Diphenylcyclopropenone (DPC) is an example of a hapten, a small organic compound that reacts with biological macromolecules including proteins to form immunogenic conjugates. It is a potent skin sensitizer (Mose et al., 2017a; Stute et al., 1981) and, since it is non-mutagenic in the AMES assay (Wilkerson et al., 1987), it has been used as an immunostimulant in the treatment of warts (Buckley et al., 1999), melanoma (Read et al., 2017), and alopecia (Ashworth et al., 1989; Karanovic et al., 2018; Lee et al., 2018). DPC can also act as an adjuvant to conventional immunization (von Moos et al., 2012). Because of its clinical applications, and because it is not commonly found in the environment, DPC serves as an interesting model to study primary and secondary responses to chemical haptens in humans, although the confounding effects of any underlying clinical condition in the treated individuals must obviously be considered. Clinically, repeated exposure results in rapid sensitization, but the magnitude of the resulting allergic contact dermatitis (ACD) does not continue to increase during repeated exposure (Mose et al., 2017a), suggesting the existence of regulatory processes. Global expression profiling at the site of elicitation suggested a predominantly Th1/Th17 type T cell infiltration, whose dynamics mirrored the clinical and histopathological picture (Mose et al., 2017b). High serum interleukin-4 (IL4) and low interleukin-12 (IL12) were observed in patients with alopecia showing a favorable response to DPC treatment.

As part of an extended project to understand more fully the immunological events leading to ACD, and hence generate better tools for chemical risk assessment (Kimber et al., 2012), we have measured the T cell receptor repertoire (TCRrep) changes in ex vivo blood samples collected from patients before and after therapeutic sensitization with DPC. In a previous study, we used a newly developed quantitative, robust experimental and computational pipeline (Oakes et al., 2017a) to measure changes in the TCRrep following in vitro re-stimulation of blood from individuals allergic to the environmental contact allergen paraphenylene diamine (Oakes et al., 2017b). A limitation of the study was that the extent of allergen exposure prior to challenge was unknown. Here we extend these studies to document in vivo changes to the TCRrep during primary and secondary sensitization with DPC in a small cohort of patients with alopecia. Exposure to DPC results in transient expansions of a set of related TCR sequences in the peripheral blood. The number of responding TCRs correlates with the extent of sensitization and is not observed in unexposed individuals. The results suggest that exposure to a contact allergen in vivo stimulates an initial polyclonal expansion of antigen-specific responding T cells which, however, does not increase on further repeated exposure to antigen.

We carried out TCRseq on unfractionated blood samples from 29 patients with alopecia (Table 1).

Samples were collected before (referred to as the pre-sensitization [PS] sample, week 0, n = 25 samples) and at one to three time points after sensitization with DPC (Figure 1A), taken at 2 weeks (PT1, n = 23), 6 weeks (PT2, n = 18), and approximately 24 weeks (PT3, n = 17) post-sensitization. As controls, we carried out TCRseq on five healthy volunteers, for which we had time points that matched the 0-, 2-, and 6-week time points of the DPC sensitized group. We obtained a total of 24,431,855 distinct TCR sequences, of which 17,451,853 were unique across all samples. The number of TCR sequences obtained varied between samples (mainly due to small differences in volume of individual libraries when preparing the pooled library for sequencing) and are shown in Supplementary file 1. The median number of TCRs per sample was 95,638 (range 23,224–690,600) and the median number of unique TCRs per sample was 75,111 (range 19,095–342,336). Analyses included every patient for whom the relevant time points (and patch test data where relevant) were available, and this number is indicated in the respective figure legends.

Figure 1

Download asset Open asset

Sensitization with DPC induces a transient expansion in the frequency of a small subset of the TCR repertoire

We next looked for evidence of changes in individual TCR sequences following exposure to DPC. We plotted the abundance of each TCR before sensitization and after sensitization (at PT1) (representative examples are in Figure 2A; all individuals in Supplementary file 3). We observed that in a number of individuals, there was a population of TCRs which were absent before sensitization and present at relatively high abundance after sensitization (indicated by the pink dots in the left panels of Figure 2A). This change is seen clearly in the abundance profile of all TCRs that were absent in the pre-sensitization sample, but present after sensitization (Figure 2B). On the basis of these profiles, we counted the proportion of TCRs absent at PS and present at PT1 with an abundance of eight times or above. The remainder of this study focuses on this population, which we refer to as PT1 expanded. The percentages of PT1 TCRs which are expanded in each individual are summarized for patients and healthy volunteers in Figure 2C. For comparison, we calculated the percentage of TCRs absent at PT1 but present at PS with an abundance of eight times or above (referred to as PS expanded). The range of percentages of expanded TCRs at PT1 is clearly much greater in the patients than in the healthy volunteers (between 0 and 2 weeks), although the medians are not significantly different (p=0.16 and p=0.47 for alpha and beta respectively, Mann–Whitney). The percentage of expanded TCRs at PT1 in the patients is significantly larger than the percentage of TCRs which were expanded at PS and absent at PT1 (p=0.02 for alpha, p=0.03 for beta, Wilcoxon signed–rank). This is not the case for the healthy volunteers (p=0.79 for alpha, p=0.59 for beta, Wilcoxon signed–rank). Similar results were obtained setting the expansion threshold at 16 or 32. The number of PT1 expanded TCR alpha and TCR beta sequences was highly correlated (Spearman’s rho = 0.85, p<0.0001, Supplementary file 3).

Figure 2

Download asset Open asset

After the initial increase post-sensitization, the number of expanded TCRs remained rather constant over the later time points. There was no evidence that the total number of expanded TCRs increased with time, despite the fact that individuals were exposed to repeated therapeutic stimulation with sensitizer on a weekly basis for the period of the study in addition to the patch test applications (Figure 1A) (Kruskal–Wallis rank sum test, p=0.92 alpha and p=0.36 beta). However, this apparent overall stability hid dramatic dynamic changes in the frequency of individual TCRs within individual patients. The majority of PT1 (2 weeks) expanded TCRs decreased or returned to baseline at subsequent time points (Figure 3A). The majority of PT2 expanded TCRs present at 6 weeks were not expanded at 2 weeks, peaked at 6 weeks, and decreased or returned to baseline by the time the PT3 sample was taken at around 24 weeks (Figure 3B). These dynamics are shown for the beta chain sequences of 10 patients for whom all four time points were available (the corresponding analysis for five healthy volunteers is in Figure 3C). The alpha chain sequences of the same patients behaved similarly (Supplementary file 4). The exposure to repeated doses of DPC therefore induced large, but predominantly transient, changes in the frequencies of a small proportion of the TCRs.

