Virtual staining overlay enabled combined morphological and spatial transcriptomic analysis of individual malignant B cells and local tumor microenvironments

Zihang Fang, Raymond Kozikowski, Kevin de Haan, Serge Alexanian, Michael E. Kallen, Alyssa Rosenbloom, Charlie Glaser, Mark Conner, Yan Liang, Kyla Teplitz, Joachim Schmid, Jaemyeong Jung, Yair Rivenson

Keywords: Deep Learning, Histology, Digital Pathology

TL;DR: A deep learning model was used to virtually stain tissue slides so that pathologists can identify and precisely sample regions of interest that would be used in a spatial transcriptomics pipeline.

Abstract: B-cell lymphomas are complex entities consisting of a component of malignant B-cells admixed in a local tumor microenvironment (TME) inflammatory milieu. Discrete characterization of both compartments can drive deeper understanding of pathophysiology, allowing more accurate diagnoses and prognostic predictions. However, limitations in both pathologist time and input tissue to generate multiple stains can greatly limit accurate identification of minute, cellular-level regions of interest necessary to achieve the full potential of spatial biology. Here, we present a novel method to perform precise sampling of cells for transcriptomic analysis using virtual staining of autofluorescence images via deep learning algorithms. We validated the performance of the model on regions of interest (ROIs) identified on chemically stained images by board certified pathologists against virtually stained images. The results confirmed the usability and accuracy of the workflow and identified distinct transcriptomic profiles across a range of virtually identified ROIs, raising the possibility of our workflow’s applicability across a broader range of pathologies and tissue types.