Development of a generalizable UMAP-based approach for comparing clinical flow cytometry data with application to NPM1-mutated AML cohorts

Abstract

Introduction AML with mutated NPM1 is associated with heterogeneous clinicopathologic features. We sought to study the association between phenotype, genetics, and clinical behavior using treatment-naïve bone marrow samples of NPM1-mutated AML. Prior phenotypic studies using flow cytometry data have primarily focused on the blast population in isolation or used less comprehensive analysis techniques, such as simple visual histogram assessment. We applied a dimensionality-reduction algorithm (UMAP) to analyze retrospective clinical flow cytometry data of tumor and non-tumor cells in an initial small cohort of NPM1-mutated samples. Methods Custom software was developed using python, FlowKit, and umap-learn to create a dictionary of the various antibody panels, detect the antibody panels that were used in each raw data (FCS) file, and determine the flow cytometer channels that should be disregarded. Subsamples from each FCS file for a given antibody panel were combined and analyzed using UMAP to create an embedding that could then be applied to all FCS files of the given antibody panel. FCS files were subsequently prepared for analysis in FlowJo, including using UMAP coordinates. The initial pilot phase included analysis of the EuroFlow AML1 panel of 11 cases, which included 3 primary refractory cases, 3 early relapsed cases, and 5 cases which achieved clinical remission without relapse. FlowJo was used to gate and examine the clusters identified by UMAP with respect to phenotypic parameters. These same UMAP gates were applied to all 11 cases for direct comparison. Results The blast count ranged from 50 to 88 in these cases. The blast phenotype was determined to be myeloid (n=3), monocytic (n=4), or other (CD34-/HLA-DR-)(n=4). Although standard CD45 by side scatter gating delineates four major cell types (lymphocytes, monocytes, granulocytes, blasts), gating using the UMAP algorithm with input data from eight phenotypic markers in conjunction with scattering parameters, produced at least 10 distinct UMAP clusters of variable cellular composition. Interestingly, applying those gates to side scatter (SSC-A) by CD45 histoplots revealed 5 distinct gates falling into the traditional “blast” region of the histoplot. The UMAP gates thus identified provided quantitative values for further statistical analysis. Conclusion We have developed a useful tool to automatically identify the antibody panels used to generate prior flow cytometry data, preprocess the data, and apply the UMAP algorithm for creating embeddings that can be applied to additional cases. Our preliminary analyses revealed significant phenotypic heterogeneity among a small cohort of NPM1-mutated cases. Ongoing work includes expansion of the cohort and number of antibody panels incorporated into the analyses to elucidate prognostic and predictive features of tumor and non-tumor populations in treatment naïve samples of NPM1-mutated AML.