Thymic selection of the T cell receptor repertoire is biased toward autoimmunity in females

Reviewer #1 (Public review):

Summary:

The goal of this paper was to determine whether the T cell receptor (TCR) repertoire differs between a male or female human. To address this, this group sequenced TCRs from double-positive and single-positive thymocytes in male and female humans of various ages. Such an analysis on sorted thymocyte subsets has not been performed in the past. The only comparable dataset is a pediatric thymocyte dataset where total thymocytes were sorted.

They report on participant ages and sexes, but not on ethnicity, race, nor provide information about HLA typing of individuals. The experiments are heroic, yet do represent a relatively small sampling of diverse humans. They observed no differences in TCRbeta or TCRalpha usage, combinational diversity, or differences in the length of the CDR3 region, or amino acid usage in the CD3aa region between males or females. Though they observed some TCRbeta CD3aa sequence motifs that differed between males and females, these findings could not be replicated using an external dataset and therefore were not generalizable to the human population.

They also compared TCRbeta sequences against those identified in the past databases using computational approaches to recognize cancer-, bacterial-, viral-, or autoimmune-antigens. They found little overlap of their sequences with these annotated sequences (depending on the individual, ranged from 0.82-3.58% of sequences). Within the sequences that were in overlap, they found that certain sequences against autoimmune or bacterial antigens were significantly over-represented in female versus male CD8 SP cells. Since no other comparable dataset is available, they could not conclude whether this is a generalizable finding in the human population.

Strengths:

It is a novel dataset that attempts to understand sex differences in the T cell repertoire in humans. Overall, the methodologies are sound and are the current state-of-the-art. There was an attempt to replicate their findings in cases where an appropriate dataset was available. I agree that there are no gross differences in TCR diversity between males and females. This is an important negative result.

Weaknesses:

Overall, the sample size is small given that it is an outbred population. This reviewer recognizes the difficulty in obtaining samples for this experiment (which were from deceased donors), and this limitation was appropriately discussed. Their analysis was limited by the current availability of other TCR sequences. These weaknesses were appropriately discussed and considered.

Reviewer #2 (Public review):

Summary:

This study addresses the hypothesis that the strikingly higher prevalence of autoimmune diseases in women could be the result of biased thymic generation or selection of TCR repertoires. The biological question is important and the hypothesis is valuable. Although the topic is conceptually interesting and the dataset is rich, the study has a number of major issues. In particular, the majority of "autoimmunity-related TCRs" considered in this study are in fact specific to type 1 diabetes (T1D). Notably, T1D incidence is higher in males, which directly contradicts the stated objective of the study - to explain the higher prevalence of autoimmune diseases in women. Given this conceptual inconsistency, the evidence presented does not support the authors' conclusions.

Strengths:

The key strength of this work is the newly generated dataset of TCR repertoires from sorted thymocyte subsets (DP and SP populations). This approach enables the authors to distinguish between biases in TCR generation (DP) and thymic selection (SP). Bulk TCR sequencing allows deeper repertoire coverage than single-cell approaches, which is valuable here, although the absence of TRA-TRB pairing and HLA context limits the interpretability of antigen specificity analyses. Importantly, this dataset represents a valuable community resource and should be openly deposited rather than being "available upon request."

Weaknesses:

I thank the authors for their detailed responses to my previous comments. Several concerns were addressed satisfactorily; however, important issues remain unresolved, and a new major concern has emerged from the revised manuscript.

Major concerns:

(1) Autoimmune specificity is dominated by T1D, contradicting the study's premise. Newly added supplementary Table 3 shows that the authors considered only 14 autoimmune-related epitopes, of which 12 are associated with type 1 diabetes (T1D) and 2 with celiac disease (CeD). (I guess this is because identification of particular peptide autoantigens is an extremely difficult task and was only successful in T1D and CeD.) Thus conclusions of this work mostly relate to T1D. However, the incidence of T1D is higher in males than in females (e.g. doi:10.1111/j.1365-2796.2007.01896.x; doi:10.25646/11439.2). This directly contradicts the stated objective of the study - to explain the higher prevalence of autoimmune diseases in women. As a result, the authors' conclusions (a) cannot be generalized to autoimmune disease as a whole as the authors only considered T1D and CeD antigens and (b) are internally inconsistent with the stated objective of the study.

(2) By contrast, CeD does show a female bias (~60/40 female/male; doi:10.1016/j.cgh.2018.11.013). However, the manuscript does not allow evaluation of how much the reported "autoimmune TCR enrichment" derives from T1D versus CeD. Despite my previous request, the authors did not provide per-donor and per-epitope distributions of autoimmune-specific TCR matches. I therefore explicitly request a table in which: each row corresponds to a specific autoimmune antigen; each column corresponds to a donor (with metadata available including sex); each cell reports the number of unique TCRs specific to that antigen in that donor. Without such data, the conclusions cannot be evaluated.

