Machine Learning Framework Identifies Tumor Microenvironment Ecotypes in Blood Samples

Researchers have created a machine learning framework that profiles spatially dependent cell states and multicellular ecosystems in the tumor microenvironment, termed spatial ecotypes. The approach integrates more than 10 million single-cell and spot-level spatial transcriptomes from human carcinomas and melanomas.

It identified nine spatial ecotypes with conserved biology, geospatial features and clinical outcome associations across cancer types. The spatial ecotypes include several linked to response to immune checkpoint inhibitors. Existing methods for profiling these ecosystems have been limited by reliance on predefined markers, lack of spatial information or inability to integrate data across samples and platforms.

Solid tumor biopsies also introduce sampling bias and are typically limited to a single diagnostic sample. Cell-free DNA has been explored as a noninvasive alternative for tumor analysis. The new framework demonstrates that spatial ecotypes are distinguishable by DNA methylation profiling and can be recovered from plasma cell-free DNA using deep learning.

In cell-free DNA from nearly 100 patients with melanoma, the levels of these ecotypes exhibited associations with immunotherapy response that aligned with tumor biopsy results.

The study compiled spatial transcriptomics data from 132 primary human tumor specimens spanning ten cancer types and six platforms. These included bulk, single-cell and single-cell-scale spatial transcriptomics assays. Single-cell RNA sequencing atlases covering 144 tumor samples from matching malignancies were also incorporated.

Integration of the datasets used a tool called CytoSPACE to reconstruct single-cell resolution profiles. This allowed analysis of gene-expression variation between tumor and adjacent stromal regions for eight tumor microenvironment cell types across ten malignancies.