Figure 3

Download asset Open asset

TCR expansion after exposure to DPC correlates with the magnitude of skin sensitization

We hypothesized that the expanded TCRs identified above might be functionally related to development of ACD. Exposure to DPC induced skin sensitization (patch test scores of + or higher) in almost all individuals but the magnitude of the response varied significantly between individuals (Figure 4A). Interestingly, the maximum response was usually observed at the first patch test, and declined or remained constant thereafter despite repeated exposure (Figure 4B). The sensitization reaction, therefore, like the number of expanded TCRs, did not continue to increase despite repeated exposure to antigen. In consequence, the magnitude of the response was very diverse at the beginning (+++, ++, or +) but converged toward a score of + or below by week 24.

Figure 4

Download asset Open asset

We plotted the number of PT1 expanded TCRs (abundance ≥8) in individuals with varying strengths of sensitization as a function of the initial skin response recorded at the site of application (Figure 4C, left panels) or at the first patch test (Figure 4C, center panels). The number of PT1 expanded TCRs was positively correlated to the patch test score. The correlation was positive but did not reach significance at an expansion threshold of 4, but remained significant when using a threshold abundance of 16 or 32. In contrast, there was no significant association between the number of PS expanded TCRs and the PT1 patch test score (Figure 4C, right panels). Although we measured the proportion of expanded TCRs, and not the absolute number, it remained possible that the total number of TCRs in each sample could affect the percentage of expanded TCRs. However, there was no significant correlation between the proportion of PT1 expanded TCRs (abundance ≥8) and the total number of TCRs sequenced in each sample (Spearman’s correlation rho = 0.18 [p=0.4] and rho = 0.22 [p=0.3] for alpha and beta respectively). We also carried out repeated (10 times) subsampling of each sample to the size of the smallest sample. We observed a mean positive correlation (Spearman’s rho) between the patch test scores and the percentage of PT1 expanded TCRs in the subsamples at a threshold of 8 (0.4 [alpha] and 0.3 [beta]), but the p-value did not reach significance (p=0.18 and 0.22 respectively). At a threshold of 16, the correlations were 0.54 (p=0.04) and 0.32 (p=0.2), and at a threshold of 32, the correlations were 0.58 (p=0.01) and 0.61 (p=0.006) for alpha and beta respectively. Overall, the qualitative pattern of greater number of PT1 expanded TCRs remained the same after subsampling although the smaller data sets did alter the magnitude of the correlations and the threshold at which they became significant.

The correlation between the number of expanded TCRs at all time points (in each case measured as change relative to pre-sensitization frequency) and the sensitization/patch test scores at all time points is summarized in Figure 4D. As illustrated above, the numbers of expanded TCRs at PT1 were correlated with the strength of the reaction recorded at the sensitization site, and with the patch test score at PT1 (2 weeks), as well as PT2 (6 weeks). The correlation was lost at PT3 (24 weeks), perhaps reflecting the greatly decreased range of patch test scores at this time point (Figure 4B).

The expanded TCRs show characteristics of antigen-driven responses

We hypothesized that the population of TCRs that are found at increased frequency post-exposure to DPC may be enriched for T cells which bind to DPC, or DPC-modified peptides. Antigen-specific sets of TCRs frequently share sequence features, including skewed use of V and J regions, and similarities in CDR3 sequence (Dash et al., 2017; Davis et al., 1995; Glanville et al., 2017; Pogorelyy et al., 2019; Sun et al., 2017; Thomas et al., 2014). The relative V and J gene usage profiles within the expanded set of TCRs compared to the pre-sensitization repertoire for that individual are shown for each patient in Figure 5. We used a non-parametric bootstrapping approach to determine which V and J genes were statistically significantly skewed in the repertoire of the expanded TCRs. We compared the observed proportion of each V and J gene in the PT1 expanded TCRs (≥8) with 1000 random sets of the same number of TCRs sampled from the pre-sensitization repertoires of the same individuals. V and J genes that ranked in the top 50 out of 1001 were considered significantly under-represented at the 0.05 significance level, and genes in the bottom 50 ranks as significantly over-represented. Several examples of skewed V and J gene usage were observed, both in TCR alpha and TCR beta (pink dots). Due to the relatively small number of expanded TCRs in some patients, under-represented genes were very common as a result of sampling, and hence less statistically robust. We therefore show only over-represented V and J genes. In addition, genes that were over-represented by a similar analysis on the set of all expanded TCRs (generated by combining the individual patient expanded sets and taking unique TCRs) are also shown (Figure 5, pink gene names). Interestingly, some V and J regions were skewed in several different individuals, suggesting some common driver of V/J gene usage, despite the fact that the individuals were not related.

Figure 5

Download asset Open asset

We also looked for potential clustering of CDR3 sequences of the expanded TCRs (Figure 6). We have previously shown that amino acid triplets (sequences of three adjacent amino acids within the CDR3 region) can predict antigen specificity (Sun et al., 2017). We therefore compared the pairwise similarity of all expanded CDR3s using a metric that quantifies the number of shared triplets between two sequences, normalized for sequence length (an example of a string kernel in the machine learning literature Shawe-Taylor and Cristianini, 2004). We converted the matrix of pairwise similarities into a network (Figure 6A), by connecting all those TCRs with a similarity above a given threshold (75%). As control, we performed the same analysis on same-sized sets randomly sampled either from the combined pre-sensitization repertoire (Figure 6B,C (ii)) or from the combined repertoire of the healthy volunteers (Figure 6B,C (iii)). The expanded TCRs formed significantly more large clusters of ‘related’ TCRs than either control sets (Figure 6B,C (i)). We selected the largest cluster of CDR3 beta sequences (Figure 6D) and carried out an alignment (summarized in Figure 6E, full alignment in Supplementary file 5). The sequences showed a high level of similarity consistent with belonging to a set of TCRs with shared specificities (Dash et al., 2017; Glanville et al., 2017).

Figure 6

Download asset Open asset

Taken together, the skewing of V and J genes, and the presence of clusters of similar TCRs were strongly suggestive that the expanded TCRs were enriched for TCRs responding to a limited number of specific epitopes.

A dynamic Bayesian network can predict sensitization based on TCR sequence

The skewed V/J usage and the TCR clustering are indicative of an antigen-specific response. We therefore explored whether the repertoire of expanded TCRs was sufficiently distinct from that of the unselected repertoire so that it could be used to distinguish DPC-expanded TCRs or sets of TCRs from the unexpanded repertoire. We used a dynamic Bayesian network (DBN) (Dagum et al., 1992; Murphy and Mian, 1999; Murphy, 2002; Pearl, 1988; Yao et al., 2008), a model we have developed to interrogate antigen specificity in TCR sequences. The DBN is a probabilistic graphical model, which provides a flexible classification tool that can incorporate heterogeneous data (e.g. sequences as well as V/J usage).

