New interpretable machine learning method for single-cell data reveals correlates of clinical response to cancer immunotherapy

Evan Greene, Greg Finak, Leonard A. D'Amico, Nina Bhardwaj, Candice D. Church, Chihiro Morishima, Nirasha Ramchurren, Janis M. Taube, Paul T. Nghiem, Martin A. Cheever, Steven P. Fling, Raphael Gottardo

Research output: Contribution to journalArticlepeer-review


High-dimensional single-cell cytometry is routinely used to characterize patient responses to cancer immunotherapy and other treatments. This has produced a wealth of datasets ripe for exploration but whose biological and technical heterogeneity make them difficult to analyze with current tools. We introduce a new interpretable machine learning method for single-cell mass and flow cytometry studies, FAUST, that robustly performs unbiased cell population discovery and annotation. FAUST processes data on a per-sample basis and returns biologically interpretable cell phenotypes that can be compared across studies, making it well-suited for the analysis and integration of complex datasets. We demonstrate how FAUST can be used for candidate biomarker discovery and validation by applying it to a flow cytometry dataset from a Merkel cell carcinoma anti-PD-1 trial and discover new CD4+ and CD8+ effector-memory T cell correlates of outcome co-expressing PD-1, HLA-DR, and CD28. We then use FAUST to validate these correlates in an independent CyTOF dataset from a published metastatic melanoma trial. Importantly, existing state-of-the-art computational discovery approaches as well as prior manual analysis did not detect these or any other statistically significant T cell sub-populations associated with anti-PD-1 treatment in either data set. We further validate our methodology by using FAUST to replicate the discovery of a previously reported myeloid correlate in a different published melanoma trial, and validate the correlate by identifying it de novo in two additional independent trials. FAUST's phenotypic annotations can be used to perform cross-study data integration in the presence of heterogeneous data and diverse immunophenotyping staining panels, enabling hypothesis-driven inference about cell sub-population abundance through a multivariate modeling framework we call Phenotypic and Functional Differential Abundance (PFDA). We demonstrate this approach on data from myeloid and T cell panels across multiple trials. Together, these results establish FAUST as a powerful and versatile new approach for unbiased discovery in single-cell cytometry.

Original languageEnglish (US)
JournalUnknown Journal
StatePublished - Jul 13 2019

ASJC Scopus subject areas

  • Biochemistry, Genetics and Molecular Biology(all)
  • Agricultural and Biological Sciences(all)
  • Immunology and Microbiology(all)
  • Neuroscience(all)
  • Pharmacology, Toxicology and Pharmaceutics(all)

Fingerprint Dive into the research topics of 'New interpretable machine learning method for single-cell data reveals correlates of clinical response to cancer immunotherapy'. Together they form a unique fingerprint.

Cite this