Original Article Computational Analysis Optimizes the Flow Cytometric Evaluation for Lymphoma Fiona E. Craig, 1 * Ryan R. Brinkman, 2,3 Stephen Ten Eyck, 1 and Nima Aghaeepour 2 1 Division of Hematopathology, Department of Pathology, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania 2 Terry Fox Laboratory, British Columbia Cancer Agency, Vancouver, British Columbia, Canada 3 Medical Genetics, University of British Columbia, British Columbia, Canada Background: Although many clinical laboratories are adopting higher color flow cytometric assays, the approach to optimizing panel design and data analysis is often traditional and subjective. In order to address the question “What is the best flow cytometric strategy to reliably distinguish germinal center B-cell lymphoma (GC-L) from germinal center hyperplasia (GC-H)?” we applied a computational tool that identifies target populations correlated with a desired outcome, in this case diagnosis. Design: Cases of GC-H and GC-L evaluated by flow cytometric immunophenotyping using CD45, CD20, kappa, lambda, CD19, CD5, CD10, CD38, were analyzed with flowType and RchyOptimyx to construct cel- lular hierarchies that best distinguished the two diagnostic groups. Results: The population CD52CD191CD101CD382 had the highest predictive power. Manual reanaly- sis confirmed significantly higher CD101/CD382B-cells in GC-L (median 12.44%, range 0.74–63.29, n 5 52) than GC-H (median 0.24%, 0.03–4.49, n 5 48, P 5 0.0001), but was not entirely specific. Diffi- culties encountered using this computational approach included the presence of CD101 granulocytes, continuously variable B-cell expression of CD38, more variable intensity antigen staining in GC-L and inability to assess the contribution of light chain restriction. Conclusion: Computational analysis with construction of cellular hierarchies related to diagnosis helped guide manual analysis of high dimensional flow cytometric data. This approach highlighted the diagnostic utility of CD38 expression in the evaluation of B-cells with a CD101 GC phenotype. In con- trast to computational analysis of non-neoplastic cell populations, evaluation of neoplastic cells must be able to take into consideration increased variability in antigen expression. V C 2013 International Clinical Cytometry Society Key words: flow cytometry; lymphoma; CD38; computational analysis How to cite this article: Craig FE, Brinkman RR, Ten Eyck, Aghaeepour N. Computational Analysis Optimizes the Flow Cytometric Evaluation for Lymphoma. Cytometry Part B 2013;00B:000–000. The distinction between germinal center B-cell lym- phoma (GC-L), such as follicular lymphoma, and germinal center lymphoid hyperplasia (GC-H) can be challenging. Although for some specimens a diagnosis can be made readily using conventional histologic evaluation, for others additional phenotypic or genotypic evaluation is required. Flow cytometric immunophenotyping is a rapid technique that can assist with this distinction, but has some limitations (1). Identification of a population of cells with immunoglobulin light chain restriction by flow cytometry provides strong support for a diagnosis of lym- phoma, but can occasionally be present in florid GC-H (2). In addition, light chain restriction can easily be over- looked if the abnormal population is small, the lym- phoma cells lack surface immunoglobulin expression, or the assay is limited by non-specific staining. Although bcl- 2 protein overexpression can be identified by *Correspondence to: Fiona E. Craig, MD, Professor, Division of Hem- atopathology, Department of Pathology, University of Pittsburgh School of Medicine, 200 Lothrop Street, Suite G300, Pittsburgh, PA 15213- 2582, USA. E-mail: [email protected]Grant sponsor: Canadian Cancer Society; Grant number: 700374; Grant sponsor: Terry Fox Research Institute and Terry Fox Foundation. Received 2 March 2013; Revised 20 May 2013; Accepted 1 July 2013 Published online in Wiley Online Library (wileyonlinelibrary. com). DOI: 10.1002/cyto.b.21115 Cytometry Part B (Clinical Cytometry) 00:00–00 (2013) V C 2013 International Clinical Cytometry Society
Transcript
Original Article
Computational Analysis Optimizes the FlowCytometric Evaluation for Lymphoma
Fiona E. Craig,1* Ryan R. Brinkman,2,3 Stephen Ten Eyck,1 and Nima Aghaeepour2
1Division of Hematopathology, Department of Pathology, University of Pittsburgh School of Medicine,Pittsburgh, Pennsylvania
2Terry Fox Laboratory, British Columbia Cancer Agency, Vancouver, British Columbia, Canada3Medical Genetics, University of British Columbia, British Columbia, Canada
Background: Although many clinical laboratories are adopting higher color flow cytometric assays, theapproach to optimizing panel design and data analysis is often traditional and subjective. In order toaddress the question “What is the best flow cytometric strategy to reliably distinguish germinal centerB-cell lymphoma (GC-L) from germinal center hyperplasia (GC-H)?” we applied a computational tool thatidentifies target populations correlated with a desired outcome, in this case diagnosis.
