To determine the CD23 cell surface marker’s expression patterns, sections of human tonsil tissue were stained with several combinations of antibodies in immunohistofluorescence experiments. In GCs, CD23 staining was restricted to the LZ. In line with the literature data (22, 23), CD21L⁺ FDCs were intensely stained ( Supplemental Figure S1A ). Using antibodies against CD23, PAX5 and PD1, we possibly distinguished CD23⁺ LZ B_GC-cells located near PD1⁺ Tfh but without certainty due to the FDC labeling ( Figure 1A and Supplemental Figure S1B ). Indeed, CD23 staining of frozen tissue sections showed that the FDCs’ extensions formed a dense mesh around B and T cells ( Figure 1B and Supplemental Figure S1C ). Overall, our data show that FDCs in LZ are strongly positive for CD23, which makes it harder to detect CD23⁺ B_GC-cells, in contrast to NBCs in the mantle zone. ( Supplemental Figure S1A ).

To determine the proportion of CD23⁺ B cells in GCs, we used flow cytometry to analyze cell suspensions obtained from tonsils and reactive lymph nodes (rLNs). Unlike NBCs and IgD⁺ B_GC-cells, IgD^- B_GC-cells, PBs and MBCs were predominantly CD23-negative ( Figure 1C and Supplemental Figure S1D ). Centrocytes are defined as CXCR4^- B_GC-cells; they belong to the LZ compartment where B cells were also described as CD83⁺ (1, 17). The mean ± standard deviation of CD23-expressing LZ B_GC-cells was 6.99% ± 1.45 in tonsils and 7.7% ± 3.99 in rLNs ( Figure 1D and Supplemental Figure S1E ). In B cells, CD23 expression is induced by Tfh-derived cytokines in general and by IL-4/STAT6 signaling in particular. CD23⁺ and CD23^- LZ B_GC-cells expressed similar levels of IL-21 and IL-4 receptors, CD40 and Ki-67, however, we found a significantly higher proportion of CCR6⁺ cells in the CD23⁺ subset consistent with Duan et al. (24) in mice. These cells could correspond to memory B cell precursors ( Figure 1E and Supplemental Figure S1F ) (25).

Given that CD23 expression is signaling-dependent, we hypothesized that CD23⁺ LZ B_GC-cells correspond to B_GC-cells that had recently engaged in a productive synapse with cognate Tfh and might therefore harbor a more restricted BCR repertoire than their CD23^- counterparts. We assessed common CDR3 clusters and variable (V) gene usage for both IgM and IgG BCRs in CD23⁺ and CD23^- LZ BGC-cells isolated from rLNs of three different subjects. We found common CDR3 clusters in all three subjects, with no significant differences in V gene usages between IgMs and IgGs. Interestingly, for one subject we noticed and enrichment for somatic mutations in CDR1, 2 & 3 in CD23⁺ cells compared to CD23^- counterparts. No definitive conclusion can be drawn, however, these results showed that CD23⁺ and CD23^- LZ B_GC-cells have very similar repertoires and so probably share a common initial activation pathway ( Supplemental Figure S1G ).

In humans, the CD23 protein is encoded by the FCER2 gene. To detect possible transcriptional differences between LZ B_GC-cells as a function of FCER2 expression, we compared previously published single-cell (sc) RNA-seq data obtained from tonsil-derived B_GC-cells (26). A comparison of 8,465 DZ cells and 11,118 LZ B_GC-cells revealed a highly significant difference between the proportions of FCER2-expressing cells (2.9% and 15.3%, respectively) ( Supplemental Figure S2A , Left panel). Among the 11,118 LZ B_GC-cells, 2,360 (21%) had a proliferative signature (i.e., S-G2-M genes expression) and among them, only 231 (9.8%) cells (primarily in the S phase cluster) expressed FCER2 ( Supplemental Figure S2A , Right panel). To investigate the heterogeneity of FCER2 expression in the LZ compartment and follow the same rationale as Holmes et al. (26), we focused our analysis on 8,758 nonproliferating LZ B_GC-cells (i.e., those in the G0-G1 stage of the cell cycle) which presented 1,465 (16.8%) cells expressing FCER2 ( Figures 2A, B ). The uniform manifold approximation and projection (UMAP) representation of the original 12 specific clusters allowed to detect the projection of FCER2-positive cells and revealed that this gene expression was significantly associated with cell activation clusters and two BCR engagement clusters. In contrast, five clusters presented significantly weaker expression; they included the PB signature, two clusters associated with B_GC-cell transition between the DZ and the LZ, and one Ribosome cluster ( Figure 2B ).

