All tissues used for this study came from subjects recruited under institutional review board approval and informed consent process according to the Declaration of Helsinki. Human stromal cells from tonsils were obtained from pediatric patients undergoing routine tonsillectomy, as previously described (35). These cells were then cultured with 10% FCS in RPMI 1640 (Thermo Fisher, Waltham, USA) supplemented with penicillin and streptomycin. Lymphoid cells were recovered by perfusion of FL LN biopsies (n = 14) ( Supplementary Table 1 ) or non-malignant tonsils (n = 12), enriched using Ficoll gradient and both magnetic and FACS sorting (see the procedure and gating strategies illustrated in the Supplementary Material and Methods ).

Samples were acquired on a Cytoflex^® flow cytometer (Beckman Coulter, Miami, FL, USA). Cell viability was determined by staining with live/dead fixable yellow dead cell stain kit (Thermo Fisher) prior to membrane marker staining and when required, fixation and permeabilization with eBiosciences Foxp3/Transcription Factor Staining Buffer Set (Life Technologies, Courtaboeuf, France). Cells were stained with antibodies referenced in Supplementary Table 2 . A minimum of 100,000 cells per sample were acquired, and the gating strategy is illustrated in Figure 1 . Subset analyses were done with FlowJo 10.4, LLC (Becton Dickinson, Franklin Lakes, USA), and the compensation matrix is provided in Supplementary Table 3 . Statistical analysis was done with GraphPad Prism 6 software suite (GraphPad Software, La Jolla, CA, USA), the Mann–Whitney test was used, and standard deviations were illustrated.

Microarray analyses were done on three samples of each following population: Tfh, non-follicular Treg, memory T, and naive T cells from the tonsils, as well as purified Tfh, Tfr, memory T, and naive T cells from FL LN, according to the gating strategies detailed in the Supplementary Material and Methods . RNA preparation and data preprocessing are also detailed in the Supplementary Material and Methods . Gene signatures were obtained using a t-test by comparison of memory T cells and each population of interest: Tfh or Tfr with limma package on R. To evaluate Tfr gene expression proximity to Tfh or Treg and activated Tfh to either FL-Tfh or tonsil Tfh, a pre-ranked Gene Set Enrichment Analysis [GSEA software and Molecular Signature Database (MSigDB)] was done. Classical GSEA was performed to determine the differentially enriched pathways between FL-Tfh and tonsil Tfh using the MSigDB hallmark collection and the resulting output to generate heatmaps (Heatmap.2 package). A similar GSEA strategy was used for centrocytes and FL-B-cell transcriptomic data previously published (36).

TSDR analysis was performed with 500 ng of genomic DNA from FL LN naive Treg (nTreg), Tfr, and Tfh (n = 3) (the sorting strategies as technical and biological bias exclusion are detailed in the Supplementary Material and Methods ), as previously described (37). The EZ DNA Methylation Kit (Zymo Research, Freiburg, Germany) was used for bisulfite conversion, and the bisulfite-converted DNA was hybridized onto the HumanMethylation 450K BeadChip kit (Illumina, Paris, France). Data from the 450k Human Methylation Array were analyzed in R using the minfi package (version: 1.6.0).

RNA was extracted from 150,000 cells of Tfh, Tfr, and Treg from FL LN using the Qiagen RNeasy Mini kit (Qiagen, Hilden, Germany). TCR sequences were amplified by unbiased RACE-PCR using the Clontech SMARTer RACE 5′/3′ kit (Clontech Laboratories, Inc., Mountain View, USA). Library preparation and the associated primer sequences are detailed in the Supplementary Material and Methods . Sequencing was done with a MiSeq Illumina and MiSeq Reagent Nano Kit v2. Pear software was used to match pair ends. FASTA sequences were then submitted to IMGT/HighV-Quest (http://www.imgt.org/) to align sequences with the reference database. Finally, clonotypes identified by IMGT were used to perform analysis with R packages (detailed analysis in the Supplementary Material and Methods ). Circular graphics were done through the Circos package run on PERL.