Design: Cases of GC-H and GC-L evaluated by flow cytometric immunophenotyping using CD45, CD20,kappa, lambda, CD19, CD5, CD10, CD38, were analyzed with flowType and RchyOptimyx to construct cel-lular hierarchies that best distinguished the two diagnostic groups.
Results: The population CD52CD191CD101CD382 had the highest predictive power. Manual reanaly-sis confirmed significantly higher CD101/CD382B-cells in GC-L (median 12.44%, range 0.74–63.29,n 5 52) than GC-H (median 0.24%, 0.03–4.49, n 5 48, P 5 0.0001), but was not entirely specific. Diffi-culties encountered using this computational approach included the presence of CD101 granulocytes,continuously variable B-cell expression of CD38, more variable intensity antigen staining in GC-L andinability to assess the contribution of light chain restriction.
Conclusion: Computational analysis with construction of cellular hierarchies related to diagnosishelped guide manual analysis of high dimensional flow cytometric data. This approach highlighted thediagnostic utility of CD38 expression in the evaluation of B-cells with a CD101 GC phenotype. In con-trast to computational analysis of non-neoplastic cell populations, evaluation of neoplastic cells must beable to take into consideration increased variability in antigen expression. VC 2013 International Clinical
How to cite this article: Craig FE, Brinkman RR, Ten Eyck, Aghaeepour N. Computational Analysis Optimizesthe Flow Cytometric Evaluation for Lymphoma. Cytometry Part B 2013;00B:000–000.
The distinction between germinal center B-cell lym-phoma (GC-L), such as follicular lymphoma, and germinalcenter lymphoid hyperplasia (GC-H) can be challenging.Although for some specimens a diagnosis can be madereadily using conventional histologic evaluation, forothers additional phenotypic or genotypic evaluation isrequired. Flow cytometric immunophenotyping is a rapidtechnique that can assist with this distinction, but hassome limitations (1). Identification of a population ofcells with immunoglobulin light chain restriction by flowcytometry provides strong support for a diagnosis of lym-phoma, but can occasionally be present in florid GC-H(2). In addition, light chain restriction can easily be over-looked if the abnormal population is small, the lym-
phoma cells lack surface immunoglobulin expression, orthe assay is limited by non-specific staining. Although bcl-2 protein overexpression can be identified by
*Correspondence to: Fiona E. Craig, MD, Professor, Division of Hem-atopathology, Department of Pathology, University of Pittsburgh Schoolof Medicine, 200 Lothrop Street, Suite G300, Pittsburgh, PA 15213-2582, USA. E-mail: [email protected] sponsor: Canadian Cancer Society; Grant number: 700374;Grant sponsor: Terry Fox Research Institute and Terry Fox Foundation.
Received 2 March 2013; Revised 20 May 2013; Accepted1 July 2013
Published online in Wiley Online Library(wileyonlinelibrary. com).DOI: 10.1002/cyto.b.21115
Cytometry Part B (Clinical Cytometry) 00:00–00 (2013)
VC 2013 International Clinical Cytometry Society
immunohistochemistry or flow cytometry in many casesof GC-L, some cases are bcl-2 negative (3,4). In addition,bcl-2 staining of GC B-cells can be difficult to distinguishfrom that of T-cells and plasma cells by single colorimmunohistochemistry and although flow cytometricanalysis can isolate the B-cells of interest, it requires cellpermeabilization, is prone to non-specific staining, andcan be difficult to interpret. Abnormal increased ordecreased expression of the surface antigens CD19,CD20, and CD10 can be found in many cases of follicularlymphoma, but evaluation for each antigen in isolationlacks sensitivity for the diagnosis of lymphoma (5,6). Theavailability of higher color flow cytometric assays facili-tates the analysis of multiple antigens simultaneously, butraises questions about which antigens to evaluate andhow best to analyze the data. In order to address thequestion “What is the best flow cytometric analysis strat-egy to distinguish GC-L from reactive lymphoid tissueusing the 8-color antibody combination: anti-kappa, anti-lambda, CD19, CD20, CD10, CD5, CD38, and CD45 anti-bodies?” we applied the computational tools flowTypeand RchyOptimyx (cellular hieraRCHY OPTIMization),which will be briefly summarized in the discussion sec-tion of the manuscript (7,8). This computational analysishighlighted the diagnostic utility of identifying CD10 pos-itive, CD38 negative B-cells in the distinction betweenGC-L and GC-H. In addition, this study revealed some ofthe limiting factors that must be considered when apply-ing computational analysis to clinical data sets.
