The peripheral blood B-cell compartment is a dynamic complex of B-cell subsets that undergo maturational steps, antigen exposure, memory cell formation, memory maintenance, and effector functions (antibody secretion). Polychromatic flow cytometry offers an extraordinary tool to investigate such a complex set of cell types. It allows for the definition and proportional assessment of B-cell subsets and also subdivides into less well-characterized intermediates. However, the analysis of complex data obtained from the six-color space is difficult. The standard analysis strategy uses the creation of series of two-dimensional gates, which makes the resulting analysis subjective, and “a priori” unknown small subsets are difficult to discover. Analysis of the B-cell compartment of common variable immunodeficiency (CVID) patients represents a challenging example of these difficulties.
CVID is a heterogeneous disorder of unknown etiology. The hallmark of the disease is a humoral deficiency characterized by low levels of serum immunoglobulin G (IgG), immunoglobulin A (IgA), and/or immunoglobulin M (IgM) and impaired specific antibody response after antigen challenge (1). Investigation of the B-cell compartment in the peripheral blood reveals heterogeneous and inconstant abnormities, such as, missing plasma cells, reduced switched memory B-cells, increased numbers of CD21low cells, etc. (2). Several research groups have found correlations between B-cell compartment abnormities and clinical presentations (1, 3, 4); however, the causal relationship between the defective B-cell function and the clinical picture is still elusive.
To test a hypothesis that different defects in the B-cell development underlie the similarities of the clinical or functional phenotypes of CVID patients, a reproducible method describing the patient's B-cell profile should be created. In concordance with expression profiles, we used computational tools to analyze the B-cell profiles of a cohort of CVID patients and healthy individuals. Our study shows that such an approach offers a tool for the comparison of B-cell profiles in an expert independent manner, thus not only preventing the subjectivity of the gating strategy but also allowing for complex data-mining from the six-dimensional space. The differences between groups can be tracked to the already defined or novel B-cell subsets.
Patients and Healthy Donors
We have obtained peripheral blood samples after informed consent from 48 CVID patients (28 females; 20 males) from two immunodeficiency centers and from 49 healthy donors (26 females; 23 males). The two cohorts were of comparable age (median CVID patients 43 years (range 11–73), healthy donors 39 years (range 18–74), no difference by Mann-Whitney test, P = 0.45). The patients fulfilled the diagnostic criteria for CVID defined earlier (4), and secondary immunodeficiency was excluded. All CVID patients were classified according to the EUROclass classification (5). Briefly, EUROclass distinguishes CVID patients based on B-cell phenotype. Group B− separating patients with ≤1% B cells and Group B+ have >1% B cells. Group B+ again divided patients into Group smB− with 2% switched memory B cells and fewer and Group smB+ with >2% switched memory B cells. Patients with CD21low B cells expansion above 10% within B cells are marked as CD21low, whereas patients with less then 10% are marked as CD21norm. Group B− patients were not examined in this study.
B-Cell Staining for Flow Cytometry
Peripheral blood mononuclear cells (PBMC) were isolated from 10 mL of heparinized blood by density gradient centrifugation Ficoll-Paque (Pharmacia, Uppsala, Sweden) and washed twice with phosphate-buffered saline supplemented with 1% bovine serum albumin. PBMCs were incubated for 15 min in the dark with anti-CD27 Pacific Blue, anti-CD38 Alexa Fluor 700 (Exbio Praha a.s., Prague, Czech Republic), anti-human IgM FITC, anti-CD21 APC (BD Pharmingen, San Jose, CA), anti-CD24 PE, and anti-CD19 PC7 (Immunotech, Marseille, France). After a final wash with PBS, two-million PBMCs from each sample were acquired in FCS 3.0 format using a Cyan ADP flow cytometer (Dako, Glostrup, Denmark) equipped with a 100 mW violet (405 nm) laser, a 25 mW blue (488 nm) laser, and a 60 mW red (642 nm) laser. Optical configuration is shown in Supporting Information 1.
Long-Term Instrument's Stability of Fluorescence Measurements
For polychromatic flow cytometry experiments, photomultiplier voltage was set above the electronic noise threshold (6) and was fixed at a target median fluorescence intensity (MFI) of the brightest peak of Rainbow 8-peak beads (Spherotech, Lake Forest, IL). MFI values were controlled daily and kept within the range of +/− 15% of the pre-set target values. Whenever the target values were missed, an alignment was performed and target values were restored. The automated compensation matrix calculation was performed using single-color stained tubes (SUMMIT 4.3) once a month. In addition, compensation was performed after each alignment or after a service visit.