Taken as a whole, these data showed that the majority of FCER2⁺ cells belong to a small number of activated, non-proliferating LZ B_GC-cells that might have been diverted from a PC fate. In contrast, FCER2-negative LZ B_GC-cells were more heterogeneous in their distribution and presented a relative prevalence in the PB subset by taking into account the significant low number of FCER2-positive cells in this compartment. In mice, it has been suggested that most pre-plasmablasts are derived from positively selected cells i.e., expressing MYC gene (13, 20), we found among the 1,518 human MYC-positive LZ B_GC-cells from Holmes et al. (26) dataset that 316 (21%) and 1,202 (79%) were, respectively, positive and negative for FCER2 expression. This result was consistent with data obtained recently in mice (24). Altogether, these data suggest that FCER2-negative LZ B_GC-cells were either unable to express FCER2 or had not received sufficient Tfh cell help to express it. In particular, FCER2-negative LZ B_GC-cells might include activated B cells that have been committed to PB differentiation and therefore have switched off the IL-4/STAT6 pathway, thereby preventing the expression of FCER2/CD23 (19).

In order to complete this in silico analysis we explored the heterogeneity of the LZ B_GC-cell compartment by comparing CD23-negative and CD23-positive cells freshly sorted from tonsils, and CD19⁺IgD^-CD38^bright PBs were included in the analysis (gating strategy, in Supplemental Figure 2B ). We applied a sensitive sc-qRT-PCR approach to profile the expression of a set of selected genes from our previous study ( Supplemental Material, Method & Tables, Table 2 ) (19). An unsupervised clustering analysis revealed three gene clusters linked to specific cell functions identified with GeneMANIA (https://genemania.org/): cell cycle regulation, B cell activation and PC signature ( Figure 2C ). Only a few genes in each cluster were significantly differentially expressed between CD23⁺ and CD23^- cells, including FCER2 and B cell activation markers CD86 and NFKB1 ( Supplemental, Material, Method & Tables, Table 6 ). Note that PRDM1 and IRF4 genes linked to PC differentiation were not differentially expressed ( Figure 2D ). However, CD23^- compared to CD23⁺ LZ B_GC-cells expressed higher levels of cell proliferation E2F1 and MKI67 genes ( Figure 2D ). The cell proliferation signature was primarily attributed to a group of CD23^- LZ B_GC-cells that also expressed MYC ( Figure 2C ). To complete our exploration on a higher number of analyzed cells, differentiation-associated transcription factors assessed by flow cytometry showed a significant enrichment in BCL6-negative and Blimp1-positive cells in CD23^- LZ B_GC-cells. In contrast, both subsets expressed similar levels of c-MYC and phospho-p70 S6 kinase (pS6), two markers related to B-cell selection ( Figure 2E ) (7, 8, 27). Altogether, these findings suggest that the CD23^- population comprises B cells committed to PC differentiation.