(3) It is scientifically inappropriate to generalize findings to "autoimmune diseases" when only T1D and CeD were analyzed. Moreover, given that T1D and CeD show opposite directions of sex bias, combining them into a single "AID" category is misleading. All analyses presented in Figure 8 and Supplementary Figure 16 should be repeated and shown separately for T1D and CeD, rather than combined.

(4) The McPAS database contains TCRs associated with other autoimmune diseases (e.g., multiple sclerosis, rheumatoid arthritis), although the exact autoantigens in these contexts are unknown. Why didn't the authors perform the search for such TCRs? I believe disease association even without particular known antigen could still be insightful.

(5) Misuse of the concept of polyspecificity. I appreciate the authors' reference to Don Mason's work; however, the concept of polyspecificity discussed there is fundamentally different from the authors' usage. Mason, Sewell (doi:10.1074/jbc.M111.289488), Garcia (doi:10.1016/j.cell.2014.03.047), and others demonstrated that individual TCRs can recognize multiple peptides, possibly around 1 million. But importantly these peptides are not random but share some sequence motif. This is a general feature of TCRs, i.e. 100% of TCRs are polyspecific in this sense.
In contrast, the authors define polyspecificity as TRB sequences annotated as specific to unrelated epitopes in TCR databases such as VDJdb. These databases are well known to contain substantial numbers of false-positive annotations (see, e.g., Ton Schumacher's preprint https://www.biorxiv.org/content/10.1101/2025.04.28.651095.abstract). The authors acknowledge that, under their definition, polyspecificity has been experimentally validated for only one (!) TCR (Quiniou et al.). In the absence of robust experimental validation, use of the term "polyspecificity" in this context is misleading. I strongly recommend removing all analyses and conclusions related to polyspecificity from the manuscript unless supported by independent functional validation.

(6) I agree that comparing specificity enrichment between sexes is meaningful. However, enrichment relative to the database composition itself is not biologically interpretable, as acknowledged by the authors in their response. I therefore recommend removing Supplementary Figure 15, which is potentially misleading.

(7) In contrast, Supplementary Figure 16 represents the most convincing result of the study (keeping in mind that the AID group should be splitted to T1D and CeD with T1D and that T1D and CeD have opposing directions of sex biases) and should be shown as a main figure, replacing Figure 8A-B which is less convincing as it doesn't show per-donor distribution.

(8) The authors argue that applying mixed-effects modeling to Rényi entropy would require assuming a common sex effect across subsets. I do not find this assumption unreasonable. For example, if sex effects are mediated through AIRE-dependent negative selection, one would indeed expect a consistent direction of effect across subsets. The lack of statistical significance in Figure 3 may reflect limited sample size rather than true absence of the difference. Moreover, the title's phrasing "comparable TCR repertoire diversity" is vague: what is the statistical definition of "comparable"?

Author response:

The following is the authors’ response to the original reviews

eLife Assessment

This study provides useful insights into addressing the question of whether the prevalence of autoimmune disease could be driven by sex differences in the T cell receptor (TCR) repertoire, correlating with higher rates of autoimmune disease in females. The authors compare male and female TCR repertoires using bulk RNA sequencing, from sorted thymocyte subpopulations in pediatric and adult human thymuses; however, the results do not provide sufficient analytical rigor and incompletely support the central claims.

The statement in the editorial assessment that our study “does not provide sufficient analytical rigor” surprised us. TCR repertoire analysis is indeed a highly complex domain, both experimentally and computationally. We consider ourselves to be leading experts in this field and have invested a great deal of effort to ensure the rigor and reproducibility of every analytical step.

Specifically, our group has previously benchmarked and published validated methodologies for the following areas: (i) TCR repertoire generation (Barennes et al., Nat Biotechnol 2021), (ii) repertoire analysis (Six et al., Frontiers in Immunol, 2013; Chaara et al., Frontiers in Immunol, 2018; Ritvo et al., PNAS, 2018; Mhanna et al., Diabetes, 2021; Trück et al., eLife, 2021; Quiniou et al., eLife, 2023; Mhanna et al., Cell Rep Methods, 2024; Mhanna et al., Nat Rev Primers Methods, 2024), and (iii) the curation and quality control of public TCR databases (Jouannet et al., NAR Genomics and Bioinformatics 2025). The current study applies these optimized and peer-reviewed pipelines, along with additional internal quality controls that we have been implemented over the years, ensuring the highest possible analytical standards for TCR repertoire studies.

We therefore respectfully feel that the phrase “insufficient analytical rigor” does not accurately reflect the methodological robustness of our work. This perception is also in contrast to the comment made by one of the reviewers, who explicitly noted that “overall, the methodologies appear to be sound.”

We would therefore be grateful if, upon reviewing our detailed point-by-point responses, the editors could reconsider this statement and tone it down in the final editorial summary.

With regard to comment that our results “incompletely support the central claims”, we will leave it to the reader’s judgement. We believe that our work provides a robust and transparent basis for future research into TCR repertoire, autoimmunity, and women’s health.