We constructed a DBN to classify TCR beta CDR3 sequences into one of two classes: DPC-related CDR3s and randomly sampled pre-sensitization CDR3s from the same individuals. To construct the training/test sets for the model, we used the PT1 expanded CDR3 sequences (≥8) for the DPC class, which were at most 22 amino acids long (2019 beta sequences from 22 patients), and for the control class, a randomly generated set of the same number of TCRs from the combined pre-sensitization sequences. CDR3 sequences in the expanded set were excluded from the pre-sensitization (control) training set. The length of CDR3 sequences used in the model was capped at 22 in order to manage the computational complexity of the algorithm, while still utilizing all expanded sequences. To account for the variable length of the CDR3 sequences, the sequences were aligned by the first Cysteine of the CDR3 on the left, up to the final Phenylalanine of the CDR3, then ‘completed’ to length 22 using a dummy variable (‘amino acid number 21’). The DBN took as input the amino acid in each position in the CDR3 (converted to an integer 0–21 for each amino acid or dummy), the V and J genes, as well as position-specific triplet scores and class scores (as explained in Materials and methods). The proposed probabilistic dependencies between these sequence features and antigen specificity are encoded into the network structure by directed edges (Figure 7A). Further details of the DBN and how it was trained and tested are provided in Materials and methods.

Figure 7

Download asset Open asset

Figure 7—source data 1

Table of expanded T cell receptors (TCRs).

The list of TCRs which is used as the DPC set to generate the DBN which is illustrated in Figure 7. Each row contains a unique combination of CDR3 sequence, V gene, and J gene annotation.

https://cdn.elifesciences.org/articles/54747/elife-54747-fig7-data1-v2.csv

Download elife-54747-fig7-data1-v2.csv

Figure 7—source data 2

Table of control T cell receptors (TCRs).

The list of TCRs which is used as the control set to generate the DBN which is illustrated in Figure 7. Each row contains a unique combination of CDR3 sequence, V gene, and J gene annotation.

https://cdn.elifesciences.org/articles/54747/elife-54747-fig7-data2-v2.csv

Download elife-54747-fig7-data2-v2.csv

One of the key features of the DBN is that each ‘slice’ of the network is position specific. This enables modeling local as well as global sequence features. For example, Figure 7B demonstrates the relationship between neighboring amino acids in specific positions in the CDR3 sequence and antigen specificity. The probability of observing a specific amino acid in a particular position given its neighbors differs between DPC-expanded sequences and control sequences. Both over and under-represented pairs of amino acids can be seen, in particular in the middle of the CDR3. Such relationships are utilized by the model in classification.

The DBN classifier could classify the full DPC/control test sets with mean accuracy of 94% from 10 repeats of 10-fold cross validation, and sets of 10 sequences with mean accuracy 77.7% (Figure 7C upper panel). Model confidence in classifying individual sequences correctly, taken as the difference in the log likelihood between the DPC and the control models for each sequence, varied between different test sets and sequences. Where the DBN was asked to classify every sequence regardless of confidence level, the mean accuracy was 54.3%. When taking the 5% or 10% most confident assignments, individual sequences were classified with mean accuracy of 64.7% and 60.6% respectively, from 10-fold cross validation (Figure 7C lower panel). The DBN provides mechanistic insights into which features are critical for classification, in addition to confidence bounds on its decisions, since the probabilities used by the model in classification are transparent.

The results presented above report the first comprehensive analysis of the TCRrep in the context of controlled exposure to a contact sensitizer in humans. The key findings are that the response to sensitizer is accompanied by a dynamic, robust polyclonal increase in abundance of a defined set of TCR genes which presumably reflects clonal expansion of antigen-specific T cells in response to DPC. Notably, the breadth of the response, as captured by the number of TCRs that increase in frequency, is very diverse between individuals and predicts the degree of sensitization. Furthermore, the response of individual TCRs, which reflect individual T cell clonal dynamics, is transient, with most TCRs returning to low frequencies, or disappearing by the subsequent time point.

The immunological mechanisms that drive ACD have been studied in detail (see for example Kimber et al., 2012). The dose and frequency of exposure are both key parameters in determining the strength of the response (Friedmann, 2006; Friedmann, 2007), although it is clear that the full impact of different factors on the observed variance in the population is still not fully understood, and is one limitation in accurate prediction of allergy to chemical exposure in the context of consumer safety (Basketter and Safford, 2016). There is widespread agreement that activation of both CD4+ and CD8+ antigen-specific T cells plays key roles, although antibody may also be important (Singleton et al., 2016).

The nature of the molecular interaction between the sensitizer, which is usually a chemical hapten that reacts with macromolecules including proteins in the skin following exposure, remains much less well understood than for microbial or viral peptide antigens. There are no MHC tetramers or multimers which can be used to identify DPC or indeed any hapten-specific T cells. Nevertheless, a number of lines of evidence support our hypothesis that the TCRs that increase in frequency following sensitization represent an antigen-specific response. The observed changes from zero to eight are unlikely to be due to sampling effects; in fact, modeling sampling as a Poisson process indicates that changes from zero to eight or more have individual probabilities of <0.05. The observed changes are also unlikely to arise from exposure to unknown antigens (e.g. infectious agents) since the number of expanded TCRs at the first time point post-sensitization is greater than the number of expanded TCRs observed in unsensitized individuals sampled at similar time points, for the great majority of the patients tested. The functional link between TCR expansion and DPC exposure is also strengthened by the observed statistical correlation between the number of expanded TCRs and the vigor of sensitization, as measured by the patch test. Moreover, the expanded set of TCRs show strong evidence of V and J gene skewing, and of increased CDR3 clustering, both well-established signatures of an antigen-specific T cell response (Thomas et al., 2014; Sun et al., 2017; Davis et al., 1995; Glanville et al., 2017; Dash et al., 2017). Finally, these features are shown to be discriminatory enough to allow classification of TCRs into DPC-related and pre-sensitization TCRs by the DBN. Taken together, this set of observations suggest that DPC, perhaps in the form of a set of modified-self peptide adjuncts, stimulates a transient expansion of T cells carrying a defined set of related TCRs. The observation that TCR clustering, V region enrichment, and DBN classification could be observed across a panel of unrelated individuals is particularly intriguing, and suggests that hapten-specific responses may be less dependent on MHC matching than conventional peptide-specific responses. However, further more detailed study is required to determine the molecular target of the DPC-responsive TCRs.

DPC is considered a potent sensitizer (Stute et al., 1981; Mose et al., 2017b). Nevertheless, we noted a wide range of quantitative responses in different individuals, at least as measured by standard patch test results. This was reflected in a wide range of expanded TCRs, and the number of expanded TCRs at the first post-sensitization sample (2 weeks) showed strong correlation with the strength of the response. The reasons for this inter-individual variation are unknown, but could reflect inter-individual differences in immune response (Brodin et al., 2015), or potentially differences in factors that control hapten penetration, such as skin thickness or composition (Reynolds et al., 2019). Understanding the factors that determine this variation may be important in developing novel approaches to ACD risk prediction (MacKay et al., 2013; Maxwell and Mackay, 2008).