MATERIALS AND METHODS
Lymphoid tissue biopsy specimens with the followingfeatures were identified from the pathology reports atthe University of Pittsburgh Medical Center (UPMC; Uni-versity of Pittsburgh Institutional Review Board IRB pro-posal PRO11060224):
� Flow cytometric immunophenotyping using a tubecontaining Bcl-2 FITC/CD10 PE/CD20 PerCP-Cy5.5(BD Bioscience, San Jose, CA).
� Presence of a population of CD10 positive germinalcenter B-cells.
� Diagnosis of reactive changes, including hyperplasticfollicles (GC-H), or GC-L, confirmed by review of alldiagnostic materials and using the criteria outlined inthe 2008 WHO classification (9).
Using these criteria, the following specimens wereidentified: GC-H (n 5 48, 25 females, 23 males, medianage 40 years) and GC-L (n 5 52, 29 females, 23 males,median age 70 years). Cases of GC-L included follicularlymphoma, Grade 1–2 (n 5 34) and follicular lymphoma,Grade 3A and/or diffuse large B-cell lymphoma (n 5 18).
Flow cytometric immunophenotyping was performedon cells extracted by manual disaggregation. Viabilitywas determined by Trypan Blue exclusion and rangedfrom 66 to 99% viable (median 84%). A suspension of 53 l05 cells/tube in phosphate buffer saline (PBS) con-taining 0.1% sodium azide and 2% fetal bovine serumwas incubated with the 8-color surface antibody combi-nation for 15–30 min at 4�C. Lysis was performed usingammonium chloride, and followed by washing with PBS.Stained cells were fixed with 2% formaldehyde. For thebcl-2 tube cells were fixed and permeabilized (Fix andPerm Kit, Life Technologies, NY) as previously described(3). Acquisition for both tubes was performed on thesame day as staining using a BD FACS Canto II flowcytometer (BD Bioscience, San Jose, CA), and collectionof 30,000 events. To ensure consistency of results instru-ment setup was standardized using target CST beads(BD Bioscience, San Jose, CA) and voltages were moni-tored with Levey–Jennings plots, settings were clonedbetween instruments, instrument spectral compensationwas set up using Compbeads (BD Biosciences) and lot-to-lot reagent checks were performed.
Data are publicly available through FlowRepository(http://flowrepository.org/id/FR-FCM-ZZ6B).
The original manual analysis was performed in theUPMC clinical flow cytometry laboratory using FACSDIVA software (BD Bioscience, San Jose, CA), with atemplate that includes the following steps: exclusion ofdoublets using a plot of forward light scatter (FSC)-areaversus FSC-height, exclusion of debris using a plot ofCD45 versus side light scatter (SSC) by gating on cellswith low SSC and staining for CD45, identify B-cellsthrough expression of CD19 and/or CD20; identify thefollowing subsets: CD381(bright) plasma cells, CD10positive GC B-cells, CD5 positive B-cells; evaluate eachsubset for immunoglobulin light chain restriction; evalu-ate each subset visually for altered expression of CD19,CD20, or CD10. A separate tube evaluating bcl-2 expres-sion was either ordered up-front or added if the 8-colorB-cell tube identified a CD10 positive germinal center B-cell population, but did not adequately distinguishbetween GC-H and GC-L. The results of this conven-tional flow cytometric analysis for the 8-color B-cell tubeand a separate flow cytometric tube containing Bcl-2FITC/CD10 PE/CD20 PerCP-Cy5.5 were reviewed.
