Leukemias - Biology, Cytogenetics and Molecular Markers in Diagnosis and Prognosis Poster I
MRD Monitoring of Childhood ALL Using Hierarchical Clustering and Support Vector Machine Learning of Complex Multi-Parameter Flow Cytometry Data.
Karel Fiser, PhD1,*,
Tomas Sieger, MSc2,*,
Julie Irving, PhD1,
Michael N Dworzak, MD3,* and
Josef Vormoor, MD, FRCPCH1
1 Northern Institute for Cancer Research, Newcastle University, Newcastle upon Tyne, United Kingdom, 2 Praha, Czech Republic, 3 Childrens Cancer Research Institute, Vienna, Austria
Flow cytometry is an important tool both for research and diagnosticsof hematologic malignancies - including monitoring of minimalresidual disease (MRD). Recent progress and more widespreadavailability of 6 and higher color flow cytometry leads to complex,information-rich datasets which are very challenging to analyze.Here, we validate a novel approach to multi-parameter flow dataanalysis in MRD flow data sets from 39 ALL patients (including23 patients from the I-BFM list mode data (LMD) ring trial).The approach combines hierarchical clustering (HCA) using anewly developed algorithm and support vector machine (SVM) learning.The algorithm employs a scale-invariant Mahalanobis distancemeasurement for merging clusters. This reflects the extendedellipsoid shape of the populations and is better suited forflow cytometric data compared with standard HCA metrics. Theresulting hierarchical tree, combined with the heatmap of theCD marker expression allows visualization of hierarchicallyclustered data of all analyzed parameters displayed in a singleplot. The clusters from HCA (representing the ALL blast populationat diagnosis) were used to train SVM classifiers which werethen applied to test for presence of a matching population inthe test sample (follow-up sample). All work was carried outin MATLAB (MathWorks, Inc.). Using HCA, we have been able todetect the leukemic blast population in diagnostic and follow-updatasets (n=81) from three centers. The correlation (Pearsoncorrelation coefficient = 0.98) between HCA and the standardgating approach was highly comparable to inter-laboratory comparisonswithin the I-BFM LMD ring trial (Dworzak MN et al; CytometryB Clin Cytom. 2008 Jun 11.). To further improve sensitivityand exact quantification of low MRD levels and to automate MRDdetection, we combined HCA with SVM learning. We have analyzed21 samples from 5 patients with MRD levels between 0.004 to57.54%. HCA plus SVM correlated better with standard gatingresults than HCA alone in particular in samples with low MRDlevels (<10–3). In summary, HCA in combination withSVM proved to be a strong analytical tool for flow cytometrywith the potential for automated MRD detection. We validatedthis approach for use in ALL diagnostics and MRD monitoringby comparison with expert-based gating analyses of I-BFM LMDring trial data.
Disclosures: No relevant conflicts of interest to declare.