Introduction

T cells play a crucial role in human adaptive immunity. The recognition of an antigenic peptide bound to the MHC molecule (pMHC) by T cell receptor (TCR) leads to proliferation of the antigen-specific T cell clones (Abbas et al. 2012; Osuna et al. 2014; Pennock et al. 2013). The expanded T cell subsets subsequently differentiate into effector T cells to generate specific immune responses. The effector CD4+ T cells produce several cytokines to activate the recruitment of other immune cells while the effector CD8+ T cells respond to intracellular infections by killing the infected cells. This cascade of events generates long-lived memory T cells that can rapidly initiate the specific immune responses upon re-encountering the same antigen (Abbas et al. 2012).

The TCR is a heterodimer consisting of either a pair of α and β chains or γ and δ chains. The α/β TCR is dominant in human T cell repertoires (Abbas et al. 2012). Each TCR chain is comprised of a variable and a constant extracellular domain. The variable domain is encoded by the germline V, D (only in β and δ chains) and J genetic segments. Within this domain, the antigen recognition site is formed by three complementarity-determining regions (CDR1, CDR2, and CDR3) (Abbas et al. 2012; Clements et al. 2006; Hou et al. Naive T cell data

The naive T-cell dataset (3,823 sequences) was derived from two healthy donors from an independent study (Sequence Read Archive, project SRP109035, data is analyzed in collaboration with Peter de Greef). The unique CDR3 clones with an absolute naive count > 1 and not observed among non-naive sequence reads were included. The CDR3 nucleotide sequences were available for these datasets and we implemented a Python (version 3.5.2) script to calculate the non-VJ peptide sequences containing unidentified D segment and random nucleotide insertions.

Statistical analysis

The frequency distributions were tested for similarity using the Kolmogorov–Smirnov (KS) test. The P value was not displayed if the P > 0.05 except in the supplementary figures in which “ns” was presented as “non-significant.” The two-sided Mann–Whitney statistics (two-sample Wilcoxon test in R) was used for quantitative comparisons such as length and hydropathy. The same criteria for P value was used as previously mentioned. An alternative method other than the two statistics was indicated if performed.

CDR3 amino acid enrichment analysis

The conserved V and J regions of the CDR3 sequences were removed at the terminal ends. The remaining sequence was defined as “non-VJ” region comprising of encoded D segment and randomly inserted nucleotides. The 20 naturally occurring amino acids were clustered into 3 subgroups based on hydropathic properties including hydrophobic (“A,” “C,” “F,” “I,” “L,” “M,” “W,” and “V”), neutral (“G,” “H,” “P,” “S,” “T,” and “Y”), and hydrophilic (“E,” “D,” “K,” “N,” “Q,” and “R”). These three groups were made following the Kyte–Doolittle hydropathy scale (Kyte and Doolittle 1982).

In order to compare CDR3 sequences varying in length, we mapped the CDR3 sequences to the 2D peptide representative of TCR structure (Lefranc 2014; Lefranc et al. 2003). The CDR3 region is defined as the region delimited by “C” residue at 104 position (C104) and “F”/“W”118 (F/W118) residues. So, the anchored residues (C104 and F/W118) and conserved terminal ends were aligned. According to the IMGT database, 80% of CDR3 are 13-amino acidlong and the added residues are located in the middle of CDR3 with the uniquely defined IMGT positions (Lefranc et al. 2003). To avoid multiple gap insertions in the sequence alignment, we selected CDR3 sequences varying in length from 10 to 15 amino acids which covered > 90% of our dataset. The alignment was performed separately in ID and SD subsets and the entire length of alignment was 17 amino acids due to the maximum CDR3 length (15 amino acids) and the two anchored residues. Next, the positional probability matrix (PPM) was derived for N aligned sequences at position j where j ∈ (1, …, 17):

$$ {PPM}_{k,j}=\frac{1}{N}\sum \limits_{i=1}^N\left({X}_{ij}=k\right) $$

The parameter i ∈ (1,…, N) indicates the set of sequences, while k is a set of 20 amino acids and gap (“-”) and Xij is the amino acid at position j, sequence i of the alignment. We then calculated a log ratio of amino acid probability of ID (PMMID) to SD PMM (PMMSD), which we called the “positional log enrichment score” (pLES). A positive ratio indicated the enrichment of amino acids at a certain position in ID compared to SD. Occasional ∞ and − ∞ values appeared in the pLES matrix due to missing amino acids in SD and/or ID. These values were replaced with 5 and – 5, respectively, since the maximum |pLES| were observed around 4. The matrix was displayed as a heatmap colored based on pLES. The positive pLES was subsequently converted into a sequence logo for improved visualization.