Computational analysis was performed on de-identifiedflow cytometric data as previously described (8). Briefly,logical transformation of immunophenotypic data wasperformed and gates were determined for each marker soas to partition positive and negative cell populations (8).Using this data, all possible phenotypes were extractedusing flowType (7,8). Receiver operating characteristic(ROC) analysis was then performed to identify the pheno-types associated with a statistically significant differencebetween GC-H and GC-L and those with the strongestpredictive power were selected for further analysis. Thephenotypes selected were analyzed using RchyOptimyxto identify their most important parent populations.
Manual re-analysis of the flow cytometric data was thenperformed using FACS DIVA software (BD Bioscience, SanJose, CA) to further explore the cell populations identifiedby computational analysis: CD101, CD101CD38(2),CD191CD101CD38(2), CD5(-)CD191CD101CD38(2);CD101CD38(2), CD101CD38(2)CD45(2), CD191CD101CD38(2)CD45(2); Lambda1CD10(2), Kappa1CD102,Lambda1Kappa1CD10(2), Lambda1Kappa1CD10(2)CD38(2), Lambda1Kappa1CD10(2)CD201. The propor-tion of events and median fluorescence intensity (MFI)was determined for each population identified in order toillustrate features previously reported using current manual
methods and compare those with the computationalresults.
RESULTS
Original Manual Flow Cytometric Analysis
Manual analysis performed in the clinical flow cytome-try laboratory revealed immunoglobulin light chainrestriction, bcl-2 expression, and aberrant weak intensityexpression of CD19, CD20, and/or bright intensityCD10, in the CD10 positive germinal center B-cells pres-ent in some GC-L specimens, but not GC-H (Table 1).
Table 1Diagnostic Utility of Parameters Identified by Conventional Flow Cytometric Analysis
FIG. 1. Computational analysis identifies cellular hierarchies with a high predictive power for distinguishing germinal center hyperplasia and germi-nal center lymphoma. The color of the nodes indicates the predictive power of the phenotype, as measured by AUC (area under ROC curve).
AUTOMATED COMPUTATIONAL ANALYSIS 3
Cytometry Part B: Clinical Cytometry
Automated Computational Flow Cytometric Analysis
A total of 5,660 phenotypes were identified and thosewith the strongest predictive power were selected andarranged into a hierarchy (Fig. 1). The phenotypes mostcapable of discriminating between GC-H and GC-L speci-mens demonstrated a range of AUC, including CD101
(higher proportion in GC-L, AUC 0.985), as demon-strated by the colors in Figure 1.
Manual Re-analysis of Flow Cytometric Data
Refined manual analysis, focusing on the cell popula-tions identified by computational analysis, confirmed thepresence of a higher proportion of CD10 positive cellsin GC-L, but demonstrated significant overlap with theproportion seen in GC-H, limiting the diagnostic utilityof this feature: GC-L median 82.15%, range 14.5–99.3%;GC-H median 8.20%, range 2.1–58.1%). In addition,review of one GC-H case with a large number of CD10positive events (67.2%) revealed that these events repre-sented neutrophils, rather than germinal center B-cells.This case highlights problems that can be encounteredwhen evaluating each marker in isolation, rather than aspart of a multiparameter approach, such as evaluatingCD10 expression on B-cells identified with CD20 and/orCD19. Plots of CD10 versus CD38 also revealed a smallnumber of spurious events, with apparent very brightstaining for CD38 and CD10 in a diagonal pattern,which likely represent aggregates from the CD38 APC-H7 reagent. Although this artifact was easy to recognizeon review of dot-plots, they could potentially misleadthe computational analysis.
A new manual clinical analysis template was createdwith focus on CD101CD38(2) events as follows: sin-glets identified using FSC-area versus FSC-height,SSC(low)CD451(bright) lymphoid cells, CD191CD201
B-cells, CD101 B-cells, CD101CD381 B-cells,CD101CD382 B-cells, Kappa1 and Lambda1
CD10(2), and CD101 B-cells. Re-analysis using this tem-plate confirmed the presence of a significantly higherproportion of CD10 positive, CD38 negative B-cells inGC-L (median 12.44%, range 0.74–63.29, n 5 52) thanGC-H (median 0.24%, 0.03–4.49, n 5 48) (Mann–Whit-ney P 5 0.0001). On visual inspection, a discrete popula-tion of CD10 positive, CD38 negative events was pres-ent in 11/48 (22.9%) GC-H specimens and 39/52 (75%)GC-L specimens. Therefore, although visual identifica-tion of a population of CD10 positive, CD38 negativeevents had a higher sensitivity for the identification ofGC-L than aberrant antigen expression of CD19, CD20,or CD10, it had limited specificity (Table 1 and Fig. 2).However, identification of a discrete population of CD10positive, CD38 negative B-cells in three of five cases ofGC-L that lacked light chain restriction (i.e., were sur-face immunoglobulin negative) could potentially beused to prompt additional investigation. Consistency of
staining was confirmed by lack of a statistically signifi-cant difference between CD38 MFI of CD10 negativegerminal center B-cells in GC-L and GC-H.