Tfh in-vitro survival was compromised without feeder cells, leading us to co-cultivate Tfh with tonsil stromal cells. In a round-bottom 96-well plate, 5,000 stromal cells were plated 1 day prior to Tfh seeding. Sorted Tfh were stained with CellTrace Far Red Kit (Thermo Fisher) to monitor their proliferation in culture. Either 60,000 Tfh or 40,000 Tfh + 20,000 Tfr were then plated in RPMI 1640 with 8% FCS (Thermo Fisher, Waltham, USA) and stimulated for 4 days with anti-CD3 (0.6 μg/ml) and anti-CD28 (0.2 μg/ml) stimulating antibodies (Sanquin, Amsterdam, Netherlands). On day 4, the supernatants were collected for cytokine quantification, while cells were trypsinized and stained with Annexin-V FITC, DAPI, and anti-CD7 BV711 allowing viable T-cell gating. Except for feeder cell usage, the same was done using CD8⁺ T cells stained with CellTrace Far Red. The proliferation index was calculated on the Annexin-V⁻ DAPI⁻ Cell Trace⁺ CD7⁺ population by ModFit LT software (Verity Software House, Topsham, USA).

To determine cytokine content in the co-culture supernatant, we used Luminex technology (Thermo Fisher, Waltham, MA, USA) with the MILLIPLEX kit: 7-Plex Human cytokine Mag kit comprising IFN-g, IL-4, IL-6, IL-10, sCD40L, MDC, and IL-21 (HSTCMAG-28SK) (Millipore, Fontenay Sous Bois, France), according to the manufacturer’s protocol. All cytokine concentrations were quantified within the same plate to avoid interexperimental variability. Results were processed with GraphPad. The Mann–Whitney test was used to determine significance.

Human tonsils, reactive LN, and FL LN were used (n = 3 for each tissue). FL grade 1–2 was selected, according to the World Health Organization criteria. Four-micrometer-thick whole-slide sections, obtained from FFPE tissue blocks, were processed by multiplex immunofluorescence staining with a U DISCOVERY 5-Plex IF. The staining and mounting procedures are detailed in the Supplementary Material and Methods . Visualization was performed with the NanoZoomer (Hamamatsu, Massy, France). FOXP3⁺/CD8⁺ and CD25⁺/CD8⁺ cell ratios were obtained in 10 follicles of each sample.

All statistics were obtained from the GraphPad Prism 6 software suite (GraphPad Software, La Jolla, CA, USA). The statistical significance of the cytometry experiments was assessed through the non-parametric Wilcoxon test for matched pairs, and the Mann–Whitney test was used for unpaired data. T-test was used when a parametric test was required, and only two groups were tested for differential gene expression and to perform a correlation curve. When more than two groups had to be compared within cytometry or transcriptomic experiments, one-way ANOVA was done with Dunn’s multiple test correction or two-way ANOVA with Bonferroni post-test. Error bars correspond to the standard deviation. *, **, and *** correspond to p < 0.05, p < 0.01, and p < 0.005, respectively.

Data are available on the GEO repository, with accession number GSE222532.

To characterize CD4 T-cell subsets in FL, we first defined these different subpopulations in the tonsils and FL lymph nodes. Among CD4 T cells, we distinguished follicular cells expressing CXCR5 from non-follicular cells which are lacking it, with their discrimination being strengthened with either PD-1 ( Figure 1E ) or ICOS ( Figure 1F ) expression as an additional segregating marker. Regulatory T cells among follicular ( Figure 1G ) and non-follicular T cells ( Figure 1H ) were distinguished by their expression of Foxp3. Therefore, follicular T cells contain Tfh defined as CD3⁺CD4⁺CXCR5^hiICOS⁺Foxp3⁻ cells, which can be discriminated from their regulatory counterpart expressing Foxp3 (CD3⁺CD4⁺CD45RA⁻XCR5⁺ICOS⁺Foxp3⁺), while non-follicular T cells comprise classical Treg cells (CD3⁺CD4⁺CXCR5⁻ICOS⁻Foxp3⁺) and other memory T cells (CD3⁺CD4⁺CD45RA⁻CXCR5⁻Foxp3⁻).