To study the effect of Tfh cell help on the membrane expression of CD23 on LZ B_GC-cells, we first cultured LZ B_GC-cells for 12 h with IL-4, CD40, IL-21 or combinations of these. IL-4 alone, CD40 alone and especially a combination of IL-4 and CD40 were effective in inducing the significant upregulation of CD23 expression ( Figure 3A ). Despite similar expression levels of IL-4R, CD40 and IL-21R on post-stimulated CD23⁺ and CD23^- LZ B_GC-cells, pSTAT6 was only induced in CD23⁺ B cells, and chemical inhibition of pSTAT6 blocked the CD23 expression ( Figure 3B and Supplemental Figures S3A, B ). Since the production of IL-21 and IL-4 by Tfh cells predominantly supports B_GC-cells (along with CD40L) (28, 29), we used these three stimuli in our subsequent experiments on primary LZ B_GC-cells. The data were reproducible, up to 50% of the cells were CD23⁺ after 12 h of culture, cell viability was over 70% (data not shown) and cell proliferation was similar in CD23⁺ and CD23^- cells ( Supplemental Figure S3C ). We next co-cultured LZ B_GC-cells with paired autologous CD4⁺CXCR5⁺ICOS⁺PD1⁺ Tfh cells for 24 h and then analyzed CD23 expression in three independent experiments. Only a weak increase in CD23 expression was observed. In contrast, when Tfh cells were activated either with α-CDA/α-CD28 antibodies or staphylococcal enterotoxin B protein, the CD23 expression increased markedly; 13.6+/-7.4% and 39.4+/-9.0% of the B cells were CD23⁺, respectively ( Figure 3C ).

Taken as a whole, our present data suggest that the low number of CD23⁺ LZ B_GC-cells is linked to a lack of effective, complete Tfh cell help in GCs - a key limiting resource for which the B cells compete (29). After early B cell-Tfh cell contact and in the absence of sustained Tfh support, CD23⁺ B_GC-cells might also undergo apoptosis; this has been described in vivo for B_GC-cells after 24 h of antigen recognition and in vitro for CD23⁺ post-activated B cells (19, 30). We confirmed the latter observation and showed that only CD23⁺ LZ B_GC-cells were more numerous when the pan-caspase inhibitor QVD-OPH was added to the culture for 12 h ( Supplemental Figure S3D ).

Since IL-4 signaling goes through the STAT6 pathway, we next determined the number of B cells receiving Tfh-derived IL-4 in a contact dependent manner at a given point in time. To this end, we looked for nuclear expression of pSTAT6 in CD20⁺ B_GC-cells by staining tonsil tissue sections. Tfh cells were detected by staining for PD1. By using a machine learning approach to automatically detect cells on microscopy images, we estimated that approximately 2 to 3% of B_GC-cells were pSTAT6⁺CD20⁺ ( Figure 3D and Supplemental Figure S3E ). By replacing CD20 with PAX5, we could estimate the number of pSTAT6⁺ cells in the vicinity of PD1⁺ T cells for three tonsils. Overall, among the average of 2880 PAX5⁺/pSTAT6⁺ B cells per tonsil, we found 18 (min 13- max 32) and 84 (min 62- max 124) of these cells located, respectively, within 10μm and 20μm of PD1⁺ T cells. This corresponds to an average of 1.2% and 5.2% of B cells in less than 10µm or within 20µm proximity of T cells that may sufficient to favor contact between cognate B and T cells ( Figures 3E, F ). For the reasons mentioned above, we could not use the CD23 marker in parallel.

Since the LZ B_GC-cell compartment was heterogeneous with regard of CD23 expression after stimulation, we decided to study CD23^- and CD23⁺ LZ B_GC-cells obtained after 12 h of culture in the combined presence of IL-21, IL-4 and CD40L (hereafter referred to as pcCD23^- or pcCD23⁺ LZ B_GC-cells, where “pc” indicates “post-culture”). The gating strategy for cell sorting is presented in Supplemental Figure S4A . A comparison of these two subsets showed that FCER2 expression was elevated only in pcCD23⁺ LZ B_GC-cells ( Supplemental Figure S4B ). With regard to the four characteristic transcription factors involved in PC differentiation, there were no differences between the subsets in BCL6, PAX5 and XBP1 expression but PRDM1 was significantly upregulated in pcCD23^- LZ B_GC-cells ( Figure 4A ). Flow cytometry analysis showed significantly higher BCL6 expressing cells in the CD23⁺ subset while the CD23^- counterpart contained more BLIMP1⁺ cells ( Figure 4B ).