Reviewer 1 (Public reviews):

Summary

The goal of this paper was to determine whether the T cell receptor (TCR) repertoire differs between a male and a female human. To address this, this group sequenced TCRs from doublepositive and single-positive thymocytes in male and female humans of various ages. Such an analysis on sorted thymocyte subsets has not been performed in the past. The only comparable dataset is a pediatric thymocyte dataset where total thymocytes were sorted.

They report on participant ages and sexes, but not on ethnicity, race, nor provide information about HLA typing of individuals. Though the experiments themselves are heroic, they do represent a relatively small sampling of diverse humans. They observed no differences in TCRbeta or TCRalpha usage, combinational diversity, or differences in the length of the CDR3 region, or amino acid usage in the CD3aa region between males or females. Though they observed some TCRbeta CD3aa sequence motifs that differed between males and females, these findings could not be replicated using an external dataset and therefore were not generalizable to the human population.

They also compared TCRbeta sequences against those identified in the past using computational approaches to recognize cancer-, bacterial-, viral-, or autoimmune-antigens. They found very little overlap of their sequences with these annotated sequences (depending on the individual, ranging from 0.82-3.58% of sequences). Within the sequences that were in overlap, they found that certain sequences against autoimmune or bacterial antigens were significantly over-represented in female versus male CD8 SP cells. Since no other comparable dataset is available, they could not conclude whether this is a finding that is generalizable to the human population.

Strengths:

This is a novel dataset. Overall, the methodologies appear to be sound. There was an attempt to replicate their findings in cases where an appropriate dataset was available. I agree that there are no gross differences in TCR diversity between males and females.

We appreciate the positive feedback from the reviewer regarding these points.

Weaknesses:

Overall, the sample size is small given that it is an outbred population. The cleaner experiment would have been to study the impact of sex in a number of inbred MHC I/II identical mouse strains or in humans with HLA-identical backgrounds.

We respectfully disagree with the reviewer’s statement. We firmly believe that the issue we are dealing with, namely sex-based differences in thymic TCR selection relevant to autoimmunity, should be investigated more thoroughly in the general human population than in inbred mouse models.

While inbred mouse strains, being MHC I/II identical, eliminate the complexity of MHC variation, this comes at the cost of biological relevance. Firstly, a discrepancy in TCR generation or selection may only become apparent under specific MHC contexts, which could easily be overlooked when studying a single inbred strain. Secondly, inbred strains frequently contain fixed genetic variants that may influence thymic selection or immune regulation. This has the potential to introduce confounding effects rather than reducing them and not solving the generalization issue.

We are in full agreement that an HLA-matched human cohort would reduce inter-individual variability. However, such sampling is impossible in practice, as our thymic tissues were obtained from deceased organ donors, a collection effort that was, as the reviewer rightly noted, “heroic”. Despite these inherent limitations, the patterns we observed were consistent across multiple analytical approaches, lending robustness to our findings.

We now explicitly acknowledge this limitation in the Discussion of the revised manuscript and explain why, despite this constraint, our study provides meaningful and biologically relevant insights into human TCR selection and sex-related immune differences.

It is unclear whether there was consensus between the three databases they used regarding the antigens recognized by the TCR sequences. Given the very low overlap between the TCR sequences identified in these databases and their dataset, and the lack of replication, they should tone down their excitement about the CD8 T cell sequences recognizing autoimmune and bacterial antigens being over-represented in females.

The three databases used in this study - McPAS-TCR, IEDB, and VDJdb - provide complementary and partially non-overlapping specificity landscapes. McPAS-TCR is enriched for pathology-associated TCRs, while IEDB and VDJdb contain a higher proportion of viral specificities. Combining them therefore broadens the antigenic spectrum accessible for analysis and represents the most comprehensive approach currently possible to capture the diversity of TCR–antigen annotations.

With regard to the limited overlap between our dataset and these databases, this observation should be interpreted with caution. While the overlap may appear minimal at first glance, it is a biologically significant phenomenon. The public databases collectively contain only a minute fraction of the total universe of TCR specificities, estimated to exceed 10^15-21 possible receptors in humans. In this context, the observation of any overlap at all, particularly with coherent biological patterns such as the overrepresentation of autoimmune- and bacterialassociated TCRs in females, is noteworthy.

We have included a short clarification in the Discussion of the revised manuscript to make this point explicit and to further temper the language describing this finding.

The dataset could be valuable to the community.

We thank the reviewer for highlighting the potential value of this dataset to the community. It will be made publicly available on the NCBI website. We would like to clarify that our intention has always been to make this dataset publicly available; therefore, we take back any incorrect suggestions made in the original submission.

Reviewer #1 (Recommendations for the authors):

I would just recommend toning down the excitement about autoimmune TCRs being overrepresented in females. Then the conclusions will be in alignment with their results.

We thank the reviewer for this constructive recommendation. We would like to express our full support for the editorial transparency policies of eLife, which allow readers to access to both the reviewers’ comments and our detailed responses, enabling them to form their own informed opinions regarding our conclusions.