Results

We obtained human CDR3 sequences from the VDJdb (Shugay et al. 2018), where TCR sequences with known epitope specificity from several studies are collected. We observed 99 (~ 1.7%) identical epitope-specific CDR3β detected both as ID and SD responses. This suggests that an ID response from one individual only in very rare cases (< 2%) is an SD response in another individual. In other words, immunodominance of a clone is more universal than originally thought. The redundant CDR3 sequences were removed to obtain a unique CDR3 dataset for ID and SD responses. After this processing, 5811 CDR3 sequences remained which were described as responses against nine virus species, namely HIV-1, CMV, influenza A virus (IAV), EBV, yellow fever virus (YFV), HCV, and four serotypes of dengue virus (DENV1–4) (Fig. 1a). Several epitopes from the same virus were usually presented by different MHC molecules, see, e.g., HIV-1 epitopes (Fig. 1a). As expected, most of the CDR3 sequences (56%, n = 3276) were restricted to HLA-A*02 molecules (Fig. 1b). The top three most abundant CDR3 sequences were also A2 restricted: GILGFVFTL (GIL, n = 1130) from IAV, NLVPMVATV (NLV, n = 979) from CMV, and GLCTLVAML (GLC, n = 738) from EBV.

Fig. 1
figure 1

Overview of selected CDR3β sequences from the VDJ db. The human CDR3 sequences (n = 5811) against 9 virus species were selected. The CDR3 sequences are color-coded based on the interacting MHC molecule involved in presenting a viral anitgen that is recognised by the CDR3. Different viral antigens derived from the same species can be presented by different MHC molecules (a). Most of the selected CDR3 responded to viral presented by HLA-A*02-encoded MHC (b)

Full size image

Influence of CDR3 length on immunodominance

The CDR3β contains the D segment interspersed between V and J regions and therefore should be longer than the CDR3α, which lacks the D segment. As expected, the CDR3β chains in our dataset were significantly longer than CDR3α chains (Fig. 2a, P < 0.001 KS test). The length distributions of both chains were normally distributed (Shapiro–Wilk test, P < 0.001) as has been observed previously (Moss and Bell 1996; Ma et al. 2016; Niemi et al. 2015). The average lengths of the CDR3α and β chains were 11.6 ± 7.4 and 12.3 ± 5.5 amino acids, respectively. Surprisingly, this result was not always consistent for paired α/β CDR3, as around 30% of unique α/β pairs contained longer α chains (supplementary Fig. 1A).

Fig. 2
figure 2

CDR3 length analysis. The distribution of CDR3α (n = 2008) and CDR3β (n = 5811) lengths are  different (KS test, P < 0.0001). The left-shifted CDR3α distribution in relative to the CDR3β suggests a significantly shorter CDR3 in the TCR α chain (a). The similar distribution of entire CDR3β length (b) and the non-VJ region is observed in ID (n = 506) and SD responses (n = 5305) (c). The CDR3β length comparison between naive T cells (n = 3823) and non-naive population reviews significantly longer in the latter (KS test, P < 0.005) (d). The violin plots of non A2-restricted CDR3β (n = 2535) with the mean CDR3β length labels show a negative correlation between the length of epitope and CDR3β (spearman correlation, r = − 0.14). The paired comparison between CDR3β length at 8- to 10-amino acid epitopes with the 11-amino acid long confirms the significantly shorter CDR3β in the longer epitope (Mann–Whitney U test, all P < 0.0001) (e)

Full size image

Next, we divided each CDR3 sequence into 3 regions; V, J, and non-VJ segments, and performed the length comparison between the α and β chains. We used the term non-VJ region to refer to the central part of the CDR3 that is not derived from the germline-encoded V and J genes (see the “Materials and methods” section). This junctional region is composed of additional nucleotides and D segment (only in the β chain); thus, we expected the longer non-VJ region in the CDR3β chain. Interestingly, we observed significantly longer J regions in the CDR3α chains while the V and non-VJ segments were longer in CDR3β chains (supplementary Fig. 1B–D).