We then evaluated the subsets’ ability to differentiate into PBs (CD38^bright) and PCs (CD138⁺). To that end, pcCD23^- and pcCD23⁺ LZ B_GC-cells were cultured separately for 48 h in the presence of IL-2, IL-4 and IL-10. Statistically significant results showed that only pcCD23^- LZ B_GC-cells were able to give rise to a significant number of CD38^bright and CD138⁺ differentiated cells ( Figure 4C ).

The stimulation experiments were performed on total LZ B_GC-cells. However, to further specifically explore the CD23⁺ minority compartment of LZ B_GC-cells, sorted tonsillar CD23⁺ LZ B_GC-cells ( Supplemental Figure S2B ) were cultured for 24 h under the same conditions as above and analyzed by sc-qRT-PCR. No marker related to PC differentiation was detected after stimulation ( Supplemental Figure S4C ).

Our in vitro results and the literature data on LZ B_GC-cells (19) confirmed that the CD23⁺ and CD23^- subsets contained cells with distinct cell fates. To further explore this difference, we tracked spatiotemporally gene expression in stimulated LZ B_GC-cells. To this end, selected genes ( Supplemental Material, Method & Tables, Table 2 ) (19) were analyzed using sc-qRT-PCR at different time points of the culture. Thus, tonsil sorted PBs were compared with paired LZ B_GC-cells after 4 h and 24 h of culture with IL-21, IL-4 and CD40L. For the 24 h time point, CD23⁺ and CD23^- cells were sorted ( Figure 5A ). To investigate the time course of changes in cell populations, we applied the Monocle trajectory inference algorithm (https://doi.org/10.1038/nbt.2859). UMAP reduction of the five cell populations highlighted two clusters that differed in their cell fate as a function of CD23 expression ( Supplemental Figure S5 ). Only 24 h CD23^- LZ B_GC-cells followed continuous, homogeneous trajectory via expression of the BCL6, PAX5, IRF4, XBP1 and PRDM1 genes associated with differentiation into PBs ( Figures 5B, C ). However, some LZ B_GC-cells had started to migrate along the differentiation trajectory as early as 4 h, and clustered with PBs. Using Monocle, the UMAP projection was color-coded according to the pseudo-time; this showed a well-ordered progression in quadrants from LZ B_GC-cells to PBs ( Figure 5D ).

The expression heatmap for Monocle-ordered cells showed a sequential transition from CD19⁺IgD^-CD38⁺CD10⁺CXCR4^- LZ B_GC-cells to CD19⁺IgD^-CD38^bright PBs ( Figure 6A ). This situation enabled us to analyze changes over time in gene expression from one quadrant to the next and excluded T24h CD23⁺ B_GC cells in agreement with the above result ( Supplemental Figure S5 ). Each quadrant is enriched with a specific subset of cells ( Figure 6B ). The computed clustering allowed to identify six gene modules annotated by GeneMANIA ( Figure 6C ) characterized by a specific expression time course and linked to transcription factors, cell identity factors, and cell functions ( Supplemental Material, Method & Tables, Table 5 ). Modules 1 and 2 were linked to PC identity genes. Modules 3, 5 & 6 were linked to a B cell activation state. In fact, Module 3 encompassed several transcription factors (including MYC) and preceded Module 6, which was linked to transient cell cycle entry prior to extinction in the fourth quadrant (Q4). In Module 4, a transient expression pattern peaked during Q3 - probably in response to upstream factors such as MYC. It included the expression of ATF5, a transcription factor involved in the CREB3L2-ATF5-MCL1 survival pathway and which acts as a stress sensor (31). Module 5 featured a transient decline in the expression of B cell identity genes in Q1, a sharp increase in Q2 and Q3, and a drastic fall in Q4.

