Description

Single-cell reference maps of tissue architecture using multiparameter imaging and unsupervised representation learning with neural networks

S. S. Bhate^1,2, J. Kennedy-Darling¹, G. P. Nolan¹

¹Baxter Laboratory, Stanford University; ²Bioengineering Dept., Stanford University

Single-cell measurements have been successful at characterizing the composition of cellular populations at the mRNA and protein level. However, cellular populations in tissues are defined not just by phenotype, but also by their single-cell spatial context, which includes morphological and local neighborhood information. Our laboratory has recently created an imaging system termed CODEX, a multiparameter microscopy technology capable of simultaneously imaging over 50 antibody markers. In order to characterize cellular populations in tissues using CODEX, we require unbiased methods to define and compare single-cells by phenotype and spatial context, as well as to use these to visualize tissue architecture in terms of constituent cells. Although most traditional image segmentation techniques can determine cell types by expression given marker sets, morphological and neighborhood analysis (including extracellular protein substrates) are actively ignored.

We have therefore developed an unsupervised method to train a deep neural network to extract features from high parameter tissue images. The extracted features capture spatial and phenotypic information about a cell directly from the image, without segmentation or human input.Applying our NN to a dataset of human immune tissues with over 2 million cells, we produced a concise graphical map describing the cellular composition of the human tonsil. This map recapitulated canonical morphological components of tonsil architecture and cellular phenotypes within them. For example, regions of the map corresponded to the B-cell follicle and its sub-compartments, including the proliferative dark-zone and the FDC rich light-zone.Using our NN on spleen CODEX images with this map as a reference, we found tissue specific differences in germinal center architecture. We also identified with this map a niche of densely packed GC B-cells, in contact with FDCs, highly expressing CD38 and CD9. We will present these and other findings in more detail at the conference.

Our method enables automatic characterization of tissue architecture and cellular niches from large-scale imaging datasets, and use of these as single-cell reference maps for understanding how cellular organization reflects function and mechanism.