It’s a favorite trope in TV crime dramas: a detective hovers over the shoulder of the resident computer genius, staring at a grainy image on a computer screen and demanding, “Enhance!” With a couple keystrokes, the pixelated image transforms layer by layer into a high-resolution photo of a suspect’s face. Thanks to “TV magic”, a modern day Nancy Drew can infer details and see the big picture at a resolution that simply wasn’t visible before. But there’s a twist – this magic isn’t constrained to TV anymore.
“BayesSpace” is a statistical computational tool developed by the Gottardo, Bielas, and Nghiem Labs and recently published in Nature Biotechnology. BayesSpace was engineered to produce highly detailed spatial transcriptomic data, or in other words, organize what genes are being expressed in precisely which cells spanning across samples with heterogeneous (mixed) cell types; brain tissue, cancers, et cetera. Consider a slice of tissue biopsied from a skin cancer like melanoma, which might contain cancer cells, healthy skin cells, infiltrating immune cells, and by-standing stroma (connective tissue) in a complex and highly heterogeneous array. Every cell in this tissue sample has a unique location and gene expression profile. Current methods for analyzing such a sample force a trade-off between maintaining spatial information and profiling large numbers of target genes in a high throughput way. “Before BayesSpace,” added Dr. Gottardo, the former Scientific Director of the Fred Hutch Translational Data Science Integrated Research Center, “we had to rely on either looking at only a few proteins at a time (via microscopy and antibodies) or looking at RNA from large clusters of cells or the whole tumor.” But now, BayesSpace can bring this kaleidoscope of data into focus computationally.
BayesSpace starts with preprocessed data from a platform like Spatial Transcriptomics (ST) or Visium, both of which are produced by 10X Genomics. These platforms rely on the use of special slides that are coated in small, RNA-binding probes arranged in a lattice formation. The probes in each spot on the slide are designed to bind the same suite of RNA from protein-coding genes but are distinguished by a unique spatial barcode. Tissue sections are placed on the slides and stained for histology, then the RNA-binding probes are used to generate a sequencing library that determines which genes are expressed in which cells and how strongly. Because it was generated on a chip, the dataset shares the chip’s resolution, meaning that a given barcoded “spot” of data might contain information from 30 up to 200 cells. Using spatial statistic methods, BayesSpace assigns the data from this chip to clusters, or groups of similar value. The method first tries to model the data very simply, then smooth the edges between clusters using the known spatial data from the spots in the preprocessed data. Next, it refines these clusters to a “subspot” resolution, essentially breaking the previous groups into sub-groups and re-evaluating the edges of each cluster according to how similar they are to their neighbors. Finally, the gene expression data is mapped onto this more detailed map and can be used to look at differential expression of genes between spots and subspots. This allows BayesSpace to evaluate gene expression at a “near single-cell resolution by computationally ‘Zooming in’,” said Dr. Gottardo.
This level of granularity is important as it allows researchers to identify very specific groups of cells within a broader landscape. Let’s return to the example of a slice of a tumor that harbors both cancer cells and infiltrating immune cells. With BayesSpace, researchers can zoom in to the immune cells and “see what they are doing in the tumor and how the cancer is avoiding destruction,” said Dr. Nghiem, a co-author on the study and Professor in the Division of Dermatology at UW. In the paper, the authors use BayesSpace to identify gene expression patterns consistent with tumor cells, fibroblasts, and immune cells such as macrophages, B cells, and T cells at a ‘subspot’ resolution and map these onto a tissue slice that previously only had annotations for melanoma, stroma, and tumor-proximal lymphoid (immune) tissue. “Understanding how immune cells talk to each other within tumor[s] has been a long-sought dream and could help develop new cancer immunotherapies,” wrote Dr. Nghiem and Dr. Pulliam, other authors on the study.
This complex project involved a diverse cast of researchers from across the Cancer Consortium, from Fred Hutch to UW and SCCA. “This team approach was critical,” said Dr. Gottardo. “The Cancer Consortium allowed us to bring together experts to gain access to data and validate our model and findings using orthogonal data and biological knowledge.” In the future, the team expects to show that BayesSpace is applicable to other platforms beyond those designed by 10X Genomics. Because of how BayesSpace works, the authors also expect it could be applied to “other data types such as [cancer] protein markers and multiomics,” extending their TV magic to multiple realms of biology.
This work was supported by the National Institutes of Health, the Immunotherapy and Data Science Integrated Research Centers at FHCRC, and the Scientific Computing Infrastructure at FHCRC.
Cancer Consortium members Dr. Raphael Gottardo (former), Dr. Jason Bielas, and Dr. Paul Nghiem contributed to this work.
E Zhao, MR Stone, X Ren, J Guenthoer, KS Smythe, T Pulliam, ST Williams, CR Uytingo, SEB Taylor, P Nghiem, JH Bielas, and R Gottardo. 2021. Spatial transcriptomics at subspot resolution with BayesSpace. Nature Biotechnology. 39: 1375-1384.