Since different T-cell subsets were found participating in FL maintenance and the role of regulatory T cells in this context is still ambiguous, we carefully analyzed CD4⁺ regulatory T-cell subsets from malignant LN (n = 14) and compared them with their tonsil counterpart (n = 12). We found a significant expansion of the Foxp3⁺ subset among CD4⁺ T cells in FL LN (p = 0.015, mean tonsil = 4.3 ± 0.499%, mean FL black and red squares = 13.7 ± 3.3%) ( Figure 2A , left). This expansion relied on the significant amplification (p < 0.001) of cells harboring a Tfr phenotype, i.e., Foxp3⁺CXCR5⁺CD4⁺ T cells, rather than Foxp3⁺ CXCR5⁻ corresponding to non-follicular Treg (mean tonsil CD4⁺Foxp3⁺CXCR5⁺ = 1.91 ± 0.206%, mean tonsil CD4⁺Foxp3⁺CXCR5⁺ = 1.695 ± 0.834%, mean FL CD4⁺Foxp3⁺CXCR5⁻ = 5.16 ± 3.313%, mean FL CD4⁺Foxp3⁺CXCR5⁺ = 8.07 ± 8.057%; Figure 2A , right). A strong correlation between Foxp3 and CD25 expression was observed in both tonsil and FL LN CD4⁺ T cells (p < 0.001, Figure 2B , left), and the CD25⁺ subpopulation was the major contributor to the CXCR5⁺Foxp3⁺ FL CD4⁺ T-cell expansion (mean tonsil CD4⁺Foxp3⁺CXCR5⁺CD25⁺ = 4.944 ± 2.738%, mean FL CD4⁺Foxp3⁺CXCR5⁺CD25⁺ = 17.50 ± 10.69%, mean FL CD4⁺Foxp3⁺CXCR5⁺CD25⁻ = 4.134 ± 1.765%; Figure 2B , right). In agreement, we validated by immunofluorescence the presence of follicular Foxp3⁺CD25⁺ cells within FL malignant follicles ( Figure 2C ). The addition of ICOS to the CD4, CXCR5, and CD25 markers demonstrated a better efficiency than PD1 to retrieve follicular CD4⁺ T cells expressing Foxp3 in FL ( Supplementary Figure 1A ). Accordingly, this CD4⁺CXCR5⁺CD25⁺ICOS⁺ population was found expanded in FL ( Figure 2D , left; mean tonsils = 1.188 ± 0.905%, mean FL = 5.377 ± 6.61%, p = 0.028). To assert the Tfr phenotypical identity of this subset, we explored its Bcl6 expression and found that it was significantly higher in Tfr (RMFI: Tfr/CXCR5⁺CD25⁺ICOS⁺ T cells = 1.5 ± 0.32, Tfh cells = 2.2 ± 0.46) than in CXCR5⁻ non-follicular Treg (non-follicular Treg = 1 ± 0.059) ( Figure 1D , middle; p < 0.05), while Foxp3 expression was similar in the two regulatory subsets ( Figure 2D , right; RMFI: Tfr/CXCR5⁺CD25⁺ICOS⁺ T cells =2.937 ± 0.3547, non-follicular Treg = 2.383 ± 0.3376) compared with the decreased one of Tfh (Tfh = 1.087 ± 0.092). Altogether, these indicated that FL-Tfr can be sorted for transcriptomic and functional studies based on their CD25 and ICOS high expression.