In our study, low responders were not boosted by repeated exposure to DPC over many months. In fact, the response was usually maximal at the first patch test, suggesting that regulatory mechanisms come into play to limit the response in vivo. Similarly the initial TCR expansion did not continue to increase despite repeated exposure to antigen, and indeed the majority of TCRs showed very transitory responses, with individual TCR frequencies falling to around pre-sensitization frequencies by the next time point. Similar rapid expansion and subsequent contraction have been observed in response to live attenuated yellow fever vaccine (DeWitt et al., 2015; Pogorelyy et al., 2019; Pogorelyy et al., 2018), one of few examples where TCR repertoire has been studied before and after challenge in a human setting. The latter study (Pogorelyy et al., 2018) precisely documents changes in TCR frequency in three pairs of twins, at several time points before and after exposure to a single dose of the live attenuated YFV 17D vaccine. The study documented the expansion and then contraction of several hundred TCR genes, which reflected the clonal expansion, and subsequent contraction of a polyclonal set of vaccine-specific T cells. However, in this case antigen was delivered as a single dose of vaccine, which gives rise to systemic but very transitory viremia. T cell contraction in the vaccine context may therefore reflect the rapid disappearance of antigen. In contrast, the patients in the current study were exposed to repeated doses of DPC weekly over the course of many months for their therapeutic benefit. The mechanisms which limit the DPC response are therefore more likely to be due to intrinsic regulatory pathways (e.g. Treg induction), or changes in intrinsic migratory pathways rather than simply reflect antigen disappearance. We noted that a second wave of TCRs appeared at increased frequency at PT2, 6 weeks post-sensitization. This set of ‘late’ TCRs was almost completely distinct from the ‘early’ peak at week 2. Furthermore, there was no correlation between the magnitude of the patch test score at PT1 and the number of TCRs expanded at PT2. Additional experiments will be needed to explore whether this set of late TCRs represents part of a regulatory mechanism, which regulates and limits further T cell expansion during chronic exposure to DPC.

In conclusion, this study is the first analysis of in vivo TCR repertoire changes in response to a chemical allergen, and is one of only a handful of studies that document in vivo longitudinal changes in TCRrep in response to immunization in humans. We cannot rule out that the immune response in this population is different from healthy individuals, since all patients suffered from some form of alopecia, which is believed to have an autoimmune etiology (Trüeb and Dias, 2018). Nevertheless, DPC is unlikely to bear any relation to the autoimmune target in alopecia, since similar therapeutic responses (hair growth) can be observed in patients with alopecia treated with other completely unrelated contact sensitizers. While the qualitative balance of immune response may be influenced by the underlying autoimmune background, the specificity of the TCR repertoire is likely to reflect fundamental features of the response to chemical haptens. The study confirms our previous in vitro findings that the response even to a simple chemical such as DPC is polyclonal involving dozens or even hundreds of TCRs. The study also highlights the dynamic nature of the TCR repertoire. Further studies will be required to unravel the complex mechanisms that regulate the immune response to chronic antigen exposure.

All DNA sequences have been submitted to the Sequence Read Archive under identifier PRJNA592875.

1. 1. Ronel T
   2. Harries M
   3. Wicks K
   4. Oakes T
   5. Singleton H
   6. Dearman R
   7. Maxwell G
   8. Chain B

   (2019) NCBI Sequence Read Archive

   ID PRJNA592875. T cell receptor sequencing of alopecia patients during skin sensitisation.

   https://www.ncbi.nlm.nih.gov/sra/PRJNA592875

1. 1. Appay V
   2. van Lier RAW
   3. Sallusto F
   4. Roederer M

   (2008) Phenotype and function of human T lymphocyte subsets: consensus and issues

   Cytometry Part A 73A:975–983.

   https://doi.org/10.1002/cyto.a.20643

   • Google Scholar
2. 1. Ashworth J
   2. Tuyp E
   3. Mackie RM

   (1989) Allergic and irritant contact dermatitis compared in the treatment of alopecia totalis and universalis A comparison of the value of topical diphencyprone and tretinoin gel

   British Journal of Dermatology 120:397–401.

   https://doi.org/10.1111/j.1365-2133.1989.tb04166.x

   • Google Scholar
3. 1. Basketter D
   2. Safford B

   (2016) Skin sensitization quantitative risk assessment: a review of underlying assumptions

   Regulatory Toxicology and Pharmacology 74:105–116.

   https://doi.org/10.1016/j.yrtph.2015.11.013

   • PubMed
   • Google Scholar
4. 1. Brodin P
   2. Jojic V
   3. Gao T
   4. Bhattacharya S
   5. Angel CJ
   6. Furman D
   7. Shen-Orr S
   8. Dekker CL
   9. Swan GE
   10. Butte AJ
   11. Maecker HT
   12. Davis MM

   (2015) Variation in the human immune system is largely driven by non-heritable influences

   Cell 160:37–47.

   https://doi.org/10.1016/j.cell.2014.12.020

   • PubMed
   • Google Scholar
5. 1. Buckley DA
   2. Keane FM
   3. Munn SE
   4. Fuller LC
   5. Higgins EM
   6. Du Vivier AW

   (1999) Recalcitrant viral warts treated by diphencyprone immunotherapy

   British Journal of Dermatology 141:292–296.

   https://doi.org/10.1046/j.1365-2133.1999.02978.x

   • Google Scholar
6. 1. Csardi G
   2. Nepusz T

   (2006)

   The igraph software package for complex network research

   Inter Journal Complex Sy 2006:1695.

   • Google Scholar
7. 1. Dagum P
   2. Galper A
   3. Horvitz E

   (1992) Dynamic network models for forecasting

   Uncertainty in Artificial Intelligence 1992:41–48.

   https://doi.org/10.1016/B978-1-4832-8287-9.50010-4

   • Google Scholar
8. 1. Dash P
   2. Fiore-Gartland AJ
   3. Hertz T
   4. Wang GC
   5. Sharma S
   6. Souquette A
   7. Crawford JC
   8. Clemens EB
   9. Nguyen THO
   10. Kedzierska K
   11. La Gruta NL
   12. Bradley P
   13. Thomas PG

   (2017) Quantifiable predictive features define epitope-specific T cell receptor repertoires

   Nature 547:89–93.

   https://doi.org/10.1038/nature22383

   • PubMed
   • Google Scholar
9. 1. Davis MM
   2. McHeyzer-Williams M
   3. Chien YH

   (1995) T-cell receptor V-region usage and antigen specificity the cytochrome c model system

   Annals of the New York Academy of Sciences 756:1–11.

   https://doi.org/10.1111/j.1749-6632.1995.tb44477.x

   • PubMed
   • Google Scholar
10. 1. DeWitt WS
    2. Emerson RO
    3. Lindau P
    4. Vignali M
    5. Snyder TM
    6. Desmarais C
    7. Sanders C
    8. Utsugi H
    9. Warren EH
    10. McElrath J
    11. Makar KW
    12. Wald A
    13. Robins HS

