Repertoire analyses reveal T cell antigen receptor sequence features that influence T cell fate

Abstract

T cells acquire a regulatory phenotype when their T cell antigen receptors (TCRs) experience an intermediate- to high-affinity interaction with a self-peptide presented via the major histocompatibility complex (MHC). Using TCRβ sequences from flow-sorted human cells, we identified TCR features that promote regulatory T cell (T_reg) fate. From these results, we developed a scoring system to quantify TCR-intrinsic regulatory potential (TiRP). When applied to the tumor microenvironment, TiRP scoring helped to explain why only some T cell clones maintained the conventional T cell (T_conv) phenotype through expansion. To elucidate drivers of these predictive TCR features, we then examined the two elements of the T_reg TCR ligand separately: the self-peptide and the human MHC class II molecule. These analyses revealed that hydrophobicity in the third complementarity-determining region (CDR3β) of the TCR promotes reactivity to self-peptides, while TCR variable gene (TRBV gene) usage shapes the TCR's general propensity for human MHC class II-restricted activation.

Extended Data Fig. 1:. Mutual information structure of the TCRβ sequence.

(a) – (e) Heatmap depicting the mutual information structure of the CDR3β amino acid sequence for CDR3βs of length 12 (a), 13 (b), 14 (c), 16 (d), and 17(e) in the discovery dataset. The lower diagonal features normalized mutual information (NMI) between each pair of TCR positions, while the upper diagonal features the maximum mutual information achieved by conditioning on any other TCR position. NMI color scale for (a)-(e) is provided in (a). (f) Probability of each amino acid in each TCR position depicted by a sequence logo. (g) Heatmap as in (a) – (e) for CDR1β and CDR2β loop positions as well as TCR features derived from the flanking regions of CDR3β (Methods). (h) Categorization of amino acids by isoelectric point and interfacial hydrophobicity (Methods).

Extended Data Fig. 2:. Consistency of TCR feature effects across individuals and clinical phenotypes.

(a) T_reg odds ratio per standard deviation increase in CDR3βmr occupancy by each of the 14 relevant amino acids, estimated separately for the T1D cases in the discovery cohort (y axis) and the controls (x axis) (b) T_reg odds ratio per standard deviation increase in CDR3βmr occupancy by each of the 15 relevant amino acids, estimated separately in each donor. (c) T_reg odds ratio for the usage of each TRBV gene relative to the reference gene TRBV05–01, estimated separately for the T1D cases in the discovery cohort (y axis) and the controls (x axis) (d) T_reg odds ratio for the usage of each TRBV gene relative to the reference gene TRBV05–01, estimated separately in each donor. P values in (a) and (c) are calculated by a two-sided t-test with Fischer transformation on Pearson’s R.

Extended Data Fig. 3:. Multicollinearity analysis.

(a)-(c) Maximum Pearson’s correlation observed between each pair of TCR features in the discovery dataset, for all possible combinations of amino acid-based TCR feature values (Methods). Heatmaps are separated by TCR region: (a) CDR3βmr, (b) TRBV-encoded (CDR1β loop, CDR2β loop, and the V-region of CDR3β) and, (c) TRBJ-encoded. (d) Feature selection for the V-region model based on variance inflation in estimated regression coefficients (Methods); each plot represents a candidate mixed effects logistic regression model jointly modeling the effects of TCR features on the x-axis. Black arrow denotes improvement from the first model to the second model via reduction of the variance inflation factor (VIF). Black horizontal line denotes the ideal VIF: zero inflation compared to a model with uncorrelated features. (e) Same as (d), for candidate J-region models.

Extended Data Fig. 4:. Thymic selection rates for TRBV and TRBJ genes.

Thymic selection rates for each TRBV and TRBJ gene in each donor in the discovery cohort and in a reference cohort of 666 healthy donors, inferred by relative gene usage in productive reads versus nonproductive reads (Supplementary Note).

Extended Data Fig. 5:. Estimated effects of physicochemical features at each TCRβ position, stratified by CDR3β length.