To get deeper insights in the FL-Tfr profile, CD4⁺CXCR5⁺CD25⁺ICOS⁺ FL-Tfr (n = 3) were sorted from selected FL LN samples where more than 90% of these cells expressed Foxp3 (depicted in red in Figure 2 , description of the corresponding patients in Supplementary Table 1 ) and their transcriptomic profiles were compared with those of FL-Tfh (n = 3), FL-CXCR5⁻Foxp3⁻ memory CD4⁺ T cells (FL-Tmem) (n = 3), inflamed tonsil (Tons)-derived Tfh (Tons-Tfh) (n = 3), Tons-non-follicular Treg (Tons-Treg) (n = 3), and Tons-Tmem (n = 3). FL T cells clustered according to their subset of origin ( Figure 3A , left, and Supplementary Figure 2A ). Tonsil-non-follicular Tmem was used as a reference population to highlight either Tfh- or Tfr-specific features as characteristics related to their physiological or malignant microenvironment. Comparison of FL-Tfr with cells highlighted the regulatory phenotype of FL-Tfr with high levels of IL2RA, IL10, CTLA-4, Foxp3, and TIGIT transcripts ( Figure 3A , middle). Furthermore, GSEA revealed that the FL-Tfr transcriptomic profile was enriched for genes expressed by Tons-Treg, like IL2RA, CTLA-4, and Foxp3, rather than genes expressed by Tons-Tfh (p < 0.0001) ( Figure 3A , right). In addition, we found a significant overlap (p = 1.025 e⁻²¹) between the FL-Tfr transcriptomic signature and the murine Tfr signature (22) ( Supplementary Figure 2B ). Finally, FL-Tfr clustered with Tons-Treg in the methylome analysis ( Figure 3B , left), in agreement with their preferential demethylation of Foxp3, Ctla-4, Tnfrsf9, and Ikzf4, in comparison with FL-Tmem or FL-Tfh ( Supplementary Figure 2C ). Finally, we studied Foxp3 CNS2 demethylation status, a marker of all Treg subsets regardless of their thymic or peripheral origin (37), in FL-Tfr ( Figure 3B , right) (n = 3). As expected, Foxp3 CNS2 of FL-Tfh was found poorly demethylated (mean = 0.58 ± 0.51%), while demethylation was high and similar in FL-Tfr (mean = 75.6 ± 9.94%), and FL I CD45RA⁺ Treg was used as a control for thymic Treg (mean = 74.08 ± 8.51%), supporting the regulatory lineage affiliation of FL-Tfr. Altogether, these data demonstrate that in FL, follicular CD4⁺ T cells co-expressing CD25 and ICOS correspond to Foxp3⁺Bcl6⁺ Tfr cells. At both the transcriptomic and methylome levels, these FL-Tfr display a regulatory phenotype similar to Treg.

To determine whether FL-Tfr originate from thymic Treg and/or from the peripheral differentiation of activated CD4⁺ T cells, we performed TCR repertoire analysis of different CD4⁺ T-cell subsets populating FL malignant LN (n = 3). FL-Tfh, FL-Tfr, and FL-non-follicular Treg were sorted, and their TCR-Vβ were sequenced following an unbiased TCR-Vβ sequence amplification by RACE-PCR. Approximately 150,000 cells per subset were sorted from each FL sample, allowing to collect 14,000 to 28,000 clones with a homogeneous repartition among subsets and clone frequencies ranging from 1 to 2,000. Although a low number of clones were shared between Tfh and Treg or even by all cell subsets, a more robust repertoire community was observed between both Tfr and Tfh, as well as between Tfr and Treg ( Figure 4A ). Remarkably, clones shared between FL-Tfh and FL-Tfr were found among the most frequent clones ( Figure 4A , Circos plot, clones frequency is represented as black histograms), suggesting that they resulted from antigen-driven amplification of a common precursor or their polarization toward an alternative phenotype (Tfh or Tfr). In line with this result, an overlap between FL-Tfr and FL-Tfh (FL-Tfh/FL-Tfr overlap = 5%) was found almost doubling when the analysis was restricted to the 1,000 most frequent clones (FL-Tfh/FL-Tfr overlap = 9%) ( Figure 4B , up and down). Furthermore, for both the FL and tonsil samples, the Horn–Morisita index, which illustrates the repertoire degree of similarity, was found higher for Tfh/Tfr than for Tfr/Treg, while the Tfh/Treg overlap was null as expected ( Figure 4C ). Altogether, these results identified Tfr as clonally related primarily to Tfh but also to Treg, particularly in FL. Whether this differential origin has an impact on their functional activity still remains to be explored.