    (2015) Dynamics of the cytotoxic T cell response to a model of acute viral infection

    Journal of Virology 89:4517–4526.

    https://doi.org/10.1128/JVI.03474-14

    • PubMed
    • Google Scholar
11. 1. Fletcher JM
    2. Vukmanovic-Stejic M
    3. Dunne PJ
    4. Birch KE
    5. Cook JE
    6. Jackson SE
    7. Salmon M
    8. Rustin MH
    9. Akbar AN

    (2005) Cytomegalovirus-specific CD4+ T cells in healthy carriers are continuously driven to replicative exhaustion

    The Journal of Immunology 175:8218–8225.

    https://doi.org/10.4049/jimmunol.175.12.8218

    • PubMed
    • Google Scholar
12. 1. Friedmann PS

    (2006) Contact sensitisation and allergic contact dermatitis: immunobiological mechanisms

    Toxicology Letters 162:49–54.

    https://doi.org/10.1016/j.toxlet.2005.10.008

    • PubMed
    • Google Scholar
13. 1. Friedmann PS

    (2007) The relationships between exposure dose and response in induction and elicitation of contact hypersensitivity in humans

    British Journal of Dermatology 157:1093–1102.

    https://doi.org/10.1111/j.1365-2133.2007.08162.x

    • PubMed
    • Google Scholar
14. 1. Glanville J
    2. Huang H
    3. Nau A
    4. Hatton O
    5. Wagar LE
    6. Rubelt F
    7. Ji X
    8. Han A
    9. Krams SM
    10. Pettus C
    11. Haas N
    12. Arlehamn CSL
    13. Sette A
    14. Boyd SD
    15. Scriba TJ
    16. Martinez OM
    17. Davis MM

    (2017) Identifying specificity groups in the T cell receptor repertoire

    Nature 547:94–98.

    https://doi.org/10.1038/nature22976

    • PubMed
    • Google Scholar
15. 1. Karanovic S
    2. Harries M
    3. Kaur MR

    (2018) Diphenylcyclopropenone for Alopecia Areata: a U.K. survey

    British Journal of Dermatology 70:514–515.

    https://doi.org/10.1111/bjd.16489

    • Google Scholar
16. 1. Karatzoglou A
    2. Smola A
    3. Hornik K
    4. Zeileis A

    (2004)

    Kernlab - An S4 Package for Kernel Methods in R

    Journal of Statistical Software 11:1–20.

    • Google Scholar
17. 1. Kimber I
    2. Maxwell G
    3. Gilmour N
    4. Dearman RJ
    5. Friedmann PS
    6. Martin SF

    (2012) Allergic contact dermatitis: a commentary on the relationship between T lymphocytes and skin sensitising potency

    Toxicology 291:18–24.

    https://doi.org/10.1016/j.tox.2011.11.007

    • PubMed
    • Google Scholar
18. 1. Larsson A

    (2014) AliView: a fast and lightweight alignment viewer and editor for large datasets

    Bioinformatics 30:3276–3278.

    https://doi.org/10.1093/bioinformatics/btu531

    • Google Scholar
19. 1. Lee S
    2. Kim BJ
    3. Lee YB
    4. Lee WS

    (2018) Hair regrowth outcomes of contact immunotherapy for patients with Alopecia Areata: a systematic review and Meta-analysis

    JAMA Dermatology 154:1145–1151.

    https://doi.org/10.1001/jamadermatol.2018.2312

    • PubMed
    • Google Scholar
20. Software

    1. Luttinen J
    2. Dietz L

    (2013) BayesNet, Comprehensive TeX Archive Network (CTAN), version 1.0

    Github.

    https://github.com/jluttine/tikz-bayesnet
21. 1. MacKay C
    2. Davies M
    3. Summerfield V
    4. Maxwell G

    (2013) From pathways to people: applying the adverse outcome pathway (AOP) for skin sensitization to risk assessment

    Altex 30:473–486.

    https://doi.org/10.14573/altex.2013.4.473

    • PubMed
    • Google Scholar
22. 1. Maxwell G
    2. Mackay C

    (2008) Application of a systems biology approach to skin allergy risk assessment

    Alternatives to Laboratory Animals 36:521–556.

    https://doi.org/10.1177/026119290803600510

    • PubMed
    • Google Scholar
23. 1. Mose KF
    2. Andersen F
    3. Skov L
    4. Røpke MA
    5. Litman T
    6. Friedmann PS
    7. Andersen KE

    (2017a) Repeated monthly epicutaneous challenges with diphenylcyclopropenone result in a clinically reproducible level of contact allergy in de novo sensitized individuals

    British Journal of Dermatology 176:1095–1097.

    https://doi.org/10.1111/bjd.14949

    • Google Scholar
24. 1. Mose KF
    2. Burton M
    3. Thomassen M
    4. Andersen F
    5. Kruse TA
    6. Tan Q
    7. Skov L
    8. Røpke MA
    9. Litman T
    10. Clemmensen O
    11. Kristensen BW
    12. Friedmann PS
    13. Andersen KE

    (2017b) The gene expression and immunohistochemical time-course of diphenylcyclopropenone-induced contact allergy in healthy humans following repeated epicutaneous challenges

    Experimental Dermatology 26:926–933.

    https://doi.org/10.1111/exd.13345

    • Google Scholar
25. 1. Murphy KP

    (2001)

    The bayes net toolbox for MATLAB

    Comput Science 33:2001.

    • Google Scholar
26. Thesis

    1. Murphy KP

    (2002)

    Dynamic bayesian networks: representation, inference and learning

    Thesis for: PhD.

    • Google Scholar
27. Book

    1. Murphy K
    2. Mian S

    (1999)

    Modelling Gene Expression Data Using Dynamic Bayesian Networks

    Tech Report, University of California.

    • Google Scholar
28. 1. Oakes T
    2. Heather JM
    3. Best K
    4. Byng-Maddick R
    5. Husovsky C
    6. Ismail M
    7. Joshi K
    8. Maxwell G
    9. Noursadeghi M
    10. Riddell N
    11. Ruehl T
    12. Turner CT
    13. Uddin I
    14. Chain B

    (2017a) Quantitative characterization of the T cell receptor repertoire of naïve and memory subsets using an integrated experimental and computational pipeline which is robust, economical, and versatile

    Frontiers in Immunology 8:1267.

    https://doi.org/10.3389/fimmu.2017.01267

    • Google Scholar
29. 1. Oakes T
    2. Popple AL
    3. Williams J
    4. Best K
    5. Heather JM
    6. Ismail M
    7. Maxwell G
    8. Gellatly N
    9. Dearman RJ
    10. Kimber I
    11. Chain B

    (2017b) The T cell response to the contact sensitizer paraphenylenediamine is characterized by a polyclonal diverse repertoire of Antigen-Specific receptors

    Frontiers in Immunology 8:162.

    https://doi.org/10.3389/fimmu.2017.00162

    • PubMed
    • Google Scholar
30. Book

    1. Pearl J

    (1988)

    Probabilistic Reasoning in Intelligent Systems : Networks of Plausible Inference

    Elsevier.