It has been suggested that shorter CDR3 sequences are more likely to be generated, as they closely resemble the encoded peptide and need very deletion events (Hou et al. 2A, B). Interestingly, we did observe a negative relationship between the length of epitope and the corresponding CDR3 sequences. This result might reflect a bias due to the dominance of A2 restricted T cell responses in our dataset (Fig. 1b) and the A2 epitopes were all 9 amino acid long. We removed A2 epitopes and we still found a weak but significant negative correlation (Spearman correlation coefficient, r = − 0.14, P < 0.001) between epitope length and CDR3 length (Fig. 2e). This trend was also observed in a set with only EBV and HIV-1 epitopes (8 to 11 amino acid long), restricted by different HLA types (supplementary Fig. 2C). This result suggests that longer CDR3β chains are not needed to recognize longer epitopes bulging from the MHC molecule, which was previously hypothesized/shown (Ekeruche-Makinde et al. 2013). Next, we compared epitope length with the combined CDR3 lengths of paired α/β CDR3 chains (no A2 epitopes included; supplementary Fig. 2D). The significant correlation we found between the epitope length and CDR3 length disappeared in this case, which might suggest that a shorter CDR3β chain may be compensated with a longer CDR3α chain to preserve the interaction with varying epitope lengths (supplementary Fig. 2E).

Amino acid composition in CDR3 sequences

The N- and C-terminus of CDR3 sequences are highly conserved due to the germline encoded V and J segments, respectively. However, the centrally variable region could hold the potential to diverse immune responses. Therefore, we selectively studied amino acid profiles of the non-VJ region of CDR3β sequences. The ID and SD responses differed slightly in their different amino acid distribution (Fig. 3a). In general, the frequently presented amino acids were small (based on the molecular volume) and neutral: glycine (“G”), serine (“S”), threonine (“T”), and proline (“P”). Additionally, arginine (“R”), glutamine (“Q”), glutamic acid (“D”), alanine (“A”), and leucine (“L”) were observed more than the expected 5%. Approximately 25% of amino acids detected in the non V-J region is “G” irrespective of being an ID or SD response. This observation was consistent across the HLA alleles (supplementary Fig. 3A). The highly enriched “G” residue in CDR3β sequences is possibly due to the guanine-rich nucleotide sequences of D segment which was estimated to be around 70% in both D1 and D2 genes (Freeman et al. 2009; Venturi et al. 2008). Therefore, the D segment might cause a bias in codons containing guanine (“g”) like “ggx” (where x stands for any nucleotide) for “G.” If this is the case, the dominance of the glycine residue should disappear in CDR3α sequences that contain no D segment. To test this hypothesis, we compared the amino acid distributions of all CDR3 α and β sequences in our dataset (Fig. 3b). Some hydrophobic residues, namely, cysteine (“C”), methionine (“M”), and isoleucine (“I”) are highly enriched in the non-VJ region of CDR3α relative to the β chain (Fig. 3b). However, the most predominant residue of the CDR3α was also “G.” The log enrichment ratio between the CDR3β and CDR3α chains of “G” was positive suggesting that the residue in CDR3β was more frequently observed than the CDR3α. However, the difference is rather small making it unlikely that the D segment is solely responsible for the enrichment of “G” in CDR3β sequences (Fig. 3b).

Fig. 3
figure 3

Amino acid composition of the non-VJ region. The hydrophilic, hydrophobic, and neutral residues distributions differ between ID and SD responses (KS test, P < 0.002, P < 0.003, and P < 0.03, respectively). The neutral amino acids are predominant in the non-VJ region and the “G” is the most enriched residue (a). The log ratio of amino acid frequency in the non-VJ CDR3β to CDR3α demonstrates highly enriched “Q” in CDR3β and hydrophobic amino acids, namely, “C,” “M,” and “I” in CDR3α (b). Guanine is frequently observed in non-VJ of naive CDR3α and CDR3β nucleotide sequences (c)

Full size image

To determine if the non-VJ segment of the antigen experienced T cells are shaped by naive T cells, we also compared the amino acid composition of the naive and antigen experienced CDR3 population. The amino acid profiles of CDR3β and CDR3α sequences were similar in both T cell populations (supplementary Fig. 3B). In general, the five most frequent amino acids (S, R, P, L, and G) observed in both CDR3 chains are the ones that can be encoded by 4 or more codons, suggesting an effect of codon degeneracy on frequency of an amino acid in non-VJ regions. The nucleotide sequences available for the naive T cell data confirmed enriched guanine even without the D segment in CDR3α (Fig. 3c). Therefore, we can conclude that enrichment of the nucleotide guanine (x), which probably results in an enrichment of the amino acid glycine (due to ggx code, see above) is not influenced by D segment, TCR chain, or antigen exposure.

