Consistent quantitative gene product expression: #1. Automated identification of regenerating bone marrow cell populations using support vector machines

Abstract

Identification and quantification of maturing hematopoietic cell populations in flow cytometry data sets is a complex and sometimes irreproducible step in data analysis. Supervised machine learning algorithms present promise to automatically classify cells into populations, reducing subjective bias in data analysis. We describe the use of support vector machines (SVMs), a supervised algorithm, to reproducibly identify two distinctly different populations of normal hematopoietic cells, mature lymphocytes and uncommitted progenitor cells, in the challenging setting of pediatric bone marrow specimens obtained 1 month after chemotherapy. Four‐color flow cytometry data were collected on a FACS Calibur for 77 randomly selected postchemotherapy pediatric patients enrolled on the Children's Oncology Group clinical trial AAML1031. These patients demonstrated no evidence of detectable residual disease and were divided into training (n = 27) and testing (n = 50) cohorts. SVMs were trained to identify mature lymphocytes and uncommitted progenitor cells in the training cohort before independent evaluation of prediction efficiency in the testing cohort. Both SVMs demonstrated high predictive performance (lymphocyte SVM: sensitivity >0.99, specificity >0.99; uncommitted progenitor cell SVM: sensitivity = 0.94, specificity >0.99) and closely mirrored manual cell classifications by two expert‐analysts. SVMs present an efficient, automated methodology for identifying normal cell populations even in stressed bone marrows, replicating the performance of an expert while reducing the intrinsic bias of gating procedures between multiple analysts. © 2016 The Authors. Cytometry Part A published by Wiley Periodicals, Inc. on behalf of ISAC.

Consistent quantitative gene product expression: #3. Invariance with age

“…The flow cytometers were cross standardized and calibrated using RCP‐30A and RFP‐30A beads (Spherotech, Lake Forest, IL) with spectral compensation performed using peripheral blood cells labeled with CD4 (SK3, BD) conjugated to fluorescein (FITC), phycoerythrin (PE), peridinin chlorophyll protein (PerCP) or allophycocyanin (APC). Eight combinations of antibodies were used as previously described .…”

Consistent quantitative gene product expression: #2. Antigen intensities on bone marrow cells are invariant between individuals

“…All SVMs demonstrated strong performance in replicating the gate of an expert while eliminating subjective analytical bias (Supporting Information Table 1). The performance of the SVMs to replicate the expert is the basis of a previous companion article (11).…”

“…This unique cell cluster includes hematopoietic stem cells (HSC) and multipotent progenitor cells, but does not include lineage specific colony forming cells (12)(13)(14). The uncommitted progenitor cells are identified by high expression of CD34 and co-expression of CD33 (11,15). The third reference population, promyelocytes (blue) (Figs.…”

Holography, machine learning, and cancer cells

“…For example, a supervised support vector machine (SVM) algorithm was used to distinguish uncommitted progenitor cells from other cell types in bone marrow aspirates from pediatric acute myeloid leukemia patients, which performed favorably compared to manual classification by an expert (10). For example, a supervised support vector machine (SVM) algorithm was used to distinguish uncommitted progenitor cells from other cell types in bone marrow aspirates from pediatric acute myeloid leukemia patients, which performed favorably compared to manual classification by an expert (10).…”

“…Machine learning techniques have automated cell classification from conventional flow cytometry and microscopy datasets in challenging situations. For example, a supervised support vector machine (SVM) algorithm was used to distinguish uncommitted progenitor cells from other cell types in bone marrow aspirates from pediatric acute myeloid leukemia patients, which performed favorably compared to manual classification by an expert (10). Similarly, an SVM was developed and trained to detect and segment pancreatic islets from microscopy images of hematoxylin and eosin-stained histopathology slides, based on color texture features extracted from clustered regions of contiguous pixels (11).…”