    • Google Scholar
31. 1. Pogorelyy MV
    2. Minervina AA
    3. Touzel MP
    4. Sycheva AL
    5. Komech EA
    6. Kovalenko EI
    7. Karganova GG
    8. Egorov ES
    9. Komkov AY
    10. Chudakov DM
    11. Mamedov IZ
    12. Mora T
    13. Walczak AM
    14. Lebedev YB

    (2018) Precise tracking of vaccine-responding T cell clones reveals convergent and personalized response in identical twins

    PNAS 115:12704–12709.

    https://doi.org/10.1073/pnas.1809642115

    • PubMed
    • Google Scholar
32. 1. Pogorelyy MV
    2. Minervina AA
    3. Shugay M
    4. Chudakov DM
    5. Lebedev YB
    6. Mora T
    7. Walczak AM

    (2019) Detecting T cell receptors involved in immune responses from single repertoire snapshots

    PLOS Biology 17:e3000314.

    https://doi.org/10.1371/journal.pbio.3000314

    • PubMed
    • Google Scholar
33. 1. Read T
    2. Webber S
    3. Tan J
    4. Wagels M
    5. Schaider H
    6. Soyer HP
    7. Smithers BM

    (2017) Diphenylcyclopropenone for the treatment of cutaneous in-transit melanoma metastases - results of a prospective, non-randomized, single-centre study

    Journal of the European Academy of Dermatology and Venereology 31:2030–2037.

    https://doi.org/10.1111/jdv.14422

    • PubMed
    • Google Scholar
34. 1. Reynolds J
    2. MacKay C
    3. Gilmour N
    4. Miguel-Vilumbrales D
    5. Maxwell G

    (2019) Probabilistic prediction of human skin sensitiser potency for use in next generation risk assessment

    Computational Toxicology 9:36–49.

    https://doi.org/10.1016/j.comtox.2018.10.004

    • Google Scholar
35. Book

    1. Shawe-Taylor J
    2. Cristianini N

    (2004)

    Kernel Methods for Pattern Analysis

    Cambridge university press.

    • Google Scholar
36. 1. Singleton H
    2. Popple A
    3. Gellatly N
    4. Maxwell G
    5. Williams J
    6. Friedmann PS
    7. Kimber I
    8. Dearman RJ

    (2016) Anti-hapten antibodies in response to skin sensitization

    Contact Dermatitis 74:197–204.

    https://doi.org/10.1111/cod.12486

    • PubMed
    • Google Scholar
37. 1. Stute J
    2. Hausen BM
    3. Schulz KH

    (1981)

    Diphenylcyclopropenone - a new strong contact sensitizer (author’s transl)

    Dermatosen in Beruf Und Umwelt. Occupation and Environment 29:12–14.

    • PubMed
    • Google Scholar
38. 1. Sun Y
    2. Best K
    3. Cinelli M
    4. Heather JM
    5. Reich-Zeliger S
    6. Shifrut E
    7. Friedman N
    8. Shawe-Taylor J
    9. Chain B

    (2017) Specificity, privacy, and degeneracy in the CD4 T cell receptor repertoire following immunization

    Frontiers in Immunology 8:430.

    https://doi.org/10.3389/fimmu.2017.00430

    • PubMed
    • Google Scholar
39. 1. Thomas N
    2. Best K
    3. Cinelli M
    4. Reich-Zeliger S
    5. Gal H
    6. Shifrut E
    7. Madi A
    8. Friedman N
    9. Shawe-Taylor J
    10. Chain B

    (2014) Tracking global changes induced in the CD4 T-cell receptor repertoire by immunization with a complex antigen using short stretches of CDR3 protein sequence

    Bioinformatics 30:3181–3188.

    https://doi.org/10.1093/bioinformatics/btu523

    • PubMed
    • Google Scholar
40. 1. Trüeb RM
    2. Dias M

    (2018) Alopecia Areata: a comprehensive review of pathogenesis and management

    Clinical Reviews in Allergy & Immunology 54:68–87.

    https://doi.org/10.1007/s12016-017-8620-9

    • PubMed
    • Google Scholar
41. 1. Uddin I
    2. Joshi K
    3. Oakes T
    4. Heather JM
    5. Swanton C
    6. TRACERx consortium
    7. Chain B

    (2019a) An economical, quantitative, and robust protocol for High-Throughput T cell receptor sequencing from tumor or blood

    Methods in Molecular Biology 1884:15–42.

    https://doi.org/10.1007/978-1-4939-8885-3_2

    • PubMed
    • Google Scholar
42. 1. Uddin I
    2. Woolston A
    3. Peacock T
    4. Joshi K
    5. Ismail M
    6. Ronel T
    7. Husovsky C
    8. Chain B

    (2019b) Quantitative analysis of the T cell receptor repertoire

    Methods in Enzymology 629:465–492.

    https://doi.org/10.1016/bs.mie.2019.05.054

    • PubMed
    • Google Scholar
43. 1. von Moos S
    2. Johansen P
    3. Waeckerle-Men Y
    4. Mohanan D
    5. Senti G
    6. Häffner A
    7. Kündig TM

    (2012) The contact sensitizer diphenylcyclopropenone has adjuvant properties in mice and potential application in epicutaneous immunotherapy

    Allergy 67:638–646.

    https://doi.org/10.1111/j.1398-9995.2012.02802.x

    • PubMed
    • Google Scholar
44. 1. Wilkerson MG
    2. Connor TH
    3. Henkin J
    4. Wilkin JK
    5. Matney TS

    (1987) Assessment of diphenylcyclopropenone for photochemically induced mutagenicity in the ames assay

    Journal of the American Academy of Dermatology 17:606–611.

    https://doi.org/10.1016/S0190-9622(87)70244-8

    • PubMed
    • Google Scholar
45. 1. Yao XQ
    2. Zhu H
    3. She ZS

    (2008) A dynamic bayesian network approach to protein secondary structure prediction

    BMC Bioinformatics 9:49.

    https://doi.org/10.1186/1471-2105-9-49

    • PubMed
    • Google Scholar

Acceptance summary:

This manuscript presents a longitudinal analysis of the whole blood Tcell repertoire dynamics in patients with Alopecia Areata upon primary, secondary and chronic exposures to an allergic chemical agent. The authors have identified signatures of polyclonal T-cell response following sensitization involving stereotypical T-cell receptor genes. Aside from the specific case of T-cell response to the allergic agent in this study, the developed methods could be used for the general and important problem of identifying reactive Tcell populations to antigenic challenges.

