Automated in-silico detection of cell populations in flow cytometry readouts and its application to leukemia disease monitoring
1 Max Planck Institute for Molecular Genetics & Berlin Center for Genome Based Bioinformatics, Ihnestrasse. 73, D-14195 Berlin, Germany
2 Dept. of Hematology, Oncology and Tumor Immunology, Robert-Roessle-Clinic at the HELIOS Klinikum Berlin, Charité Medical School, Berlin, Germany
3 EMBL – European Bioinformatics Institute, Wellcome Trust Genome Campus, Hinxton, Cambridge CB10 1SD, UK
BMC Bioinformatics 2006, 7:282 doi:10.1186/1471-2105-7-282Published: 5 June 2006
Identification of minor cell populations, e.g. leukemic blasts within blood samples, has become increasingly important in therapeutic disease monitoring. Modern flow cytometers enable researchers to reliably measure six and more variables, describing cellular size, granularity and expression of cell-surface and intracellular proteins, for thousands of cells per second. Currently, analysis of cytometry readouts relies on visual inspection and manual gating of one- or two-dimensional projections of the data. This procedure, however, is labor-intensive and misses potential characteristic patterns in higher dimensions.
Leukemic samples from patients with acute lymphoblastic leukemia at initial diagnosis and during induction therapy have been investigated by 4-color flow cytometry. We have utilized multivariate classification techniques, Support Vector Machines (SVM), to automate leukemic cell detection in cytometry. Classifiers were built on conventionally diagnosed training data. We assessed the detection accuracy on independent test data and analyzed marker expression of incongruently classified cells. SVM classification can recover manually gated leukemic cells with 99.78% sensitivity and 98.87% specificity.
Multivariate classification techniques allow for automating cell population detection in cytometry readouts for diagnostic purposes. They potentially reduce time, costs and arbitrariness associated with these procedures. Due to their multivariate classification rules, they also allow for the reliable detection of small cell populations.