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

Abstract

7We introduce a non-parametric method for unbiased cell population discovery in single-cell flow 8 and mass cytometry that annotates cell populations with biologically interpretable phenotypes 9 through a new procedure called Full Annotation Using Shape-constrained Trees (FAUST). 10 We used FAUST to discover novel (and validate known) cell populations associated with 11 treatment outcome across three cancer immunotherapy clinical trials. In a Merkel cell carcinoma 12 anti-PD-1 trial, we detected a PD-1 expressing CD8+ T cell population -undetected both by 13 manual gating and existing computational discovery approaches -in blood at baseline that was 14 associated with outcome and correlated with PD-1 IHC and T cell clonality in the tumor. We also 15 validated a previously reported cellular correlate in a melanoma trial, and detected it de novo 16 in two independent trials. We show that FAUST's phenotypic annotations enable cross-study 17 data integration and multivariate analysis in the presence of heterogeneous data and diverse 18 immunophenotyping staining panels, demonstrating FAUST is a powerful method for unbiased 19 discovery in single-cell data. 20 1 Introduction 21 Cytometry is used throughout the biological sciences to interrogate the state of an individual's 22 immune system at a single-cell level. Modern instruments can measure approximately thirty (via 23 fluorescence) or forty (via mass) protein markers per individual cell [1] and increasing throughput 24 can quantify millions of cells per sample. In typical clinical trials, multiple biological samples are 25 measured per subject in a longitudinal design. Consequently, a single clinical trial can produce 26 hundreds of high-dimensional samples that together contain measurements on millions of cells. 27To analyze these data, cell sub-populations of interest must be identified within each sample. 28 The manual process of identifying cell sub-populations is called "gating". An investigator gates a 29 single sample by sequentially inspecting bi-variate scatter plots of protein expression and grouping 30 cells with similar expression profiles together. Each sample is gated according to the same scheme, 31 ⇤ Corresponding author 2 FAUST Method 3 and samples are usually compared on the basis of the frequencies of cells found within each cell 32 sub-population. 33Manual gating introduces the potential for bias into cytometry data analysis [1, 2]. One source 34 of bias is the choice of gating strategy, since it is fixed in advance and is only one of many possible 35 strategies to identify a cell phenotype. A different strategy can lead to different gate placements 36 and consequently different cell counts. A more serious source of bias arises from the fact that 37 manual gating only identifies cell populations deemed important a-priori by the investigator. Since 38 the number of possible populations grows exponentially with the number of measured protein 39 markers, manual identification cannot be used to perform unbiased discovery and analysis...

1-Methylnicotinamide is an immune regulatory metabolite in human ovarian cancer

“…2f-i). These results were confirmed through automated assignment of cell phenotypes using machine learning method 14 , which further revealed an abundant myeloid cell population (CD3-/CD4+) predominately in patient ascites that displayed the highest glucose uptake and mitochondrial activity of any identified cell type (Extended Data Fig 3). These results underscore strong metabolic differences across different cell types found in the ascites and tumors of HGSC patients.…”

“…To supplement our manual gating strategy for the above flow panel, we used Full Annotation Using Shape-constrained Trees (FAUST) 14 to automatically assign cells to populations, after dead cell exclusion in FlowJo. We manually curated outputs to merge populations that appeared to be mis-assigned (merged PD1+ with PD1-tumor cells), and retained populations comprising, on average, more than 2% of cells in each sample, for a total of 11 populations.…”

muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data

“…It is worth noting that extensive workflows for DS analysis of high-dimensional cytometry data have been established [12][13][14][15] , along with a rich set of visualization tools and differential testing methods 13,[16][17][18] , and applied to, for example, unravel subpopulation-specific responses to immunotherapy 19 . Notably, aggregation-based methods (e.g., representing each sample as the median signal from all cells of a given subpopulation) compare favorably in (cytometry) DS analysis to methods that run on full cell-level data 17 ; however, in the cytometry case, only a limited range of cell-level and aggregation approaches were tested, only simplistic regimes of differential expression were investigated (e.g., shifts in means), and the number of features measured with scRNA-seq is considerably higher (with typically fewer cells).…”

New cytometry tools for immune monitoring during cancer immunotherapy

“…For clinical studies, combining these tools into robust, reproducible, and easy‐to‐use data analysis pipelines will be critical to identify verifiable markers of patient response (Conrad et al, 2019). For example, the machine learning algorithm Faust was used to discriminate and annotate meaningful cell clusters, match biologically comparable clusters between samples, and discover which clusters are the best predictors of response to therapy across multiple CIT studies (Greene et al, 2019).…”