Decision letter after peer review:

Thank you for submitting your article "The clonal structure and dynamics of the human T cell response to an organic chemical hapten" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Aleksandra Walczak as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Yuval Elhanati (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

This manuscript presents a longitudinal analysis of the whole blood TCR repertoire dynamics in patients with Alopecia Areata upon primary, secondary and chronic exposure to DPC treatment. The authors have identified signatures of polyclonal T-cell response following sensitization involving stereotypical T-cell receptor genes. Using a Dynamic Bayesian Network, the authors characterize a sequence based classifier that with a high accuracy can predict sensitization based on TCR sequence. Data collection and analysis are performed carefully and the results are presented clearly, but there are still some concerns that we would like to see addressed.

1) Most of the analysis rely on recognizing the expanded clones. For this purpose TCRs were counted based on their fold change – 2, 4, 8 and 16. However, the fold change can be very sensitive to the sample size at each time point – smaller samples tend to have lower diversity, which translates into bigger clone frequencies. This bias in particular, is present for TCR clones with low multiplicities (i.e., the majority of TCRs in your data). For example, consider two replicate blood samples such that one is sequenced ten times deeper than the other. A lot of singletons (TCRs with count 1) should be present in the deep sample, whereas almost all of them should be absent in the other sample. To estimate frequencies however, authors adjust these absent counts from 0 to 1 and divide the adjusted counts by the total counts in each sample. As a result, these fluctuations due to sample depth would appear as 10-fold difference in frequency of the singleton clones between samples.

– To gauge the validity of the analysis, it would be insightful if authors show more basic statistics for the fold change in clone sizes: what are pre-sensitization frequency and counts for expanding clones? Are there a lot of clonotypes that grow from 0 to 8 counts? What fraction of pre-sensitization high-frequency clones expand? One way to show this is to make scatterplots for clone counts in PS vs counts in PT1, ideally in log-scale; see DeWitt et al., 2015 Figure 2 as an example.

– Given the concerns above, it is also necessary that authors provide a table with details on sample size in data (e.g. unique clonotypes and UMIs for each sequenced sample) and show that the results are robust to these differences. For example, do observed differences in Figure 2 hold, if all time points are down-sampled to the same number of UMI per sample?

2) A more detailed description of the methods for analysis of clone size dynamics is necessary: Are clonal frequencies estimated after adding 1 to zero counts or are missing clones assigned a minimum frequency in a given sample (see the example in comment 1)?

3) From the comparison of Figure 2 and Supplementary file 2 it seems that expansion analyses in α and β repertoires do not quite agree. In particular, for the α chain, there are no significant differences between control and patient groups. If there are strong clonal expansions, one expects a correlated fold increase in α and β chains that belong to the same clone. Can authors explain this discrepancy? Is there a correlation between the number of expanded clones in α and the β repertoires?

4) Presentation of Figure 3 should be improved. For examples, clonal dynamics can be shown by connected lines over time. For comparison, it would be useful to show similar statistics in healthy individuals.

5) Figure 4 is difficult to read, specifically the left panels in C, D and E. These plots highlight the dependence of the number of expanded TCRs on the expansion fold change, which is rather trivial and not the point the authors are making here. On the other hand, the differences between the different sensitization groups are difficult to see. A different visualization should be considered.

6) Figure 5: Although V/J usage can reflect a response to DPC, the authors should first analyze whether the V/J usage is patient dependent or treatment dependent. For instance, this can be done by performing the same analysis on all clonotypes from patients versus healthy volunteers, expanded clonotypes in healthy individuals versus 1000 random clones in healthy individuals, etc… In addition, the use of box-plot is not readable for all the V and J. Better summarizing the information as heatmap.

7) For DBN, it is not clear from the Materials and methods how the sequences are aligned since TCRs have variable length.

8) To train a DBN, the authors construct the negative set from TCRs of healthy donors which were not in the expanded positive set. A natural negative set would be the patient-specific TCRs that were present in the pre-sensitized sample but did not expand. Authors should check how their results are robust to this choice of negative set.

Also, TCRs are somewhat individual-specific and so using a negative set from different donors might be problematic and create an "easy" task for classification. This can be fixed by using donor specific training sets.

What is the ratio of positive to negative set? Again, how robust is the performance of DBN to this training ratio. Authors should provide access to their training and test sets as a separate supplementary file.

9) Can authors discuss in more detail the TCR features that are informative for DPC classification. Do the authors expect same classifier performance for other tasks (e.g. distinction of TCRs specific for certain viral epitope from TCRs specific for other epitopes), or high accuracy of DBN is limited to the special case of DPC-recognizing TCRs?

This manuscript presents a longitudinal analysis of the whole blood TCR repertoire dynamics in patients with Alopecia Areata upon primary, secondary and chronic exposure to DPC treatment. The authors have identified signatures of polyclonal T-cell response following sensitization involving stereotypical T-cell receptor genes. Using a Dynamic Bayesian Network, the authors characterize a sequence based classifier that with a high accuracy can predict sensitization based on TCR sequence. Data collection and analysis are performed carefully and the results are presented clearly, but there are still some concerns that we would like to see addressed.

1) Most of the analysis rely on recognizing the expanded clones. For this purpose TCRs were counted based on their fold change – 2, 4, 8 and 16. However, the fold change can be very sensitive to the sample size at each time point – smaller samples tend to have lower diversity, which translates into bigger clone frequencies. This bias in particular, is present for TCR clones with low multiplicities (i.e., the majority of TCRs in your data). For example, consider two replicate blood samples such that one is sequenced ten times deeper than the other. A lot of singletons (TCRs with count 1) should be present in the deep sample, whereas almost all of them should be absent in the other sample. To estimate frequencies however, authors adjust these absent counts from 0 to 1 and divide the adjusted counts by the total counts in each sample. As a result, these fluctuations due to sample depth would appear as 10-fold difference in frequency of the singleton clones between samples.

We agree with this very important point. Indeed, the estimation of ratios is fraught with potential confounders of the type the reviewers raise. On the basis of the suggestions below, we have re-examined the raw data (what the reviewers refer to as “basic statistics”). We note (see below, and new Figure 2) that actually almost all the repertoire activity we observe following sensitisation is focused on TCRs ABSENT pre-sensitisation. Rather than calculate frequencies, and impute missing values, we therefore now focus on those TCRs absent at PS, and expanded at PT1. For comparison, we also analyse the complementary group – those absent at PT1, and expanded at PS. In this way, we do not have to impute any missing values, do not have to calculate fold expansion, but can focus directly on the proportion of TCRs present at an expanded threshold at a particular timepoint. In all quantitative analyses we subsample all the repertoires to the same number of TCRs (defined by unique UMIs), so as to add additional robustness to the analyses.

– To gauge the validity of the analysis, it would be insightful if authors show more basic statistics for the fold change in clone sizes: what are pre-sensitization frequency and counts for expanding clones? Are there a lot of clonotypes that grow from 0 to 8 counts? What fraction of pre-sensitization high-frequency clones expand? One way to show this is to make scatterplots for clone counts in PS vs counts in PT1, ideally in log-scale; see DeWitt et al., 2015, Figure 2 as an example.

