Quantifying changes in the T cell receptor repertoire during thymic development

  1. Francesco Camaglia
  2. Arie Ryvkin
  3. Erez Greenstein
  4. Shlomit Reich-Zeliger
  5. Benny Chain
  6. Thierry Mora
  7. Aleksandra M Walczak
  8. Nir Friedman

  • Curated by eLife

    eLife logo

    eLife assessment

    This paper addresses an important question within adaptive immunity, namely whether the T cell receptor (TCR) repertoire of negatively selected thymocytes shares common features. The authors analyze T cell receptor sequences from mice as they progress through positive selection, CD4/CD8 lineage commitment, and negative selection, to find small but consistent differences between the repertoires at these selection stages. They argue that their findings do not indicate any sequence-specific selection; however, some of the conclusions drawn are currently incompletely supported by the performed analyses.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

One of the feats of adaptive immunity is its ability to recognize foreign pathogens while sparing the self. During maturation in the thymus, T cells are selected through the binding properties of their antigen-specific T-cell receptor (TCR), through the elimination of both weakly (positive selection) and strongly (negative selection) self-reactive receptors. However, the impact of thymic selection on the TCR repertoire is poorly understood. Here, we use transgenic Nur77-mice expressing a T-cell activation reporter to study the repertoires of thymic T cells at various stages of their development, including cells that do not pass selection. We combine high-throughput repertoire sequencing with statistical inference techniques to characterize the selection of the TCR in these distinct subsets. We find small but significant differences in the TCR repertoire parameters between the maturation stages, which recapitulate known differentiation pathways leading to the CD4 + and CD8 + subtypes. These differences can be simulated by simple models of selection acting linearly on the sequence features. We find no evidence of specific sequences or sequence motifs or features that are suppressed by negative selection. These results favour a collective or statistical model for T-cell self non-self discrimination, where negative selection biases the repertoire away from self recognition, rather than ensuring lack of self-reactivity at the single-cell level.

Article activity feed

  1. Version published to 10.7554/elife.81622 on eLife
  2. eLife

    eLife assessment

    This paper addresses an important question within adaptive immunity, namely whether the T cell receptor (TCR) repertoire of negatively selected thymocytes shares common features. The authors analyze T cell receptor sequences from mice as they progress through positive selection, CD4/CD8 lineage commitment, and negative selection, to find small but consistent differences between the repertoires at these selection stages. They argue that their findings do not indicate any sequence-specific selection; however, some of the conclusions drawn are currently incompletely supported by the performed analyses.

  3. eLife

    Reviewer #1 (Public Review):

    In this manuscript, the authors analyze the impact of thymic selection on the TCR repertoire in Nur77-mice by studying the repertoire at different developmental stages using high-throughput sequencing. In combination with different statistical methods and analytical approaches, they searched for specific TCR patterns that could be characteristic for different stages of T cell differentiation. Based on their methods and analyses, they found that there was no evidence for a selection of specific sequences at different stages of development, proposing that negative selection to avoid self-recognition is mainly performed on the collective level rather than at the single-cell level.

    The authors use a very interesting and reasonable set of analytical approaches to compare the TCR repertoires at different stages of development. The performed analyses lead to the conclusion that there is no specific pattern of sequences or sequence motifs that are suppressed by negative selection. Their comparisons are valid, but as the authors already point out in their discussion, there could be some aspects that could mask the ability to detect characteristic signatures of TCR repertoires with regard to developmental stage. This includes the separate analyses of alpha- and beta-chain repertoires without considering their combination, as well as the selected experimental system that could affect identification of clearly non-selected cell populations, but also potentially the pooling of the read outs from several mice that could mask signatures on an individual level. The authors provide reasonable arguments for the performed approaches, but some additional analyses might be helpful to corroborate the claims put forward within this manuscript.

  4. eLife

    Reviewer #2 (Public Review):

    This is an original and carefully argued study on a key question of immunology. The authors detect statistical differences in the TRA and TRB repertoires between negatively selected thymocytes and mature T cells. Discrimination does not work for individual T cell receptor chains, but starts to become reasonably sensitive and specific for quora of 30 alpha chains (I did not find ROCs for beta chains; see below). These results, including the more detailed statistics based on CDR3 sequence, are technically sound and make a unique conceptual contribution to quantitative immunology.

    In terms of interpretation, the premise of the paper - that negatively selected and peripheral T cell repertoires should systematically differ in some characteristics because thymocytes could scan only a tiny fraction of self-peptides - is not based on experimental evidence. Experimental data allow for the possibility that a thymocyte scans a much larger fraction of self-peptides than the number given by the authors. Hence this point cannot be maintained as a premise, while the underlying question is key and worth discussing. In this context, I also recommend that the title give a factual account of the main finding, rather than propose a particular hypothetical interpretation (hypotheses will be better placed in the Discussion, possibly the Abstract). These suggested edits do not impact the originality and importance of the experiments and computational results; these will be of wide interest.