Nevertheless, we have moderated some of our wording.

Reviewer #2 (Public review):

Summary:

This study addresses the hypothesis that the strikingly higher prevalence of autoimmune diseases in women could be the result of biased thymic generation or selection of TCR repertoires. The biological question is important, and the hypothesis is valuable. Although the topic is conceptually interesting and the dataset is rich, the study has a number of major issues that require substantial improvement. In several instances, the authors conclude that there are no sex-associated differences for specific parameters, yet inspection of the data suggests visible trends that are not properly quantified. The authors should either apply more appropriate statistical approaches to test these trends or provide stronger evidence that the observed differences are not significant. In other analyses, the authors report the differences between sexes based on a pulled analysis of TCR sequences from all the donors, which could result in differences driven by one or two single donors (e.g., having particular HLA variants) rather than reflect sex-related differences.

Strengths:

The key strength of this work is the newly generated dataset of TCR repertoires from sorted thymocyte subsets (DP and SP populations). This approach enables the authors to distinguish between biases in TCR generation (DP) and thymic selection (SP). Bulk TCR sequencing allows deeper repertoire coverage than single-cell approaches, which is valuable here, although the absence of TRA-TRB pairing and HLA context limits the interpretability of antigen specificity analyses. Importantly, this dataset represents a valuable community resource and should be openly deposited rather than being "available upon request."

We thank the reviewer for highlighting the potential value of this dataset to the community. It will be made publicly available on the NCBI website. We would like to clarify that our intention has always been to make this dataset publicly available; therefore, we take back any incorrect suggestions made in the original submission.

Weaknesses:

Major:

The authors state that there is "no clear separation in PCA for both TRA and TRB across all subsets." However, Figure 2 shows a visible separation for DP thymocytes (especially TRA, and to a lesser degree TRB) and also for TRA of Tregs. This apparent structure should be acknowledged and discussed rather than dismissed.

We thank the reviewer for this careful observation. Discussing apparent “trends” rather than statistically significant results is indeed a nuanced issue, as over-interpretation of visual patterns is usually discouraged. We agree that, within the specific context of TCR repertoire analyses, visual structures in multivariate projections such as PCA can provide useful contextual information.

However, we have not identified a striking trend in our representation. We therefore chose to avoid overemphasizing these visual impressions in the text.

Supplementary Figures 2-5 involve many comparisons, yet no correction for multiple testing appears to be applied. After appropriate correction, all the reported differences would likely lose significance. These analyses must be re-evaluated with proper multiple-testing correction, and apparent differences should be tested for reproducibility in an external dataset (for example, the pediatric thymus and peripheral blood repertoires later used for motif validation).

As is standard in exploratory immunogenomic studies, including TCR repertoire analyses, our objective was to uncover broad biological patterns rather than to establish definitive statistical associations. In analyses that are discovery-oriented, correction for multiple testing, while essential in confirmatory contexts, is not mandatory and may even obscure meaningful trends by inflating type II error rates. Our objective was therefore to highlight consistent directional patterns across analytical layers, to guide future confirmatory work rather than to make categorical claims.

We also note that this comment somewhat contrasts with the earlier suggestion to discuss trends that are not statistically significant.

With regard to the proposal to verify our observations using an external dataset, we are in full agreement that independent confirmation would be beneficial. However, as reviewer 1 rightly emphasized, the generation of such datasets from sorted human thymocyte subsets is “heroic” and has rarely, if ever, been achieved. We are aware of no existing dataset that provides comparable material or analytical depth.

The available single-cell thymic dataset (Park et al., Science 2020) includes only a few hundred sequences per donor, which is significantly less than the number of sequences in our study. This limited dataset is not adequate for cross-validation or for representing the full complexity of thymic TCR repertoires.

As with the pediatric thymus dataset, the lack of statistical power in the dataset due to the small number of female subjects (only three) means that sex-related differences in V/J usage cannot be evaluated.

Finally, the peripheral blood dataset is not appropriate for validating thymic generation or selection processes, as it reflects post-thymic selection and antigen-driven remodeling, making it impossible to distinguish peripheral effects from thymic influences.

For these reasons, none of the currently available datasets provides a sufficiently clean or powerful framework to test the reproducibility of subtle sex-associated effects on thymic TCR repertoires. Nevertheless, we fully agree that confirmation in an independent and larger cohort will be an important next step to refine these exploratory findings and assess their generalizability to a broader human population.

Supplementary Figure 6 suggests that women consistently show higher Rényi entropies across all subsets. Although individual p-values are borderline, the consistent direction of change is notable. The authors should apply an integrated statistical test across subsets (for example, a mixed-effects model) to determine whether there is an overall significant trend toward higher diversity in females.

We agree that Rényi entropies tend to show a consistent direction of change across subsets, with slightly higher values observed in females. In this section, our objective was to provide a descriptive overview of diversity patterns for each thymic subset. This is because these subsets are biologically distinct and therefore require individual analysis, as we previously demonstrated using the same dataset (Isacchini et al, PRX Life. 2024). Therefore, while a mixed-effects approach could in principle be applied to test for an overall trend, such an analysis would rely on the assumption of a common sex effect across heterogeneous cell types.

