A deep learning method to map tissue architecture

Abstract

A new study in Nature Methods describes a computational method named UTAG (unsupervised discovery of tissue architecture with graphs) that aims to identify and quantify higher-level tissue domains from biological images without previous knowledge.

Changes in patterns at the tissue and organ level can be a sign of disease development, but combining single-cell data with maps of tissue structure to reveal these patterns relies on time-consuming manual annotation. A new computational method named UTAG (unsupervised discovery of tissue architecture with graphs) promises to identify and quantify organ-specific micro-anatomical domains from biological images without previous knowledge.

The central concept of UTAG is that it leverages multiplexed images to extract and then combine single-cell phenotypical data (such as cell morphology and gene or protein expression) with information on the physical proximity of cells in their native tissue context. Images can be obtained by a range of highly multiplexed single-cell imaging methods, including CODEX (codetection by indexing), CyCIF (cyclic immunofluorescence), IMC (imaging mass cytometry), MIBI (multiplexed ion beam imaging) and multiplexed spatial platforms.

Coolstuff/Alamy Stock Photo

UTAG uses unsupervised deep learning to convert cellular phenotypes, such as expression intensities, into a numeric feature matrix, which comprises every marker for each cell. In addition, the model constructs a graph of physical cellular interactions based on the spatial location of cells in the image. This graph is then converted to a binary adjacency matrix (0 for no interaction, 1 for interaction between cells). The two matrices are combined to create a new matrix comprising spatially aggregated phenotypic information, which can be used by clustering methods to group phenotypically and spatially related cells into tissue domains.

The team first tested UTAG on 26 highly multiplexed IMC images of healthy lung tissues, comprising 28 markers. To serve as a reference, micro-anatomical domains, such as airways and vessels, were manually annotated, and UTAG was able to accurately identify corresponding micro-anatomical domains.

Applying their method to diseased lung tissues, using a dataset of 239 IMC images with 37 markers, identified 6 micro-anatomical domains across different disease states, representing airways, vessels and connective tissue, among others. The relative abundance of the domains was dependent on disease state, with, for example, a large proportion of connective tissue in samples of a patient with late-stage COVID-19, which is indicative of fibrosis. The authors were also able to quantify physical interactions between domains, revealing structural differences between healthy and diseased tissues.

“group phenotypically and spatially related cells into tissue domains”

The authors note that the successful application UTAG depends on “the interpretation of the discovered topological domains in terms of their identity and biological relevance”, and thus recommend the involvement of field experts such as pathologists when applying their approach.