  5. eLife

    Reviewer #3 (Public Review):

    The authors aimed to quantify changes in the (CDR3beta) T cell receptor (TCR) repertoire as the cells go through the successive stages of thymic selection. To this end, they used Nur77 reporter mice and Annexin V to detect activated and/or dying cells, allowing them to some extent to identify cells that had undergone positive and/or negative selection. The authors appear to set out to prove the absence of major sequence-specific differences between these repertoires to support a stochastic model of thymic selection, in which T cells experience mild sequence-specific biases rather than being strongly pushed towards a specific fate. Indeed, since the ground-breaking results by Davis et al (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4455602/), such a stochastic model is now commonly assumed rather than the older "text-book view" of thymic selection removing most or all auto-reactive cells; as such, it is a reasonable starting point.

    The dataset generated for this paper is very interesting and will no doubt be useful to the wider community. To my knowledge, this is the first time this combination of Nur77 and Annexin V was used to aim to pinpoint cells that were deleted. The authors use state-of-the-art generative statistical models for TCR repertoires to conduct their analyses. The initial analyses shown in Figure 2 are promising in that they indicate that there are indeed visible systematic differences between these subsets, even if they might be small.

    A limitation of the Annexin V based approach is that the fraction of cells expressing Annexin V is small, and there appears to be no clear "cutoff" value separating negative from positive cells. This means that the negatively selected subpopulations, probably the most interesting ones for this study, are also the smallest at 1000 cells or less per sample. This limits the ability to detect specific "signatures" of detection. Indeed, from the initial analyses (Figure 2), it appears that the difference between Annexin V+ and Annexin V- populations is just barely detectable. Unfortunately, this means that not too much can be concluded from the absence of clear signals when comparing these subpopulations, as there may simply not be enough statistical power. This would make it very important for the authors to state clearly which signals they can and cannot expect to detect in datasets of this size. For instance, it may well be that some of the TCRs that are specific for a small number of ubiquitously expressed proteins (such as beta-Actin) are reliably removed during negative selection, but these TCRs may be a small minority of the overall pool, and they may not share common sequence features as there would presumably be many different peptides that they could respond to. As such, this kind of sequence-specific selection would likely go undetected by the analyses shown in this paper.

    The authors show in Figure 3 that while individual TCRs coming from the different populations cannot be distinguished reliably, we can still distinguish these populations if we instead look at larger groups of TCRs. The authors interpret this as evidence for the idea that T cells collectively distinguish self from nonself by quorum sensing. However, the fact that several noisy predictions of a class can be combined to obtain a better prediction is not specifically related to TCR sequences, and a similar phenomenon would appear in any classification task (in machine learning, this phenomenon is known as "boosting"). It is a consequence of the law of large numbers -- an average taken from several values (TCR sequence predictions in this case) will be closer to the true population average than one taken from few values. Thus, as soon as there is *any* difference between the mean predicted class probabilities for the two classes, then this phenomenon will occur.
    The authors do not clearly explain how this basic fact substantiates the idea of quorum sensing, which is a phenomenon involving several T cells that are specific to the same antigen.

    In Figure 4, the authors show that there are differences in amino acid usage between the populations (further detailed in supplementary figures) and that similarity in amino acid usage corresponds to closeness in the lineage. This is an interesting observation, which raises the question whether it is really necessary to look at 3-mers to get this result or whether simple 1-mers (i.e., simply the usage of amino acids without considering contiguity of positions) would already be sufficient. Several results show that differences do exist at the 1-mer level already, so it remains unclear whether going to k-mers is really necessary.

    In Figure 5, the authors argue that the data are inconsistent with a model in which two fates for the same T cell receptor are mutually exclusive (or at least sufficiently strongly biased towards mutual exclusion), as they would expect a negative correlation between the class probabilities of these two fates in this case. However, the scenario shown in Figure 5A is not comparable to the data. For example, even if the CD4SP and CD8SP fates were mutually exclusive, we might still not expect a negative correlations between the quantities E_CD4SP-E_DPPRE and E_CD8SP-E_DPPRE because the cells need to reach the DPPOS stage first. Therefore, E_DPPOS would be a common cause of E_CD4SP and E_CD8SP, inducing a positive correlation which may well be stronger than the expected negative correlation.

    Overall, this is a relevant paper based on an interesting dataset and sophisticated methodology. However, I was not convinced of some of the authors' conclusions due to the aforementioned issues in the methodology. Generally speaking, the paper is also still rather difficult to parse since it is not always clear what exactly the authors are trying to achieve with their quite sophisticated analyses, and simpler baselines are not considered to show that these complex analyses are truly necessary; certainly for the analyses shown in Figure 3B and Figure 5A, it was not entirely clear why these were performed and what we might conclude from them. Therefore, in its current state, I worry that the paper might not yet be very accessible to the broader community and that the motivation behind its methodology might remain somewhat obscure to many readers.

  6. Version published to 10.1101/2022.02.14.480309v1 on bioRxiv