It is important to note that the complete dataset has now been made publicly available, enabling interested researchers to perform additional integrative or model-based analyses to further explore these diversity trends.

Figures 4B and S8 clearly indicate enrichment of hydrophobic residues in female CDR3s for both TRA and TRB (excluding alanine, which is not strongly hydrophobic). Because CDR3 hydrophobicity has been linked to increased cross-reactivity and self-reactivity (see, e.g., Stadinski et al., Nat Immunol 2016), this observation is biologically meaningful and consistent with higher autoimmune susceptibility in females.

We thank the reviewer for this insightful comment.

As correctly noted, increased hydrophobicity at specific CDR3β positions has been linked to enhanced cross-reactivity and self-reactivity, as described by Stadinski et al. (Nat Immunol 2016), and we reference this work in the manuscript.

In our analysis corresponding to Figure 4B (TRB), hydrophobicity was quantified at the sequence level by computing, for each unique CDR3β sequence, the overall proportion of hydrophobic amino acids across the CDR3 loop. This approach aligns with that of Lagattuta et al. (Nat Immunol 2022), whose code we adapted to accommodate longer CDR3s. This global hydrophobicity metric captures overall composition, but, by its construction, does not account for positional context, the key mechanism implicated by Stadinski et al.

As outlined in our original Figure 4C, the results were obtained through a position-based amino acid analysis. For each CDR3β sequence, we extracted the amino acid at every IMGTdefined CDR3 position (p104–p118) and quantified, at each position, the percentage of unique sequences containing each amino acid. Positions p109 and p110 correspond to the p6–p7 sites highlighted by Stadinski et al. as functionally relevant for self-reactivity. This analysis evaluates positional composition independently of clonotype frequency, focusing specifically on hydrophobic amino acid classes.

Following the recommendation of the reviewer, the revised manuscript has removed alanine (which is only weakly hydrophobic) has been excluded from the hydrophobic residue set. With this refined definition, we observe a significant enrichment of hydrophobic amino acids at p109 in CD8 T cell repertoires from females, with similar but non-significant trends at p109 in DP and CD4 Teff cells and at p110 in CD8 cells (see new Figure 4C).

As outlined in the revised Methods, Results, and Discussion sections, Figure 4C focuses exclusively on positional hydrophobic amino acid usage. This was previously implicit, although it was noted in the legend and visually represented in the plots.

The majority of "hundreds of sex-specific motifs" are probably donor-specific motifs confounded by HLA restriction. This interpretation is supported by the failure to validate motifs in external datasets (pediatric thymus, peripheral blood). The authors should restrict analysis to public motifs (shared across multiple donors) and report the number of donors contributing to each motif.

We fully agree that donor-specific and HLA-restricted motifs represent a major potential confounder in repertoire-level comparisons. To minimize this potential bias, our analysis was explicitly restricted to public motifs, as clearly stated in the Materials and Methods section:

“Additional filters were applied so that: (i) a motif includes public CDR3aa sequences (shared by at least two individuals); (ii) a significant enrichment is detected (Fisher’s exact test, p < 0.01); and (iii) a usage difference between groups of at least twofold (Wilcoxon test, p < 0.05).”

Accordingly, every motif reported in the manuscript is supported by at least two independent donors, ensuring that no motif reflects an individual- or HLA-specific effect (see Supplementary Figures 10-13[previously Supplementary Figure 9]). We have now added a more explicit mention of the number of donors contributing to each motif in the figure legend and have clarified this point in the revised Methods and Results sections to make this criterion more visible to readers.

When comparing TCRs to VDJdb or other databases, it is critical to consider HLA restriction. Only database matches corresponding to epitopes that can be presented by the donor's HLA should be counted. The authors must either perform HLA typing or explicitly discuss this limitation and how it affects their conclusions.

We respectfully disagree with the assertion that HLA typing is necessary for the type of comparative analysis we have conducted. While it is true that HLA molecules present peptides to TCRs and thereby contribute to the tripartite interaction determining T cell activation, extensive evidence indicates that the CDR3 region, particularly CDR3β, is the dominant determinant of antigen specificity. This finding is supported by structural and computational studies (Madi et al., eLife, 2017; Huang et al., Nat. Biotech., 2020; MayerBlackwell et al., Methods Mol. Biol., 2022) showing that CDR3β residues are responsible for the majority of peptide contacts, whereas CDR1 and CDR2 primarily interact with the MHC framework.

As emphasized in several recent benchmarking studies (e.g., Springer et al., Front Immunol, 2021), CDR3β sequence composition alone captures most of the information required for specificity inference. Consequently, widely used and validated computational tools such as GIANA (Zhang et al. Nat. Commun. 2021), iSMART (Zhang et al. Clin. Cancer Res. 2020), and ATMTCR (Cai et al. Front. Immunol. 2022) rely exclusively on CDR3β aminoacid sequences and still achieve high predictive performance.