Review of manual analysis also highlighted limita-tions in the computational analysis that related to thepartition of positive and negative cell events. Althoughthere were discrete CD5, CD10, CD19, and CD20 posi-tive and negative populations in most specimens, GC-Ldemonstrated more variable intensity of antigenexpression than GC-H (Fig. 3). Indeed, as reported pre-viously, there were significant differences in the MFIfor CD19 (lower in GC-L), CD20 (lower in GC-L),CD10 (higher in GC-L), and CD38 (lower in GC-L)(Table 2 and Fig. 3). Therefore, when a single thresh-old value was applied to partition positive and nega-tive cell populations, some GC-L cells with weak inten-sity expression of CD19 were considered “negative.”However, our computational approach, with a fixedthreshold value, was not an effective way of identify-ing alterations in intensity. In addition, some otherantigens did not demonstrate discrete positive and neg-ative populations. For example, CD45 staining was uni-form and of relatively weak intensity and interpretedas “negative” by computational analysis but was con-sidered “positive” by manual analysis because of theexpected staining of leukocytes. Although this differ-ence in interpretation did not affect the analysisbecause CD45 staining was similar in lymphoma andreactive cases, it could potentially mask detection ofloss of staining in other data sets. In contrast, CD38demonstrated continuously variable intensity of stain-ing. Therefore, the threshold chosen by computationalanalysis divided cells into those with relatively higherand lower intensity staining for CD38, with lowerintensity staining that crossed the threshold being clas-sified as “negative.” Repeat computational analysis uti-lizing the MFI for CD19 and CD38, rather than thepercent of each cell population, did not improve theROC values (data not shown).
Review of the manual analysis for kappa and lambdahighlighted some difficulties encountered by the compu-tational analysis. The computational approach applieddid not effectively assess for immunoglobulin light chainrestriction because it evaluated each phenotype inde-pendently (e.g., kappa1/lambda2, lambda1/kappa2),rather than identifying a dominant population of B-cellsexpressing one light chain, either kappa or lambda,along with lack of another population of B-cells express-ing the other light chain. In addition, there was poorseparation of surface immunoglobulin light chain posi-tive and negative events, in part due to cytophilic anti-body staining of non-B-cells, which likely led to the iden-tification of events apparently positive for both kappaand lambda. Interestingly, the population “kappa1,lambda1, CD102, CD201” was identified as one of thekey discriminators between GC-L and GC-H, andappeared to relate to the higher prevalence of non-germinal center B-cells, without dominant light chainexpression, in GC-H.
4 CRAIG ET AL.
Cytometry Part B: Clinical Cytometry
DISCUSSIONAdvances in instrumentation and reagent technologies
have led to the widespread use of high-level multicolorflow cytometry. However, effective strategies for storing,representing, and interpreting the increasingly complexdata have been lacking. Recognition of this need has ledto the recent development of many automated gatingtools, some of which are now employed in researchapplications and to support high-throughput technology(10–12). There is increasing interest in employing thesetools in clinical flow cytometry laboratories to supportthe use of higher color systems, larger panels ofreagents, and more sophisticated analysis strategies withcomplex hierarchical gating (13,14). Manual analysis offlow cytometric data remains one of the largest variablesin flow cytometric immunophenotyping (15) often relieson personal experience, is time intense, error prone,and difficult to standardize. Automated analysis is poised
not only to remove the burden of manual gating (16),but also to take the next step and identify biologicalchanges associated with disease. However, these toolshave the potential to also assist with manual analysis bymaximizing the information obtained from a single mul-ticolor tube, evaluating the relative importance of infor-mation in reaching a final interpretation, suggesting opti-mized gating strategies and potentially decreasing thesize of marker panels (13,14).
