probabilities to actually do the mapping," he said. "It neglects all of the other cells that actually still belong in the specimen."
CytoSPACE forces every cell to map. "The method is able to capture nuanced expression differences in a given cell type that are spatially different that other methods just aren't able to resolve well," Newman said.
CellTrek and Tangram improve on low-resolution deconvolutional methods, but still fall short in Newman's view, in part because they calculate in isolation, without considering how cells interact with each other. "CytoSPACE … is able to optimally map individual cells and ensure that the mapping is essentially globally optimal," he said.
"We know that gene expression is impacted by the microenvironment and the surrounding community of cells. No cell can exist in isolation," Newman added.
Tangram, developed by Aviv Regev's group at Genentech and the Broad Institute, integrates scRNA-seq and spatial transcriptomic data, then makes predictions through a technique called non-convex optimization. CellTrek, a method from Nicholas Navin and colleagues at MD Anderson Cancer Center, identifies shared embedding between scRNA-seq and spatial transcriptomics data, then relies on random-forest modeling.
The latter requires co-embedding from Spatial Seurat, which the Stanford computational biologists said "can erase subtle yet important biological signals (for example, cell state differences)," citing a preprint article from the Icahn School of Medicine at Mount Sinai.
Tommaso Biancalani, director of artificial intelligence and machine learning at Roche-owned Genentech and Regev's co-lead developer of Tangram, still uses that software in his daily work, and has no plans to change, according to an emailed statement. A new version, called Tangram2, will support extraction of intercellular communication mechanisms as well as better data integration when it comes out later this year.
"A major obstacle to understanding spatial organization in cellular tissues is that we do not currently have the technology to resolve full spatial gene expression," Biancalani said in the email. "It's always exciting to see new methods advance and progress our ability to integrate single-cell and spatial transcriptomics data, and we expect more progress in spatial genomics methods in the near future."
CellTrek developers did not respond to a request for comment.
CytoSPACE is freely available on GitHub for academic developers to add their own inputs and to estimate cell-type fractions. Stanford also put up a web portal for those who just want to run the application.
Newman and colleagues are also developing plans to commercialize CytoSPACE in the next six to 12 months, likely through a licensing agreement between Stanford and a new company. The software will remain free and open to academic users, Newman said.
Newman said that understanding spatial variation between cells will be the "next frontier in precision cancer medicine," helping researchers understand interactions and, ultimately, potential therapeutic targets and pathways.