Our analysis aligns with this well-established paradigm. While we agree that integrating donor HLA typing would refine epitope-level annotation and reduce potential noise, the absence of HLA data does not invalidate the comparative framework we used, which focuses on relative representation of annotated specificities across groups rather than on individual TCR–HLA–peptide triads.

Although the age distributions of male and female donors are similar, the key question is whether HLA alleles are similarly distributed. If women in the cohort happen to carry autoimmuneassociated alleles more often, this alone could explain observed repertoire differences. HLA typing and HLA comparison between sexes are therefore essential.

To address the issue of any potential differences in HLA background, we examined the subset of adult donors for whom HLA typing information was available (HLA-A, HLA-B, HLADR, and HLA-DQB; n = 16). Within this subset, the distribution of HLA alleles was relatively balanced between males and females (as illustrated by the heatmap showing HLA class II expression patterns and HLA class I family grouping in Author response image 1). This analysis suggests that the sex-associated differences in the repertoire observed in our study are unlikely to be driven solely by unequal representation of autoimmune-associated HLA alleles.

We acknowledge, however, that complete HLA information was not available for all donors, which remains a limitation of the dataset.

Author response image 1.

In some analyses (e.g., Figures 8C-D) data are shown per donor, while others (e.g., Fig. 8A-B) pool all sequences. This inconsistency is concerning. The apparent enrichment of autoimmune or bacterial specificities in females could be driven by one or two donors with particular HLAs. All analyses should display donor-level values, not pooled data.

While Figures 8A–B present pooled data to summarize global trends, the corresponding donor-level analyses were provided in Supplementary Figures 15B and 16 (previously Supplementary Figures 11B and 12). In these, each individual is shown separately, with each point representing an individual. It is important to note that these donor-resolved plots do not reveal any sample-specific driver: the patterns observed in the pooled data remain consistent across donors, without any single individual accounting for the apparent enrichments. As outlined in the revised manuscript, readers now directed to the relevant supplementary figures for further clarification.

The reported enrichment of matches to certain specificities relative to the database composition is conceptually problematic. Because the reference database has an arbitrary distribution of epitopes, enrichment relative to it lacks biological meaning. HLA distribution in the studied patients and HLA restrictions of antigens in the database could be completely different, which could alone explain enrichment and depletions for particular specificities. Moreover, differences in Pgen distributions across epitopes can produce apparent enrichment artifacts. Exact matches typically correspond to high-Pgen "public" sequences; thus, the enrichment analysis may simply reflect variation in Pgen of specific TCRs (i.e., fraction of high-Pgen TCRs) across epitopes rather than true selection. Consequently, statements such as "We observed a significant enrichment of unique TRB CDR3aa sequences specific to self-antigens" should be removed.

We respectfully disagree with the conclusion that our enrichment analysis lacks biological meaning. Our approach directly involves a direct comparison of the same set of observed TCR sequences between males and females. Consequently, any potential biases related to generation probability (Pgen), which affect all sequences equally, cannot account for the observed sex-specific differences. To summarize, because the comparison is performed on the same set of sequences, changes in the probability of generation across epitopes cannot explain the differences seen between the sexes.

We do agree, however, that the composition of the reference databases may influence apparent enrichment patterns, as these resources contain uneven distributions of epitope categories and often incomplete information regarding HLA restriction. It should be noted that this limitation is inherent to all database-based annotation approaches, a fact which is explicitly acknowledged in the revised Discussion.

The overrepresentation of self-specific TCRs in females is the manuscript's most interesting finding, yet it is not described in detail. The authors should list the corresponding self-antigens, indicate which autoimmune diseases they relate to, and show per-donor distributions of these matches.

We thank the reviewer for this constructive suggestion.

As recommended, we have expanded the description of the self-specific TCRs identified in our dataset and now provide this information in Supplementary Table 2 of the revised manuscript. Specifically, the table lists the corresponding self-antigens and the autoimmune diseases with which they are associated. In our curated database, these annotations primarily correspond to celiac disease and type 1 diabetes, which were the two autoimmune contexts explicitly defined in the manually curated reference datasets.

For the “cancer” specificity group, we have clarified that antigen assignments were established based on (i) annotations available in the original databases (IEDB, VDJdb, McPAS-TCR) and (ii) cross-referencing with additional resources, including the Human Protein Atlas, the Cancer Antigenic Peptide Database (de Duve Institute), and the Cancer Antigen Atlas (Yi et al., iScience 2021), to ensure consistency in the classification of cancer and neoantigen specificities. Please refer to the Materials and Methods section for a full description of the procedure for this specific assignment.

Donor-level distributions of these self-specific matches are now shown in Supplementary Figures 15B and 16 (previously Supplemental Figures 11B and 12), allowing direct visualization of inter-donor variability. Importantly, these plots confirm that the observed enrichment in females is not driven by a single individual, further supporting the robustness of the finding.