(a) Estimated log odds ratio for T_reg per standard deviation of each physicochemical feature at each CDRβ(1–3) loop position in each CDR3β length; features with an estimate > 0 are positively associated with T_reg fate while features with an estimate < 0 are negatively associated. For each CDR3β length, all effects were estimated jointly in an L2-regularized logistic regression with a penalty weight tuned via 10-fold cross-validation (Methods). (b) T_reg odds ratio per standard deviation increase in each physicochemical feature at each CDR3βmr position for each CDR3 length (Methods, Supplementary Table 9). Error bars denote 95% confidence interval for the estimated odds ratio.

Extended Data Fig. 6:. Cell type identification for thymic T cells.

(a) scRNAseq thymic dataset cells arranged in a 2-dimensional embedding by UMAP and colored by normalized expression level of select transcripts; gray (low) to red (high). (b) Transcriptional cluster assignments. (c) Average normalized expression of cell-type-relevant transcripts per cluster.

Extended Data Fig. 7:. Cell type identification for tumor microenvironment T cells and reference T cells.

(a) Log-normalized CD8A, CD4 and FOXP3 mRNA expression in T cells from breast tumor biopsies in Azizi et al. 2018, organized into a 2-dimensional embedding by Uniform Maniform Approximation and Projection (UMAP). (b) Louvain clustering of breast tumor microenvironment T cells. Broad cell type labels are indicated for each cluster in the surrounding legend. (c) Expression levels of key surface proteins measured by CITE-seq in the CD4+ reference single cell dataset (low = purple, high = light green). Protein levels are normalized by the centered log-ratio (CLR) transformation (Methods). (d) LogCP10K-normalized expression levels of key mRNA transcripts in the CD4+ reference single cell dataset (low = purple, high = light green).

Extended Data Fig. 8:. Symphony mapping details.

(a) Tumor microenvironment T cells mapped into the reference embedding by Symphony, colored by donor to reveal successful integration of donors. (b) same as (a), colored by cancer type to reveal successful integration of cohorts. (c) Tumor microenvironment T cells mapped into the reference embedding by Symphony, colored by cell types derived from internal clustering (by Yost et al. for the SCC and BCC samples, and as depicted in Extended Data Figure 7a−b for the BRCA samples) to show the extent of concordance with Symphony’s cell type solutions. (d) same as (a), colored by the TiRP score of their TCR. TiRP is scaled such that 0 corresponds to the mean score and one unit corresponds to one standard deviation of held-out bulk sequencing TCRs (Figure 5c). (e) FOXP3 expression differences between T_regs and T_convs within mixed clones of three representative donor samples. Each mixed clone is represented by a line connecting the average FOXP3 expression of Tregs within the clone to the average FOXP3 expression of T_convs within the clone. Each P value is computed by a two-sided paired t-test comparing the mean FOXP3 expression in Tregs to that in T_convs within each mixed clone.

Extended Data Fig. 9:. Further analysis of principal components, murine Tregs, and human memory Tconv.

(a) 67 samples from the replication cohort colored by donor ID and arranged by principal component space according to variation in TCR sequence feature frequencies. (b) Same as (a), colored by donor clinical phenotype. (c) Replication of CDR3βmr percent composition of amino acid effects in mice. Error bars correspond to 95% confidence intervals for ORs. (d) Lack of mouse-human correspondence for position-specific TCR feature effects. TCR features are colored by type; error bars denote OR 95% confidence intervals. Murine TRBV genes were mapped to their human homologs for comparison, only those with a human homolog are shown (Methods). (e) Mean TiRP component scores for CD4⁺ expanded pure T_conv, pure T_reg, and mixed clones in the tumor microenvironment^,. Error bars denote standard error of the mean. T_conv mTiRP compared to mixed clone mTiRP two-sided Wald test P = 2.9 × 10⁻⁴, all other comparisons nonsignificant. (f) Overall lack of correspondence between Treg-Tconv OR and memory-naïve OR for CDR3βmr percent composition of amino acids. Error bars correspond to 95% confidence intervals, and amino acids are colored by the scheme in (c). (g) Replication of memory T_conv – naive T_conv TRBV gene odds ratios in an independent dataset of sorted memory and naïve T cells from 4 healthy donors. TRBV genes are colored by their T_reg-T_conv odds ratios. For (c), (d), (f), and (h), R = Pearson’s correlation coefficient and P values are computed by a two-sided t-test with Fischer transformation. For (e)-(g), human T_reg-T_conv OR result from fixed-effect meta-analysis across the discovery and replication cohorts.