Turning the Six-Color Flow Cytometry Measurement into Dataset Using Probability Binning Algorithm
The probability binning algorithm was extensively described by Roederer et al. (7). The aim of our study was to obtain a B-cell profile using the entire six-color space. Coordinates of all CD19+ lymphocytes were exported as a “scale value” text-file using FlowJo 8.8.4 software (Tree Star, Ashland, OR). We have implemented the probability binning algorithm in Fortran (Fortran 95; GNU Fortran 95 Compiler 4.3). To create a reference matrix, only healthy donors' samples were used. In short, 20,000 B-cells from each donor were combined together, 1,024 bins of an equal size of 957 cells were generated by successive division. The algorithm begins by calculating the variance of all the data for each of the parameters included in the file. It first chooses the parameter with the largest variance, and divides the events in symmetrical halves (bins) based on the median value of that parameter and is repeatedly applied until the desired number of bins is created (7). When 1,024 bins were created, we ran all individual samples through the same network of bins in six-parameter space and obtained a table with each patient in one column and each bin in a row containing a number of cells in each respective bin.
For each bin we take expected number of cells (E),
E = (total number of cells in sample)/1,024 (throughout the samples in the experiment E ≈ 20).
Taking the measured number of cells in the bin (M), we assign to each bin a value (F),
F = M/E− 1.
Thus, if the number of cells in the bin was equal to the reference dataset of healthy controls, the value F was zero, if the bin was “over-fed” with more cells, F > 0, and “under-fed” bins were assigned values F < 0. We considered this data set to represent a B-cell profile and we further analyzed it by applying approaches used in expression profiling.
This approach enables six-color flow cytometry data to be evaluated without any bias introduced by subjective choice of a position and sequence of bi-dimensional gates made by an expert. Moreover, it takes into account the complete cellular information contained in the six-color space.
B-Cell Profile Data Analysis and Hierarchical Clustering
The dataset was ordered as 97 columns (one column per individual) and 1,024 rows (one row per bin). The dataset was analyzed by tools developed for the DNA microarray analysis. Hierarchical clustering was performed in the Bioconductor package of the R-project (Available at: http://www.r-project.org/). The Dynamic Tree Cut (dynamicTreeCut R-package) (8) algorithm was used to define clusters. The clusters were validated using the internal and the stability measures designed in clValid R-package (9). The nonrandomness of patient distribution among clusters was tested by Fisher's Exact Test with simulated P-values (107 replicates).
Evaluation and Visualization of Significantly Different Cell Subsets Between Clusters
Selected clusters were compared with each other to find significantly different bins in a search for cellular subsets that would serve as hallmark of the cluster. Multiple testing procedure (MTP) (10) was used in the search for differently “fed” bins, which is designed for testing a hypothesis with unknown dependence structures among variables.
MTP was set to find significantly different bins at the level of P < 0.01.
An output table with significant bins represented by artificial events (i.e., each significant bin was filled by 100 random events generated from uniform distribution on the bin) was analyzed in MeV-TM4 software (11) by principal component analysis (PCA) (12) to reveal the spatial relationship of “over-fed” and “under-fed” bins. Thus, for each cluster comparison, we generated a PCA with a graphical representation and eigenvectors describing the values of each component. Eigenvectors were used to construct a polyvariate plot (“PV plot” in FlowJo) on the original CD19+ gated listmode data of patients and healthy donors. Over- or under-fed spaces displayed by PCA were gated on the PV plot.
Clusters of Patients
The B-cell profile of the whole cohort of CVID patients and healthy controls was analyzed by hierarchical clustering using a Pearson correlation. Clusters were defined by dynamic branch cutting method implemented in R-package (8). We allowed only groups of five or more patients to be considered as a cluster. The hierarchical clustering algorithm on this dataset was validated by clValid R-package (9). This approach uses all information contained in all six parameters from 1,024 bins to create a similarity matrix. The cluster dendrogram is shown in Figure 1a, with the 12 clusters depicted in red.
The objective computational method was able to cluster together healthy controls (in clusters 3, 5, 6, and 10–12). Furthermore, the composition of clusters in relation to the EUROClass CVID classification (5) has a nonrandom pattern (Fisher's Exact Test P < 10−6; Fig. 1b), where the most of the switched memory B-cell lacking (smB−CD21norm) patients are within clusters 7 and 2, smB+CD21low patients are in clusters 8 and 9 and most of the smB−CD21low patients are in cluster 4. Surprisingly, although patients with switched memory B cells are scattered in many clusters, the smB+CD21low and smB+CD21norm patients never locate to a common cluster. SmB+CD21norm patients frequently locate with healthy control subjects, which is in line with the preservation of some B-cell function in this patient group.