The concept of poly-specificity is controversial. The authors should clearly explain how polyspecific TCRs were defined in this study and highlight that the experimental evidence supporting true polyspecificity is very limited (e.g., just a single TCR from Figure 5 from Quiniou et al.).

We certainly agree (and regret) that the concept of TCR polyspecificity remains a subject of debate and often underappreciated in the field of immunology. As Don Mason famously discussed in his seminal essay “A very high cross-reactivity is an essential feature of the TCR” (doi: 10.1016/S0167-5699(98)01299-7) published over 25 years ago, both theoretical and experimental evidence indicates that each TCR can, in principle, recognize millions of distinct peptides, albeit with variable avidity.

Although this principle is widely accepted, it is frequently overlooked in the field of experimental immunology. In this area, anything that deviates from strict monospecificity is often disregarded as noise.

In our own analyses of large-scale TCR repertoires, we have repeatedly observed that many CDR3 sequences are annotated with multiple specificities across different databases, often corresponding to peptides from unrelated organisms. As demonstrated in Quiniou et al. (eLife 2023), such polyreactive TCRs exhibit distinctive features, including biased physicochemical composition, and tend to be enriched in various biological contexts. In our preliminary study of such TCRs, which have the capacity to be specific for multiple viral- and self- epitopes, we hypothesized that they may serve as a first line of defense against pathogens and also be involved in triggering autoimmunity. We therefore consider it important to report this phenomenon rather than omit it, especially given its potential relevance to both protective immunity and autoimmunity.

In the present study, polyspecific TCRs were defined operationally as TRB CDR3aa sequences associated with a minimum of two distinct specificity groups, corresponding either to different microbial species or to multiple antigen categories within the curated database. Therefore, our definition captures broader antigenic groupings rather than epitope-level binding events.

We fully acknowledge that direct experimental evidence for true molecular-level polyspecificity remains limited. Indeed, as the reviewer notes, only a single TCR with multiepitope reactivity has been rigorously demonstrated to date (Quiniou et al.2023). Consequently, our analysis does not make claims about structural promiscuity; instead, it uses database-annotated cross-reactivity as a proxy to explore broader repertoire-level patterns.

As outlined in the Methods section, this definition has been clarified and its discussion expanded in the Discussion to explicitly address these conceptual and methodological nuances.

Minor:

Clarify why the Pgen model was used only for DP and CD8 subsets and not for others.

As noted, computing Pgen values involves two steps: (i) training a generative model of V(D)J recombination using IGoR, and (ii) estimating generation probabilities with OLGA based on that model. Both steps require a significant amount of computing power, especially when applied to large repertoires across multiple subsets. For this reason, we focused the analysis on DP thymocytes, which represent the repertoire prior to thymic selection, and CD8 T cells after CD8 selection.

The Methods section should define what a "high sequence reliability score" is and describe precisely how the "harmonized" database was constructed.

Briefly, the annotated database used in this study was constructed in accordance with the procedure established in our previously published work (Jouannet et al., NAR Genomics and Bioinformatics, 2025). The study integrates three publicly available resources, IEDB, VDJdb, and McPAS-TCR, which were collected as of October 2023. These three datasets were then merged into a single harmonized compendium, undergoing extensive standardization. When entries shared identical information across databases (same V–CDR3–J for both TRA and TRB, same epitope, organism, PubMed ID, and cell subset), only one representative was kept; discrepant or incomplete entries were retained to preserve information. We then assigned a sequence reliability score, the Verified Score (VS), following the verification strategy used by IEDB. The scale ranges from 0 to 2 and reflects the concordance between calculated and curated TRA/TRB CDR3 sequences (2 = both TRA and TRB present are verified, 1.1 = only TRA verified, 1.2 = only TRB verified, 0 = no verified chain). A second score, the Antigen Identification Score (AIS), is used to rank antigen-identification methods on a scale of 0 to 5, according to the strength of the experimental evidence supporting them.

In the present study, “high reliability” refers to sequences with a verified TRB CDR3aa chain (VS ≥ 1.2) and an AIS score corresponding to T cells in vitro stimulation with a pathogen, protein or peptide, or pMHC X-mer sorting (> 3.2, excluding categories 4.1 and 4.2), ensuring that downstream analyses were performed on a rigorously curated and biologically trustworthy dataset. The Methods section now explicitly details these criteria.

The statement "we generated 20,000 permuted mixed-sex groups" is unclear. It is not evident how this permutation corrects for individual variation or sex bias. A more appropriate approach would be to train the Pgen model separately for each individual's nonproductive sequences (if the number of sequences is large enough).