Extended Data Fig. 10:. TiRP scoring of autoreactive T cell receptors.

TiRP scores of McPAS and VDJdb autoimmune TCRs (points) compared to memory T_convs and T_regs from the replication dataset held out for testing (boxplots). Each point in the autoimmune category represents one TCR from McPAS or VDJdb. Error bar denotes standard error of the mean TiRP for autoreactive TCRs, which is higher than reference memory T_convs (P = 1.5 × 10⁻⁹, two-sided Wald test), but not significantly different from reference T_regs (P = 0.43, two-sided Wald test). Within each boxplot, the horizontal lines reflect the median, the top and bottom of each box reflect the interquartile range (IQR), and the whiskers reflect the maximum and minimum values within each grouping no further than 1.5 × IQR from the hinge. T1D = Type 1 Diabetes CD = Celiac Disease IBD = Inflammatory Bowel Disease MS = Multiple Sclerosis

Figure 1.. Study design.

(a) We first examined the structure of the T cell receptor (TCR) sequence to define 1080 sequence features. Depicted is a T cell receptor (TCR) β chain in complex with antigenic peptide (red) and human MHC II molecules (brown). The TCR is colored by region: V-region (including CDR1β and CDR2β loops) in green, CDR3β middle region (CDR3βmr) in orange, and J-region in pink. We used mutual information analysis and mixed effects model comparisons to select 606 nonredundant TCR features that best explained variance in T cell state. We fit mixed effects logistic regression models for 70% of the data in the discovery and replication cohorts separately, and combined the effect sizes for each TCR feature across the two cohorts by meta-analysis. TiRP was calibrated to include only 208 of the 606 TCR features that had Bonferroni-significant meta-analytic P values. (b) We then applied TiRP to the TCRs to tumor-infiltrating CD4⁺ cells in order to study mixed clones: groups of T_regs and T_convs with the same TRB and TRA sequences observed in the same individual. These mixed clones likely represent lineages of T cells that have undergone a peripheral conversion between the regulatory and conventional phenotypes. Such clones may include induced or iT_regs (T_conv cells that have acquired a regulatory phenotype), exT_regs (T_reg cells that have lost the regulatory phenotype), or both. (c) Finally, we investigated the drivers of TiRP by separately examining the two elements of the human T_reg TCR ligand: the self-peptide and the human MHC II molecule. Figure created with BioRender.com.

Figure 2.. TCR sequence structure.

(a) Probability of each amino acid in each CDR3β position depicted by a sequence logo, with a heatmap of normalized mutual information (NMI) between each pair of CDR3β residues for the most frequent CDR3β length, 15 amino acids. Based on this mutual information structure, we partitioned the CDR3β sequence into a Vmotif within a V-region, a CDR3β middle region (CDR3βmr), and a Jmotif within a J-region. (b) Schematic showing TCRs of multiple lengths aligned to the TCR β chain structure. Three complementary-determining regions within the TCR β chain protrude as loops into the pMHC-TCR complex: CDR1β, CDR2β, and CDR3β. CDR1β and CDR2β are encoded by the TRBV gene, while CDR3β spans TRBV-encoded residues, random nucleotide insertions (CDR3βmr) and TRBJ-encoded residues. Random nucleotide insertions from VDJ recombination occur at the V/D and D/J junctions, creating variation in CDR3βmr length. Regions suggested by mutual information structure are not drawn to scale. NMI: Normalized mutual information

Figure 3.. Broad differences exist between the TCRs of T_regs and T_convs.