Physiologically, Tfr are known to modulate Tfh and GC-B-cell responses. In the FL context, Tfr functions have not yet been estimated. Autologous culture of sorted Tfr and Tfh from the tonsils or FL LN was thus performed in the presence of anti-CD3/-CD28 stimulation (n = 3). As expected, the proliferation of activated tonsil Tfh was significantly reduced (p < 0.001) in the presence of Tons-Tfr. A similar trend was observed when FL-Tfh were co-cultured with FL-Tfr ( Figure 5A , top). Interestingly, in-vitro stimulated FL-Tfh displayed a lower proliferation index than Tons-Tfh, and the addition of FL-Tfr abrogated the residual proliferative capacity of FL-Tfh. This finding was in line with the significantly lower percentage of KI67⁺ cells found within ex-vivo FL-Tfh (6.796 ± 1.963%) compared with Tons-Tfh (13.06 ± 0.33%) (p < 0.05) ( Figure 5A , bottom), and the specific downregulation of pathways related to cell cycle and growth ( Figure 5B , left part) was investigated by the transcriptomic approach. mTORC1 and glycolysis pathways were also found downregulated in FL-Tfh, matching the signature of in-vitro murine Tfh suppressed by Tfr (34). These data argue for a FL-Tfh regulation by FL-Tfr in vivo. Of note, the same pathways were found downregulated in FL-B cells compared with healthy centrocytes ( Figure 5B , right part), hypothesizing that FL-B cells were subjected to the same regulatory environment as FL-Tfh. However, both FL-Tfh and FL-B cells exhibited an increased effector profile. The ICOSL and IL-6 genes were found highly expressed in FL-B cells ( Figure 5C , left and middle, respectively), while FL-Tfh exhibited more IL-4, IL-2, IFN-g, and TNFRSF9 mRNAs than their respective tonsil counterparts ( Figure 5C , right). Since the prognostic value of Treg in FL has been linked to their capacity to inhibit tumor-infiltrating cytotoxic CD8⁺ T cells (19), we decided to study the function of CD8⁺ T cells sorted from either the tonsils or FL LN. Anti-CD3/-CD28 antibody stimulation induced CD8⁺ T-cell proliferation ( Figure 5D , left, circles) (n = 3). Nevertheless, FL CD8⁺ T cells exhibited a lower proliferation index than tonsil CD8⁺ T cells. The addition of autologous Tfr significantly reduced the proliferation capacity of CD8⁺ T cells in both tonsil and FL contexts (p < 0.05) ( Figure 5D , left, triangles). In accordance, less IFN-γ (mean CD8 = 233 ± 122.3 pg/ml vs. mean CD8+Tfr = 8.857 ± 6.869 pg/ml) and IL-2 Tfr (mean CD8 = 95.56 ± 57.73 pg/ml vs. mean CD8 + Tfr = 6.527 ± 2.797 pg/ml) were found in the supernatant of activated CD8⁺ T cells co-cultured with FL-Tfr ( Figure 5D , middle and right). Finally, to confirm that FL-Tfr indeed impacted CD8⁺ T cells in vivo, we assessed intrafollicular Foxp3⁺CD25⁺/CD8⁺ cell ratio in FL LN in comparison to reactive LN and inflamed tonsils (n = 3). We observed a significant increase (FL: 0.45 ± 0.089, reactive LN: 0.17 ± 0.05, tonsil: 0.04 ± 0.02, p < 0.0001) of the intrafollicular Foxp3⁺CD25⁺/CD8⁺ cell ratio in FL LN samples ( Figure 5E ), and a similar trend was observed by flow cytometry ( Supplementary Figure 3 ), strengthening our hypothesis of a CD8⁺ T-cell control by FL-Tfr in vivo. Therefore, in light of these results, FL-Tfr likely exert their regulatory function on lymphoid cell proliferation in malignant follicles, but FL-Tfh and B cells somehow escape Tfr functional control in contrast to CD8⁺ T cells.