Cell type identification in spatial transcriptomics data can be improved by leveraging cell-type-informative paired tissue images using a Bayesian probabilistic model

Abstract

The disorganization of cell types within tissues underlies many human diseases and has been studied for over a century using the conventional tools of pathology, including tissue-marking dyes such as the H&E stain. Recently, spatial transcriptomics technologies were developed that can measure spatially resolved gene expression directly in pathology-stained tissues sections, revealing cell types and their dysfunction in unprecedented detail. In parallel, artificial intelligence (AI) has approached pathologist-level performance in computationally annotating H&E images of tissue sections. However, spatial transcriptomics technologies are limited in their ability to separate transcriptionally similar cell types and AI-based pathology has performed less impressively outside their training datasets. Here, we describe a methodology that can computationally integrate AI-annotated pathology images with spatial transcriptomics data to markedly improve inferences of tissue cell type composition made over either class of data alone. We show that this methodology can identify regions of clinically relevant tumor immune cell infiltration, which is predictive of response to immunotherapy and was missed by an initial pathologist’s manual annotation. Thus, combining spatial transcriptomics and AI-based image annotation has the potential to exceed pathologist-level performance in clinical diagnostic applications and to improve the many applications of spatial transcriptomics that rely on accurate cell type annotations.

Cell composition inference and identification of layer-specific transcriptional profiles with POLARIS

“…Imaging information, however, has been under-utilized. Recent work [18, 19] showed the potential of histology images accompanying ST data. We believe that histology images can be further leveraged for ST inference.…”

“…However, cell composition inference hasn’t benefited from these models yet. In recent literature, histology images have been utilized to improve deconvolution accuracy [18, 19] but methods that can predict cell composition solely from histology images are currently unavailable.…”

EnDecon: cell type deconvolution of spatially resolved transcriptomics data via ensemble learning

“…The lack of ground truth for the cell type composition of spots in ST or 10× Visium experiments makes evaluation using real data challenging ( Chen et al , 2022 ; Lopez et al , 2022 ; Zubair et al , 2022 ). To compare the performance of EnDecon with other methods, we conduct an experiment following ( Zubair et al , 2022 ).…”

“…The lack of ground truth for the cell type composition of spots in ST or 10× Visium experiments makes evaluation using real data challenging ( Chen et al , 2022 ; Lopez et al , 2022 ; Zubair et al , 2022 ). To compare the performance of EnDecon with other methods, we conduct an experiment following ( Zubair et al , 2022 ). We use a dataset that measures gene expression in an adult mouse brain using the 10× Genomics Visium platform and performs immunofluorescent (IF) staining on the reverse side of tissue sections for two proteins (GFAP and RBFOX3), which are protein markers specific to glial and neurons cells, respectively ( Fig.…”

“…3a ). In a similar way to Zubair et al (2022) , we use the estimated spot-level intensity of two marker proteins as ground truth for each matched spot ( Fig. 3a ).…”