Wubing Zhang, Erin L. Brown, Abul Usmani, Noah Earland, Minji Kang, Chibuzor Olelewe, Anushka Viswanathan, Prof. Pradeep S. Chauhan, Chloé B. Steen, Hyun Soo Jeon, Susanna Avagyan, Irfan Alahi, Nicholas P. Semenkovich, Janella C. Schwab, Chloe M. Sachs, Faridi Qaium, Peter K. Harris, Qingyuan Cai, Andrew J. Gentles, James Knight, Rondell P. Graham, Antonietta Bacchiocchi, Peter C. Lucas, Ryan C. Fields, Mario Sznol, Ruth Halaban, Di Chen, Aadel A. Chaudhuri, Aaron M. Newman
Multicellular programs in the tumour microenvironment (TME) drive cancer pathogenesis and response to therapy but remain challenging to identify and profile clinically1–3. Here, we present a machine-learning framework for multi-analyte profiling of spatially dependent cell states and multicellular ecosystems, termed spatial ecotypes (SEs). By integrating over 10 million single-cell and spot-level spatial transcriptomes from diverse human carcinomas and melanomas, we identified nine SEs with broad conservation, each of which has unique biology, geospatial features and clinical outcome associations, including several linked to immunotherapy response. Notably, SEs were distinguishable by DNA methylation profiling and were recoverable from plasma cell-free DNA (cfDNA) using deep learning. In cfDNA from nearly 100 patients with melanoma, SE levels exhibited striking associations with immunotherapy response. Our data reveal fundamental units of TME organization and demonstrate a multimodal platform for profiling solid and liquid TMEs, with implications for improved risk stratification and therapy personalization. Multimodal machine learning reveals that tumour microenvironments can be decomposed into spatially organized multicellular ecosystems, termed spatial ecotypes, that can be accessed non-invasively via liquid biopsy and used to profile individual cancers and target treatments.