In the current study, we used computational analysisto identify features that can best distinguish germinalcenter lymphoma from reactive germinal center cells.The utility of our computational pipeline has previouslybeen demonstrated using 13-color flow cytometric datafrom T-cell subset evaluation of HIV positive subjects(7,8). This pipeline consists of two independent open-source tools. In this setting, flowType was used to iden-tify surrogate cell surface marker phenotypes that could
FIG. 2. Example phenotypes identified in germinal center hyperplasia and lymphoma. Dark blue events 5 B-cells identified with staining for CD19and CD20. Green events 5 CD10 positive B-cells. Red events 5 CD101CD38(2). GC-H Case 56 (upper row) displays many CD101 B-cells (58.1%total) that demonstrate typical staining for CD38 and polytypic staining for lambda and kappa immunoglobulin light chains. GC-H Case 31 (centerrow) displays small population of CD10 positive events with typical expression of CD19 and CD20, and polytypic staining for lambda and kappaimmunoglobulin light chains, but partial lack of staining for CD38. GC-L Case 77 (lower row) demonstrates a large population of CD10 positive B-cells with decreased intensity staining for CD19 and CD20, lambda light chain restriction and lack of staining for CD38. Note the presence of a fewevents with apparent very bright, diagonal, staining for CD10 and CD38 that likely represent reagent aggregates.
AUTOMATED COMPUTATIONAL ANALYSIS 5
Cytometry Part B: Clinical Cytometry
overcome the need to detect intracellular markers andRchyOptimyx was used to simplify the gating strategyfor the identified phenotypes (e.g., using CD45RO2
CCR52 CCR71 instead of the more complicatedCD281 CD45RO2 CD572 CCR52 CD271 CCR71)and summarize a large list of phenotypes to identifythree strong predictors of progression to AIDS (7). Inthe current study, we utilized RchyOptimyx to objec-tively identify the optimal analysis strategy for the 8-color B-cell directed antibody combination: anti-kappa,anti-lambda, CD19, CD20, CD10, CD5, CD38, and CD45.Application of this computational tool highlighted thediagnostic utility of identifying CD38 negative, CD10positive germinal center B-cells in association with lym-phoma, a feature not emphasized in the original manualanalysis.
CD38 is a glycoprotein that is expressed by precursorB-cells, germinal center B-cells, and plasma cells, but isabsent from na€ıve B-cells and memory B-cells (17–19).Its expression appears to be under tight control, with B-cells undergoing synchronous gain and loss of CD38 andCD10 expression as they enter and exit the germinalcenter (20). Non-neoplastic B-cells with a CD38 nega-tive, CD10 positive phenotype are not well recognizedand therefore, the presence of this phenotype in GC-Lmay reflect aberrant antigen expression. This interpreta-tion is supported by a previous study that reported sig-
nificantly lower CD38 expression by the neoplastic cellsin follicular lymphoma, as determined by MFI, whencompared with reactive germinal center B-cells (21).However, in both the previous and current study, evalua-tion of CD38 expression alone was insufficient to estab-lish a diagnosis of lymphoma (21). Although revision ofthe manual analysis strategy in the current study toemphasize CD38 negative, CD10 positive B-cells identi-fied most of the cases of germinal center lymphoma, asmall population of cells with this phenotype was alsoidentified in some cases of reactive hyperplasia. Theidentification of a small population of CD38 negative,CD10 positive B-cells in reactive lymphoid tissue sug-gests that in GC-L this phenotype may represent expan-sion of this subset rather than aberrant antigen expres-sion. It will be of interest to explore additionalphenotypic findings than can assist in the distinctionbetween normal and neoplastic CD10 positive, CD38negative B-cells, and determine whether non-neoplasticB-cells with this phenotype reflects a transitional stagefrom na€ıve or immature/transitional B-cells to germinalcenter B-cells, or germinal center B-cells to memory B-cells (19).