Replicate Samples from Patients are Properly Assigned as the Closest Match
To test the reliability of this method, we obtained samples from 21 donors repeatedly 98–336 days apart. This approach tests the whole procedure including sampling, antibody staining, acquisition by flow cytometry, and the data binning, and similarity matrix generation. When we matched the B-cell profile of replicates to the original cohort using the similarity matrix of the Pearson correlation, 15 of replicates matched to the same individual, three replicates matched to a different individual within the same cluster, and three replicates matched to a different cluster (Table 1).
Table 1. Replicates of blood drawn from 17 patients and four healthy controls were obtained to test the reproducibility of the B-cell profile (Pearson correlation)
ORIGINAL SAMPLE CODE
TIME TO REPLICATE (DAYS)
The closest match for the replicate in the original cohort is listed in the second column. Time (in days) from the original draw to the replicate draw is shown.
Match to the replicate
Match within cluster
Match to different cluster
Thus, we conclude that the procedure is highly accurate and reproducible when comparing cellular profiles in complex six-parameter flow cytometry data.
B-Cell Phenotype Stability
The replicate sample analysis also revealed unprecedented stability of the B-cell profile among CVID patients. Although the time between replicates was longer for the CVID patients than for the healthy controls (mean 249 vs. 126 days), 14 of 17 of the CVID patients' replicates (82%) matched to their original samples, whereas only one of four healthy controls matched exactly (25%) Table 1. This observation suggests a tight regulation of the abnormal B-cell compartment in CVID patients.
Principal Component Analysis Reveals Significantly Different Cell Subsets
To reduce the data back to a level where significantly different B-cell subsets could be visualized by the common flow cytometry software tools, we set out to find significant differences in the cell composition between the CVID clusters and healthy controls. Thus, we tested clusters 2b, 4, 7, 8, and 9 against controls (controls were all healthy donors from groups 3, 5, 6, and 10). Using the MTP procedure, we selected differentially filled bins. Bins were annotated as “over-fed” or “under-fed” as compared with controls. We received a list of 15–311 different bins for each group. Because it is impossible to plot six parameters in two-dimensional plots, the coordinates of the “over-fed” or “under-fed” bins' means were subjected to PCA to find the distinct portions of the six-parameter space where the “over-” or “under-fed” bins reside. PCA was able to display distinct spaces of “over-” or “under-fed” bins (e.g., Cluster 7 vs. controls on Fig. 2). Moreover, PCA computed the best resolving composition of components values (eigenvectors), which could be used to create a “PV plot” in FlowJo software. However, when we displayed differentially filled bins for cluster 12 versus controls, the PCA did not allow spatial separation of the “over-fed” and “under-fed” bins (Supporting Information 2).
Differentially Represented Cell Subsets can be Visualized and Verified in the Original Listmode Data
Using the same eigenvectors, a PV plot was created in the FlowJo software using the original listmode data of donors in cluster 7 and healthy controls (Fig. 3a). The PV plot enabled a display of the major differences between the two clusters, similar to the PCA of the clusters' differences, but also allowed for further analysis of each individual CVID patient back in the listmode file. When we gated on the prominent PCA (7) over-fed region, we could track the significantly different over-represented cells as transitional B-cells, Figure 3b. The percentage of the PV plot gated cells (cluster 7 med = 21.6%, SD = 8.6%; healthy med = 2.35%, SD = 2.6) and classically gated transitional B-cells (cluster 7 med = 13.4%, SD = 7%; healthy med = 1.5%, SD = 2.3) was higher in cluster 7 patients than in the healthy donors (P < 0.01 for both comparisons). Classical gating is shown in Figure 3c.
Cellular subsets that were hallmarks of the defined clusters were tracked back. However, the majority of them did not correspond with the B-cell subsets described so far (13–17), (Table 2, Supporting Information 3).2
Table 2. Location of different bins tracked to the six-dimensional space by PCA
Cluster number is shown in parenthesis. “over” and “under” denotes over- and under-fed bins. Symbols “−” or “+” or “++” stand for the level of expression on the subset (negative or dim or bright). Expression range “−” or “+” means the bin spanned across from negative to positive cells. Only the PCA(7)-over gate was clearly overlapping with the known subset of transitional B-cells. Compare with Figure 3b and Supporting Information 3.
