Old data, new tricks: Comprehensive computational analysis of 10 years of multi‐center
EuroFlow
acute myeloid leukemia diagnostic immunophenotypic data
Sarah Bonte, Rosan Olsman, Sofie Van Gassen, Sergio Matarraz, Neus Villamor, Stefan Nierkens, Paula Fernandez, Elaine Sobral da Costa, Carmen‐Mariana Aanei, Alberto Orfao, Jacques J. M. van Dongen, Yvan Saeys, Mattias Hofmans, Vincent H. J. van der Velden Abstract
Acute myeloid leukemia (AML) is characterized by high genotypic and immunophenotypic heterogeneity. We collected an extensive dataset containing 5366 flow cytometry files from 885 AML patients, stained with the EuroFlow acute leukemia orientation tube (ALOT) and AML/MDS panel, acquired in a standardized way at eight centers over a period of 10 years. Unsupervised clustering identified groups of patients based on FlowSOM‐derived cell population percentages. In addition, we investigated immunophenotypic patterns in World Health Organization (WHO) patient classes and NPM1 mut subclasses, both at the cell population and the individual marker level. Some WHO classes, for example, AML with t(8;21)(q22;q22)/ RUNX1::RUNX1T1 or t(15;17)(q24;q21)/ PML::RARA , showed homogeneous immunophenotypes. Characterization of maturation arrest using FlowSOM confirmed maturation arrest at early stages in distinct WHO classes. Finally, a machine learning model was trained to predict WHO genetic classes from immunophenotypic data. The model allowed accurate prediction in 77% of cases, reproducible in an independent validation cohort. In conclusion, we show that EuroFlow standardized protocols allow analysis of multi‐centric data measured over an extended period of time. Computational analysis demonstrated inter‐ and intrapatient immunophenotypic heterogeneity and allowed prediction of genetic abnormalities.