The objective of this analysis was to determine whether the enrichment of TRBV06-5 in females was due to random grouping of individuals or whether it was attributable to sex itself. To do so, we generated all possible perfectly mixed groups of donors (i.e., groups containing an equal number of male and female donors) for the concerned thymocyte subset, and then performed 20,000 random pairwise comparisons between such mixed groups. For each comparison, we tested the TRBV06-5 usage between the two mixed groups. This procedure directly evaluates whether group composition (independent of sex) could spuriously generate differences in TRBV usage. Notably, none of these 20,000 comparisons between the two mixed groups yielded a statistically significant difference in TRBV06-5 usage. In contrast, when comparing the true male and female groups, a significant difference was identified. This demonstrates that the signal we observe is not driven by random donor grouping or individual-level variation, but is specifically associated with sex. It is important to note that this analysis, which is designed to exclude spurious group effects, is rarely performed in published repertoire studies, yet it provides an important internal control for robustness.

Reviewer #2 (Recommendations for the authors):

(1) Data availability "upon request" is unacceptable. All raw and processed data, as well as scripts used for analysis and figure generation, must be publicly deposited before publication.

We would like to clarify that our intention has always been to make this dataset publicly available. It was a mistake to suggest otherwise in the original submission.

(2) At the beginning of the Results section, include a brief description of the dataset: number of donors, sex ratio, age range, number of samples per subset, and sorting strategy. Although Figure 1 shows this, the information should also be mentioned in the main text.

In line with the recommendation, we have now added a summary of the cohort characteristics at the beginning of the Results section. This includes the number of donors, sex ratio, age range, number of samples per subset, and the sorting strategy used. While this information was already included in Figure 1, we concur that including it directly in the main text enhances readability.

(3) Report the number of cells and unique clonotypes analyzed per individual. Rank-frequency plots (in log-log coordinates) would be helpful.

We have now added, for each donor and each subset, the number of cells, and additionally for each chain, the number of total and unique clonotypes analyzed. This information is provided in the revised manuscript in a new supplementary table (Supplemental Table 1).

These plots have been integrated into the revised manuscript as Supplementary Figure 2.

(4) For analysis in Figure 4B, the total fraction of hydrophobic amino acids should be calculated for each patient separately, and values for men and women should be compared (analogously to Figure 4C, but for the whole CDR3 and excluding alanine).

Please note that the TRB CDR3aa composition in Figure 4B has already been quantified at the individual level. For each unique TRB CDR3aa sequence, we computed the proportion of each of the 20 amino acids across the CDR3β loop, then summarized these values per donor (mean per individual). The log2 fold change displayed in Figure 4B (and supplemental Figure 9 for TRA) is calculated from the median donor-level values for females versus males, rather than from pooled CDR3s. It is intended as descriptive, “global” view of amino acid usage within the central CDR3 region. Hydrophobicity was not used directly in the computation, but is indicated only by bar color, based on the Kyte-Doolittle- derived IMGT classification. This provides an observational overview of amino acid composition in the central CDR3 region.

As the mechanistic link between hydrophobicity and self-reactivity described by Stadinski et al. is explicitly position-dependent, we consider positional analyses to be the most appropriate method for formally interrogating this hypothesis, as we did in Figure 4C. Here, our primary focus was on the position-specific usage of hydrophobic amino acids at IMGT positions p109-p110. These positions correspond to the central p6-p7 positions described by Stadinski et al. For each individual, we computed the proportion of unique TRB CDR3aa sequences carrying a hydrophobic amino acid at a given position.

Accordingly, in the revised manuscript we refined the Figure 4C by excluding alanine due to its weak hydrophobic property (as recommended by the reviewer) This positional composition analysis now reveals a statistically significant increase in hydrophobic usage at p109 in female CD8 repertoires, with similar, though non-significant, trends at p109 in DP and CD4Teff ad at p110 in CD8 cells. Figure 4B is therefore retained as an exploratory overview of amino acid composition usage along the CDR3 loop, while Figure 4C is used for the more specific question of hydrophobicity and potential cross-reactivity.

The Methods section has been expanded to provide clearer descriptions of these computations, and the Results and Discussion sections corresponding to Figures 4B-C (and supplemental Figure 9) have been revised to make the rationale, implementation, and interpretation of these hydrophobicity analyses more explicit.

(5) Figure 6 shows a trend toward higher clustering of Treg TCRs in males, which could relate to the lower incidence of autoimmunity in men. The authors could test whether specific Treg clusters are male-specific and shared among male donors.

As shown in Figure 6, a clear trend towards higher similarity among Treg CDR3aa sequences in males is evident, as indicated by the proportion of sequences included in clusters and in the overall similarity density. However, identifying “male-specific clusters” shared across donors is not straightforward in our analytical framework.

In our approach, for each cell subset, CDR3aa sequences were downsampled 100 times to the smallest sample size, and clustering was repeated at each iteration. Therefore, the clusters’ identities are not consistent across iterations. The clusters depend on the specific subset of sequences selected at each downsampling step, as well as on their underlying Pgen distribution. Therefore, it is not possible to reliably assess whether specific clusters are systematically “male-shared”. This is because cluster composition is a function of stochastic resampling rather than of biological structure. For this reason, a comparison of cluster identities across donors would not produce interpretable results.