+ OR ++
+ OR ++
− OR +
− OR +
− OR +
− OR +
− OR +
− OR + OR ++
− OR +
− OR +
− OR +
− OR +
− OR +
This is, to our knowledge, the first attempt to use unsupervised computational methods to define clusters of individuals with similar immunophenotype cellular profiles and to apply the methodology for the analysis of the B-cell compartment of CVID patients. Furthermore, we were able to prove the reproducibility of this approach by analysis of biological replicates. The acquired number of B-cells in a sample (20,000) related to the number of bins (1,024) created was sufficient for the description of the overall B-cell profile, and also for the definition of small hallmark subsets that comprised only 1.5% of the healthy control B-cells (transitional B-cells). While we have evaluated the usefulness of the matrix composed of 64–4,096 bins, clustering according to the profile was very similar between 256 and 4,096 bins (data not shown). We decided to use 1,024 bins to ensure the recognition of small subsets (the number of bins was such that one bin contained 0.1% of the B-cells). A crucial precondition of the described approach is the long-term stability of the fluorescence measurements. Typically, defocusing of the alignment of the lasers was the major cause of our instrument stability failure, which occurred about once a month. Careful instrument stability monitoring ensured early detection of technical problems and setting the target MFI on the brightest peak of Rainbow beads was instrumental to restore the fluorescence measurements to the standardized level. Whether or not such an approach can be adopted for a multicenter study remains to be evaluated. We can foresee that some adaptations of this process might be necessary; for example, the use of a percentile scale for bin definition would reflect distribution of a signal measured on an individual instrument.
The amount and quality of information gained by this approach is critically dependent on the composition of the measured markers and also on a careful fluorochrome choice that ensures sufficient sensitivity for each marker. Selected markers were tested to be able to recognize important known B-cell subsets before sample acquisition in this study. Using the six-color panel, we were able to define: transitional B-cells as CD19+CD27−CD24++CD38++ (14), follicular B cells as CD19+CD38+CD24+CD21+CD27−IgM+ (16), marginal zone-like B-cells as CD19+CD38−CD24++CD21int CD27+IgM++ (17), class switched memory B cells as CD19+CD38+CD24++CD21+CD27+IgM− (15), and plasmablasts as CD19+CD38+++CD24−CD27++ (13).
A factor limiting the number of markers used in this study was the availability of key B-cell markers conjugated to less common fluorochromes (Pacific Orange, PE-TxRed, PerCP, and APC-H7) that would be expressed strongly enough for a meaningful analysis.
Unsupervised hierarchical clustering analysis enabled the definition of clusters of individuals with similar B-cell profiles. We chose the Pearson correlation as a distance metric to recognize trends (correlations) in the individual profiles. On the contrary, Euclidian distance evaluates the absolute value of the distance without regard to the direction of change from normality. Both methods yielded meaningful results, but Euclidian distance was better in organizing the healthy controls together and separating healthy controls from CVID patients. The Pearson correlation enhanced formation of CVID patient clusters. The differences are logical as we have chosen the value F = 0 for the “average” or “normal” profile. Thus, healthy controls have low absolute distance from each other, but can be negatively correlated.
The described approach allows for a search of common features of each cluster. Although the cohort of patients is too small to reveal common blocks in B-cell maturation in such a highly heterogeneous syndrome, we were able to prove the principle and show the capacity for its future use. We have shown that there is an incomplete overlap between the EUROClass trial classification and unsupervised clustering. When we reduce the information again in search of a cluster's hallmarks, we can find well defined subsets (e.g., “over-fed” transitional B-cell subset in cluster 7) or less well described subsets (see Table 2, Supporting Information 3). This opens a way for functional characterization of the revealed subsets and the determination of their role in the CVID pathogenesis. A larger study of B-cell profile, T-cell profile, and clinical course of CVID patients is underway.
The major contribution of this study is that it exploits advanced technology (polychromatic flow cytometry) and combines it with new approaches in high-throughput data analysis (well-known in the field of expression profile microarray analysis) to analyze a syndrome that goes unresolved despite several decades of research effort.
There have been some pilot studies of novel approaches in data analysis going in a similar direction; notably, Pedreira et al. (18) used computational approaches to define minimal residual disease in B-cell chronic lymphoproliferative disorders and Steinbrich-Zollner et al. (19) classified the pattern of inflammatory diseases. However, while Pedreira et al. use the whole profile for PCA analysis, they do not use polychromatic data and they do not aim to form clusters of similar profiles. Steinbrich-Zollner et al. perform clustering the results of analysis but necessitate manual expert gating of the listmode data.
Although expression profiling (also known as DNA microarrays) provide information about tens of thousands genes, they suffer from problems arising from the mixed cell populations they analyze. Thus, small cell subsets, if not purified, will not change the expression profile and will go unnoticed. Strengths of the presented “cellomics” approach lay in the analysis of single cells and their surface protein expression, the ease of use, and the rather wide availability of the 3-laser flow cytometry instrumentation. The pitfalls are in the proper choice of defining antigens. The presented proof of principle study sets the stage for more focused, in depth studies of B-cell functional subsets in CVID patients and is applicable to many other research questions in hematopoietic cell development.
The authors thank Daniel Thűrner and Ladislav Król for superior technical assistance.