The main utility of the computational strategy utilizedin this study was the ability to identify phenotypes thatwere associated with an outcome, in this case diagnosis,and thereby assess the relative utility of different anti-bodies and analysis strategies. However, although thecomputational analysis employed in the current studyhighlighted a cell population that had not been empha-sized in the manual analysis, it did not supersede any ofthe existing analysis strategies. One factor that contrib-uted to the decision to retain the other analysis compo-nents is the low tolerance for any false positive or falsenegative results in the clinical diagnostic setting. In addi-tion, some of the diagnostically useful parameters uti-lized in the conventional analysis could not be identifiedwith the computational analysis because of independentconsideration of each marker and use of fixed thresh-olds. For example, the independent partition of eachmarker into positive and negative cell populations led todifficulty in distinguishing CD10 staining of granulocytesfrom that of germinal center B-cells, and could havebeen minimized by initial gating on B-cells or possiblyuse of side light scatter to gate out granulocytes. Simi-larly, this strategy could not distinguish B-cells and non-B-cell cytophilic staining for kappa and lambda, and wasunable to assess for immunoglobulin light chain restric-tion, i.e., more homogeneous staining of B-cells for one
FIG. 3. Variation in germinal center B-cell antigen expression inhyperplasia and lymphoma. MFI indicated for CD10 positive B-cells inGC-H and GC-L.
Table 2Comparison of germinal center B-cell antigen expression in hyperplasia and lymphoma
MFI 5 Median fluorescence intensity of CD10 positive B-cells.
6 CRAIG ET AL.
Cytometry Part B: Clinical Cytometry
light chain only. Other strategies for automated analysisof flow cytometric data, such as vector quantization,dimension reduction, and clustering algorithms have theadvantage of utilizing multiple parameters simultane-ously to identify and evaluate populations of cells, asshown in the referenced examples (22–24). Multipara-metric approaches, such as these, or an ensemble ofcomputational strategies, as highlighted through theFlowCAP challenges, might be more successful at popu-lation identification and characterization, and serve as abetter comparison with manual analysis (16). However,these multivariate approaches are limited in their abilityto incorporate biological guidance for identifying cellpopulations, and often require a complex and subjectivemeta-clustering step for matching high-dimensional cellpopulations across different patients (14,25). Anotherchallenge in the design of automated computationalstrategies is the presence of technical and biologic vari-ability in the data. Although the flow cytometric datautilized in the current study was generated employingclinical laboratory procedures to ensure consistency ofdata, there were some artifacts, such as the presence ofnon-specific antibody staining, doublet formation, andreagent aggregates, which could mislead automatedcomputational analysis. As the development of clinicallyapplicable computational tools progresses, it will beimportant to address these issues using controlled clini-cal data sets with associated outcome data, such as theone we describe here.
LITERATURE CITED
1. Craig FE, Foon KA. Flow cytometric immunophenotyping for hema-tologic neoplasms. Blood 2008;111:341–3967.
2. Kussick SJ, Kalnoski M, Braziel RM, Wood BL. Prominent clonal B-cell populations identified by flow cytometry in histologically reac-tive lymphoid proliferations. Am J Clin Pathol 2004;121:464–472.
3. Cook JR, Craig FE, Swerdlow SH. bcl-2 expression by multicolorflow cytometric analysis assists in the diagnosis of follicular lym-phoma in lymph node and bone marrow. Am J Clin Pathol 2003;119:145–151.
4. Cornfield DB, Mitchell DM, Almasri NM, Anderson JB, Ahrens KP,Dooley EO, Braylan RC. Follicular lymphoma can be distinguishedfrom benign follicular hyperplasia by flow cytometry using simulta-neous staining of cytoplasmic bcl-2 and cell surface CD20. Am JClin Pathol 2000;114:258–263.
5. Ray S, Craig FE, Swerdlow SH. Abnormal patterns of antigenicexpression in follicular lymphoma: A flow cytometric study. Am JClin Pathol 2005;124:576–583.
6. Almasri NM, Iturraspe JA, Braylan RC. CD10 expression in follicularlymphoma and large cell lymphoma is different from that of reactivelymph node follicles. Arch Pathol Lab Med 1998;122:539–544.
7. Aghaeepour N, Chattopadhyay PK, Ganesan A, O’Neill K, Zare H,Jalali A, Hoos HH, Roederer M, Brinkman RR. Early immunologiccorrelates of HIV protection can be identified from computationalanalysis of complex multivariate T-cell flow cytometry assays. Bioin-formatics 2012;28:1009–1016.
8. Aghaeepour N, Jalali A, O’Neill K, Chattopadhyay PK, Roederer M,Hoos HH, Brinkman RR. RchyOptimyx: Cellular hierarchy optimiza-tion for flow cytometry. Cytometry A 2012;81A:1022–1030.
9. Swerdlow SH, International Agency for Research on Cancer, andWorld Health Organization. WHO Classification of Tumours of Hae-matopoietic and Lymphoid Tissues, 4th ed. World Health Organiza-tion Classification of Tumours, 2008. Lyon, France: InternationalAgency for Research on Cancer. 439 p.