Our model of LZ B_GC-cell differentiation as a function of CD23 expression makes it possible to calculate how levels of transcription factors change in space and over time. The changes over time observed here were in line with the literature data ( Figure 6D ) (32). The IRF4 gene was expressed in two phases, with an initial peak preceding the increase in expression of the B cell identity genes PAX5, BCL6, BACH2, and SPIB. The second peak occurred when the four factors were no longer expressed. PAX5 expression showed a striking increase and peaked at the end of Q3 before dropping sharply and thus de-repressing the expression of PRDM1. This result is consistent with the fact that LZ B_GC-cells enter an activation state before they switch to the PB differentiation pathway (32). Although PRDM1 and XBP1s are late transcription factors that seal the commitment to PBs, XBP1 expression starts as soon as B cells activate. Furthermore, our model highlights the MYC imprint involved in the differentiation of B cells; our results are in agreement with MYC’s description as a mediator of B_GC-cell survival and cell-cycle re-entry and as a marker of positive selection (7, 32, 33, 34). MYC expression is absent in sorted LZ B_GC-cells but rises quickly in the timeline followed by the expression of early MYC-target genes TP53, CDK4 and CDK6 ( Figure 6E ) (35, 36) and then the proliferative genes from Module 6 plus CCNB1 (Module 4) – marking cell cycle re-entry.

The pseudo-time inferred from the Monocle algorithm provided an overview of the signal delivered by IL-4 (present in our differentiation cocktail) through the expression of STAT6, IL4R, FCER2 and the well-established IL-4/STAT6 target CCL22 (37) ( Figure 7A ). In line with our previous results (19), LZ B_GC-cells are unable to upregulate FCER2 once they have committed to the PBs pathway but do produce a clear STAT6 response marked by the expression of IL4R and CCL22. We then positioned on the sc-RNA-seq UMAP plot of total B_GC-cells from Holmes et al. (26) genes a B cell-specific IL-4 response signature ( Figure 7B ) (38). Positive cells for this signature were predominantly located in an area between pre-MBC and PB clusters, the latter two representing the two distinct fates for B_GC-cells ( Figure 7C ). A similar distribution was obtained for cells expressing FCER2 with higher density on the far right of the map, drawing a ridge line extending from the intermediate (INT) 6 cluster to the pre-MBC cluster ( Figure 7D ). Statistical analysis showed a striking enrichment for the IL-4 signature in INT5, INT6, light-zone (LZ) and pre-memory clusters compared to other clusters ( Figure 7E ). IL-4 signaling was significantly enriched in FCER2⁺ cells compared to FCER2^- counterparts in some clusters including INT6, LZ and pre-memory ( Figure 7F ). Interestingly, these three latter also exhibited the highest proportion of FCER2⁺ cells compared to other clusters ( Figure 7G ). Finally, unlike pre-MBCs, most differentiated PBs were negative for both IL-4 signature and FCER2 expression, confirming that in last steps of the PB commitment, cells turned off their IL-4/STAT6/FCER2 signaling (19).

Tonsils obtained from children undergoing routine tonsillectomy were obtained following approval by the French Ministry of Higher Education and Research (reference: AC-2014-2315). The children’s parents provided their informed consent to use of the samples for research purposes. The study was performed in accordance with the principles of the Declaration of Helsinki. Tonsil-derived B_GC-cells were enriched by negative selection via magnetic cell separation, using the MojoSort Human Pan B Cell Isolation Kit (Biolegend, San Diego, CA) with the addition of biotinylated IgD antibody (BD Biosciences, San Jose, CA) and the AutoMACS deplete program (Miltenyi Biotech, Bergisch Gladbach, Germany). LZ B_GC-cells (CD19⁺IgD^-CD10⁺CD38⁺CXCR4^lo) and PBs (CD19⁺IgD^-CD38^bright) were then sorted using a FACSAria system (BD Biosciences). All antibodies used for flow cytometry are listed in Supplemental Material, Method & Tables, Table 1 .