To test position specific enrichment of amino acids in CDR3 sequences, we aligned unique CDR3 sequences of ID and SD separately. Prior to the alignment, CDR3 sequences were mapped to the IMGT positions corresponding to the 2D peptide chains representing the functional TCR structure (Lefranc 2014; Lefranc et al. 2003). As a consequence, the conserved terminal ends were aligned and gaps were allowed in the central variable domain. We then created the positional probabilistic matrices (PPM) to present the observed frequency of each amino acid throughout the CDR3 alignment. From this PPM, we also calculated positional probability ratios of ID to SD responses. This log ratio value was used as a positional log enrichment score (see the “Materials and methods” section). In most of the positions, all amino acids were presented equally in ID and SD. However, in positions 106, 112.1, and 115 to 117, several amino acids were enriched in SD sequences, suggesting that certain amino acids promote the subdominance probably because of the weak interactions with pMHC complexes (supplementary Fig. 4).

To refine the possible motifs in analysis, we repeated the same procedure at the level (Fig. 4a). We selected three with the most CDR3 sequences (GIL, NLV, and GLC) in order to have enough data. All three were presented by the HLA-A*02 molecule. The SD-associated CDR3 sequences were abundant and only a small fraction of around 2–6% of CDR3 sequences were defined as ID for each epitope. We computed the matrices for each and displayed the matrices as heat maps and sequence logos (Fig. 4b–g). The negative (blue), inferring an enrichment in SD, was observed dominantly (Fig. 4b–d), while few highly enriched amino acids were found in ID (red) responses. The sequence logos of the positive amino acids (Fig. 4e–g) displayed very different patterns for the different, suggesting that a simple motif/pattern is very difficult to find even among CDR3 sequences recognizing different on the same HLA molecule. Although one can detect some distinctive patterns between ID and SD per epitope, a general amino acid signature that defines CDR3 being an ID response was not found.

Fig. 4
figure 4

pLES heatmaps and corresponding sequence logos of CDR3 against the three prevalent epitopes: GIL, GLC, and NLV. Amino acid distributions (ID and SD combined) of CDR3 sequences against the three epitopes are similar (tested by Anderson–Darling k-sample test (Canada 1986)) (a). Log ratio of amino acid frequency in ID to SD at each CDR3 position from C104 to W/F118 are displayed in pLES heatmap. The positive (red) indicates enriched ID and the negative (blue) referred to enrichment in SD relative to ID. The gray color illustrates the absence of an amino acid at a certain position. GIL, GLC, and NLV (b–d) were converted into sequence logos which were created from positive pLES from antigen-specific CDR3 alignment (e–g)

Full size image

Discussion

The recently developed VDJdb contains a vast amount of epitope-specific CDR3 sequences. These sequences were derived from multiple research groups working on shared CDR3 sequences between individuals, recognition of dominant epitopes, binding motifs in CDR3 region, V(D)-J recombinations, and repertoire diversity (Shugay et al. 2018). To our knowledge, the immunodominancy of T cell responses had not yet been explicitly addressed with these data, despite its importance in sha** immune responses (Osuna et al. 2014). We here present the first attempt to do this.

We first focused on common properties like the length and of CDR3β sequences, since they have been shown to impact immune responses (Hou et al. 2017; Freeman et al. 2009; Moss and Bell 1996; Lundegaard et al. 2010; Ma et al. 2016; Song et al. 2017). We observed uniquely biased V-J usage for each epitope and found that the combined V and J genes in ID clones were largely shared with the SD responses (data not shown). To predict the generation probabilities of ID and SD responses (based on their VDJ recombinations), we made use of the OLGA server (Sethna et al. 2019). A per-epitope comparison of the generation probabilities for ID and SD responses did not indicate a significant difference, suggesting that the immunodominance of a T cell clone cannot be explained by high generation probability of its TCR sequence.

We mainly performed our analysis on CDR3β and when available on CDR3α sequences. However, the binding site of TCR to pMHC is composed of CDR1, CDR 2, and CDR3 regions from both TCR chains (Abbas et al. 2012; Clements et al. 2006; Hou et al. 2016; Hughes et al. 2003). The entire complex interaction can therefore not be studied completely using the available data on CDR3 sequences, if the other CDR regions play unexpectedly pivotal role in TCR-pMHC engagement (Miyazawa and Jernigan 1996). Thus, our analysis provides only a preliminary view of the interaction between CDR3 and bounded epitopes. When more data become available on CDR2 and CDR3 sequences, our analysis should be repeated.

In conclusion, we found that the global properties of CDR3 sequences between ID and SD are highly similar. However, several features are distinctive regarding epitope specificity and, thus, can enable classification of epitope-specific CDR3 from a diverse T cell response. Most interesting findings of our study, though slightly outside of our initial research question, were differences between antigen experienced and naive T cell clones. The results raised by our analysis ask for larger sets of naive T cell repertoires for validation.