We thank the reviewer for this excellent suggestion. This way of visualisation is indeed very instructive and these plots are included in Figure 2, and in Supplementary file 3.

– Given the concerns above, it is also necessary that authors provide a table with details on sample size in data (e.g. unique clonotypes and UMIs for each sequenced sample) and show that the results are robust to these differences. For example, do observed differences in Figure 2 hold, if all time points are down-sampled to the same number of UMI per sample?

We include unique TCRs and total UMIs for each sample in Supplementary file 1. We show that the proportion of expanded TCRs is not correlated to the number of UMIs per sample. We also carry out repeat subsampling of the data to the lowest number of UMIs and calculate the correlation between expansion and patch score. The qualitative picture is the same, although as discussed in more detail in the manuscript, the expansion threshold at which the mean correlation becomes statistically significant changes a little bit.

2) A more detailed description of the methods for analysis of clone size dynamics is necessary: Are clonal frequencies estimated after adding 1 to zero counts or are missing clones assigned a minimum frequency in a given sample (see the example in comment 1)?

This issue is no longer relevant since, as discussed above, we do not use fold change any more, and do not have to assign values to missing points.

3) From the comparison of Figure 2 and Supplementary file 2 it seems that expansion analyses in α and β repertoires do not quite agree. In particular, for the α chain, there are no significant differences between control and patient groups. If there are strong clonal expansions, one expects a correlated fold increase in α and β chains that belong to the same clone. Can authors explain this discrepancy? Is there a correlation between the number of expanded clones in α and the β repertoires?

Expanded α and β TCR numbers are indeed closely correlated (rho – 0.85, p-value<0.0001) (now included in Supplementary file 3). Occasionally there are small differences in statistics for α and β; we presume this reflects sampling effects or some other technical variation between samples.

4) Presentation of Figure 3 should be improved. For examples, clonal dynamics can be shown by connected lines over time. For comparison, it would be useful to show similar statistics in healthy individuals.

We have redrawn these clonal dynamics as lines and included healthy volunteers.

5) Figure 4 is difficult to read, specifically the left panels in C, D and E. These plots highlight the dependence of the number of expanded TCRs on the expansion fold change, which is rather trivial and not the point the authors are making here. On the other hand, the differences between the different sensitization groups are difficult to see. A different visualization should be considered.

We have completely redrawn this figure, focusing on the correlation between number of expanded TCRs and sensitization scores. We believe the results are now clear.

6) Figure 5: Although V/J usage can reflect a response to DPC, the authors should first analyze whether the V/J usage is patient dependent or treatment dependent. For instance, this can be done by performing the same analysis on all clonotypes from patients versus healthy volunteers, expanded clonotypes in healthy individuals versus 1000 random clones in healthy individuals, etc… In addition, the use of box-plot is not readable for all the V and J. Better summarizing the information as heatmap.

We agree with the reviewers that V/J usage may be patient dependent. For this reason, we have compared the expanded set of each individual against unexpanded sets from the same individual. This shows that certain genes are over-represented in the expanded TCRs compared to background for the individual. We have rearranged the plots to make them more readable.

7) For DBN, it is not clear from the Materials and methods how the sequences are aligned since TCRs have variable length.

We have added an explanation of how the sequences are aligned.

8) To train a DBN, the authors construct the negative set from TCRs of healthy donors which were not in the expanded positive set. A natural negative set would be the patient-specific TCRs that were present in the pre-sensitized sample but did not expand. Authors should check how their results are robust to this choice of negative set.

Also, TCRs are somewhat individual-specific and so using a negative set from different donors might be problematic and create an "easy" task for classification. This can be fixed by using donor specific training sets.

We have changed the control set to TCRs from the pre-sensitization repertoires of the same donors as the expanded set. As suggested by the reviewer, this is indeed a more difficult classification task, and we have found that the DBN is very good at telling individuals apart !!! With the new control set, classifying individual sequences is somewhat harder for the DBN, but it is still very successful at classifying sets of sequences. We have included both the results for sets and single sequences in the new Figure 7C.

What is the ratio of positive to negative set? Again, how robust is the performance of DBN to this training ratio. Authors should provide access to their training and test sets as a separate supplementary file.

The positive and negative sets are the same size, and running the model with shuffled class labels produces around 50% accuracy. We have added this to Materials and methods. The positive and negative sets are submitted as Supplementary files.

9) Can authors discuss in more detail the TCR features that are informative for DPC classification. Do the authors expect same classifier performance for other tasks (e.g. distinction of TCRs specific for certain viral epitope from TCRs specific for other epitopes), or high accuracy of DBN is limited to the special case of DPC-recognizing TCRs?

We are writing a manuscript fully focused on the DBN and we would prefer to leave the discussion of other data sets etc. to this other manuscript, as it is not really relevant to this manuscript.

Author details

1. Tahel Ronel

   1. Division of Infection and Immunity, University College London, London, United Kingdom
   2. Cancer Institute, University College London, London, United Kingdom

   Contribution

   Conceptualization, Software, Formal analysis, Visualization, Methodology, Writing - original draft

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9513-9181

2. Matthew Harries

   1. Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom
   2. Salford Royal NHS Foundation Trust (Dermatology Centre), Salford, United Kingdom

   Contribution

   Supervision, Investigation, Project administration, Writing - review and editing

   Competing interests

   No competing interests declared

3. Kate Wicks

   Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

   Contribution

   Investigation, Writing - review and editing

   Competing interests

   No competing interests declared

4. Theres Oakes

   Division of Infection and Immunity, University College London, London, United Kingdom

   Contribution

   Investigation, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

5. Helen Singleton

   Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

   Contribution

   Investigation, Writing - review and editing

   Competing interests

   No competing interests declared

6. Rebecca Dearman

   Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

   Contribution

   Resources, Supervision, Writing - review and editing

   Competing interests

   No competing interests declared

7. Gavin Maxwell

   Safety and Environmental Assurance Centre, Unilever, Colworth Science Park, Bedford, United Kingdom

   Contribution

   Conceptualization, Funding acquisition, Project administration, Writing - review and editing

   Competing interests

   GM is an employee of Unilever PLC. Apart from GM's contribution (see author contributions), the funder was not involved in the study design, collection, analysis, and interpretation of data, the writing of this article or the decision to submit it for publication.

8. Benny Chain

   1. Division of Infection and Immunity, University College London, London, United Kingdom
   2. Department of Computer Science, University College London, London, United Kingdom

   Contribution

   Conceptualization, Formal analysis, Supervision, Funding acquisition, Writing - original draft, Project administration

   For correspondence

   b.chain@ucl.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-7417-3970

• 574

  Page views

• 80

  Downloads

• 0

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.