Single-cell experiments were performed using the Fluidigm^® C1 ™systems according to manufacturer instructions., Briefly, sorted cells were captured with CI Single-Cell Auto Prep integrated fluidic circuits (IFC) 5-10 µm (Fluidigm; 100-5757). Cells were then lysed, and reversed transcription and pre-amplification (Ambion Single cell-to-Ct Kit; 4458237) was done within the C1 system. Gene expression levels was then assessed by qPCR for selected taqman assays using Taqman Gene expression master mix Life technologies; 4369016) on 96.96 Dynamic Arrays IFC (Fluidigm; BMK-M-96.96) within the Fluidigm BioMark™ HD system. The list of TaqMan assay-on-Demand™ used is provided in Supplemental Material, Method & Tables, Table 2 .

Human tonsils and reactive lymph nodes were embedded in Optimal Cutting Temperature Compound (OCT, Sakura) and conserved at -80°C. Cryostat sections (18 µm thick) were fixed with 4% paraformaldehyde for 15 min at room temperature (RT). Sections were then incubated during 1 h with a blocking solution (PBS, 2% Bovine Serum Albumin, 4% donkey serum and 0.1% saponin) at RT and incubated in a humidified chamber overnight at 4° C with primary antibodies. Sections were washed with PBS 0.1% saponin and incubated with secondary antibodies for 1 h at RT. Finally, tissue sections were mounted with Mowiol (Merck) antifade reagent containing 20µM of SytoxBlue nucleic acid stain (Thermo Fischer) and analyzed by confocal microscopy on a SP8 (Leica Microsystems). ImageJ software (National Institutes of Health) was used for image analysis. The list of primary and secondary antibodies used for immunohistofluorescence is provided in Supplemental Material, Method & Tables, Tables 3, 4 .

For multiplex fluorescence microscopy and analysis, three FFPE samples of human tonsil were provided by the Pathology department of Rennes. Four-micrometer-thick whole-slide sections, obtained with a microtome (Histocore multicut Leicabiosystems, Nanterre, France) from FFPE tissue, were transferred onto plus-charged slides (VWR international), followed by multiplex immunofluorescence staining with a U DISCOVERY 5 plex immunofluorescence (Roche Diagnostics, Meylan, France). Four sequential rounds of staining were performed each including heat deactivation step, followed by incubation with primary antibody and corresponding HRP secondary antibody. Hence, primary antibodies expressions (as described in Supplemental Material, Method & Tables, Table 3 ) were visualized on the same section. HRP enzyme mediated deposition of the tyramide; coupled to respectively rhodamine, DCC, cyanine-5 and FAM fluorophores species (kits Ventana Medical Systems, Tucson, Arizona) that covalently bound to the tissue at the site of the reaction. After four sequential reactions, sections were counterstained with DAPI and cover slipped using fluoromount (Enzo Life Sciences, Farmingdale, NY, USA). Visualization was performed with the Nanozoomer (Hamamatsu Photonics, Massy, France) equipped with the multicolor fluorescence module.

For the pSTAT6-positive B_GC-cell percentage assessment, an automated analysis by machine learning with the HALO software was performed. After cell segmentation and nuclear detection pSTAT6, CD20, PAX5 and PD1 threshold intensities were set up. Six GCs in three different tonsils were analyzed. Double CD20 or PAX5 and pSTAT6-positive cells and total CD20- or PAX5-positive cells were quantified and a percentage of positive pSTAT6 B cells was assessed.

Bulk RNAseq datasets used to select genes analyzed by sc-qPCR in this study are available in the Gene Expression Omnibus database under accession no. GSE136990.

Single-cell-RNAseq datasets from Holmes et al.’s paper (26) are available in the Gene Expression Omnibus database under accession no. GSE139833. sc-gene expression data are available under accession no. GSE139891.

Quantitative variables were expressed as the mean ± standard deviation (SD). Statistical analyses were performed with Prism software (version 5, GraphPad Software, San Diego, CA) and R software (version 3.6.0). Statistical significance was assessed using the Mann-Whitney nonparametric U test and a one-way Anova and Kruskal-Wallis multiple comparison test (*P ≤ .05; **P < .01; ***P < .001; ****P < .0001).