10. Bashashati A, Brinkman RR. A survey of flow cytometry data analy-sis methods. Advances in Bioinformatics 2009;584603, 19 p.
11. Maecker HT, McCoy JP, Nussenblatt R. Standardizing immunopheno-typing for the Human Immunology Project. Nat Rev Immunol 2012;12:191–200.
12. Preffer F, Dombkowski D. Advances in complex multiparameterflow cytometry technology: Applications in stem cell research.Cytometry B Clin Cytom 2009;76B:295–314.
13. Costa ES, Arroyo ME, Pedreira CE, Garc�ıa-Marcos MA, TaberneroMD, Almeida J, Orfao A. A new automated flow cytometry data anal-ysis approach for the diagnostic screening of neoplastic B-cell disor-ders in peripheral blood samples with absolute lymphocytosis. Leu-kemia 2006;20:1221–1230.
14. Zare H, Bashashati A, Kridel R, Aghaeepour N, Haffari G, ConnorsJM, Gascoyne RD, Gupta A, Brinkman RR, Weng AP. Automatedanalysis of multidimensional flow cytometry data improves diagnos-tic accuracy between mantle cell lymphoma and small lymphocyticlymphoma. Am J Clin Pathol 2012;137:75–85.
15. Jaimes MC, JMaecker HT, Yan M, Maino VC, Hanley MB, Greer A,Darden JM, D’Souza MP. Quality assurance of intracellular cytokinestaining assays: Analysis of multiple rounds of proficiency testing. JImmunol Methods 2011;363:143–157.
16. Aghaeepour N, Finak G; FlowCAP Consortium; DREAM Consortium,Hoos H, Mosmann TR, Brinkman R, Gottardo R, Scheuermann RH.Critical assessment of automated flow cytometry data analysis tech-niques. Nat Methods 2013;10:228–238.
17. Deaglio S, Malavasi F. The CD38/CD157 mammalian gene family: Anevolutionary paradigm for other leukocyte surface enzymes. Puriner-gic Signal 2006;2:431–441.
18. Deaglio S, Mehta K, Malavasi F. Human CD38: A (r)evolutionarystory of enzymes and receptors. Leuk Res 2001;25:1–12.
19. Perez-Andres M, Paiva B, Nieto WG, Caraux A, Schmitz A, AlmeidaJ, Vogt RF Jr, Marti GE, Rawstron AC, Van Zelm MC, et al. Humanperipheral blood B-cell compartments: A crossroad in B-cell traffic.Cytometry B Clin Cytom 2010;78B (Suppl 1):S47–S60.
20. Ridderstad A, Tarlinton DM. Kinetics of establishing the memory Bcell population as revealed by CD38 expression. J Immunol 1998;160:4688–4695.
21. Mantei K, Wood BL. Flow cytometric evaluation of CD38 expressionassists in distinguishing follicular hyperplasia from follicular lym-phoma. Cytometry B Clin Cytom 2009;76B:315–320.
22. Aghaeepour N, Nikolic R, Hoos HH, Brinkman RR. Rapid cell popu-lation identification in flow cytometry data. Cytometry A 2011;79A:6–13.
23. Pedreira CE, Costa ES, Almeida J, Fernandez C, Quijano S, Flores J,Barrena S, Lecrevisse Q, Van Dongen JJ, Orfao A. A probabilisticapproach for the evaluation of minimal residual disease by multi-parameter flow cytometry in leukemic B-cell chronic lymphoproli-ferative disorders. Cytometry A 2008;73A:1141–1150.
24. Costa ES, Pedreira CE, Barrena S, Lecrevisse Q, Flores J, Quijano S,Almeida J, del Carmen Garc�ıa-Macias M, Bottcher S, Van Dongen JJ,Orfao A. Automated pattern-guided principal component analysis vsexpert-based immunophenotypic classification of B-cell chronic lym-phoproliferative disorders: A step forward in the standardization ofclinical immunophenotyping. Leukemia 2010;24:1927–1933.
25. Azad A, Pothen A. Multithreaded algorithms for matching in graphswith application to data analysis in flow cytometry. In: Parallel andDistributed Processing Symposium Workshops & PhD Forum(IPDPSW); Shanghai, China, 2012 IEEE 26th International, 2012.