(a) Percentage of select amino acids in the CDR3βmr, plotted as the mean for each donor sample in the discovery cohort, separated by cell type and colored by amino acid groups. P values are computed by a two-sided Wald test on the coefficient for each amino acid term in a mixed effect logistic regression model (Methods). (b) Incremental variance explained by the addition of labeled TCR features to the V-region (left), CDR3βmr (middle), and J-region (right) mixed effect logistic regression models. The addition of each TCR feature increased model complexity by adding one degree of freedom for each quantitative feature and k - 1 degrees of freedom for each qualitative feature, where k is equal to the number of possible values for the qualitative feature (k = 58 for 58 possible TRBV genes; k = 8 for 8 possible Vmotifs). For each region, the primary modeling approach was compared to the alternative modeling approach, and the modeling approach that explained greater variance was selected. Colored horizontal lines depict the total percent of explained variance attributable to each TCR region, summing to 100%. (c) Percent of explained variance by each TCR feature type, summing to 100% for each length of CDR3β. (d) Variance explained by each TCR region for different CDR3β lengths. As CDR3β length increases, CDR3βmr occupies a greater proportion of the TCR (fraction of amino acid residues), at the expense of V and J region proportions. Line of best fit is drawn for each TCR region; 95% confidence interval shaded in gray, with each point is labeled by CDR3β length. X-axis corresponds to the proportion of TCR β chain amino acids derived from the V, J, and middle regions (summing to 100 for each CDR3β length, Methods), while the Y-axis corresponds to the absolute variance explained (scale: 0 −100%). VGSR = V gene selection rate (Supplementary Note). CDR3βmr %AAs = percent composition of amino acids in the CDR3βmr.. VGSR = V gene selection rate (Supplementary Note). CDR3βmr %AAs = percent composition of amino acids in the CDR3βmr.

Figure 4.. T_regs exhibit position-specific TCR sequence features.

(a) Estimated odds ratio (per standard deviation) for each physicochemical feature at each CDRβ(1–3) loop position; features with an estimate > 1 are positively associated with T_reg fate while features with an estimate < 1 are negatively associated. Odds ratios denote the change in T_reg odds per standard deviation increase in the given physicochemical feature at the given TCR position. Within each CDR3β length, all effects were estimated jointly in an L2-regularized logistic regression with a penalty weight tuned via 10-fold cross-validation (Methods). Shown are the odds ratio estimates for each position-feature averaged across the six CDR3β lengths. Vertical lines denote the boundaries of each CDRβ loop. (b) Correspondence between TRBV gene isoelectric point at p37 (apex of CDR1β) and TRBV gene odds ratio for T_reg fate compared to the reference gene, TRBV05–01. Each TRBV gene is labeled with its amino acid residue at p37 and the 95% confidence interval for its odds ratio. (c) Distribution of CDR3βmr hydrophobicity in T_convs compared to T_regs in the discovery dataset. Hydrophobicity values are averaged over the CDR3βmr for each TCR, and then scaled to have mean 0 and variance 1. Horizontal lines depict mean for each population (T_reg mean CDR3βmr hydrophobicity = 0.05, T_conv mean hydrophobicity = −0.03, Wald test P value = 2.3 × 10⁻⁵²³). (d) Sequence logo depicting the effects of amino acids in the highly entropic CDR3βmr residues, sized proportionally to the associated change in T_reg odds, with amino acids more frequent in T_regs above the horizontal line and amino acids more frequent in T_convs below.

Figure 5.. T_reg TCR sequence biases replicate in independent cohorts.

(a) Correspondence between the discovery and replication cohort odds ratios for CDR3βmr compositional amino acids (AAs); OR corresponds to the change in T_reg odds associated with one standard deviation (SD) increase in CDR3βmr percentage for a given AA. Colors for amino acids correspond to Extended Data Figure 1h. (b) Comparison in (a) for all other TCR sequence features; OR corresponds to the change in T_reg odds associated with the presence of the given feature compared to the reference feature (Supplementary Table 1). For (a)-(b), R = Pearson’s correlation coefficient and P values are computed by a two-sided t-test with Fischer transformation. (c) Validation of the TCR-intrinsic regulatory potential (TiRP) score in held-out donors of the discovery and replication datasets (n = 3,277,036 TCRs). Each SD increase in TiRP was associated with a 23% increase in the odds of T_reg status (OR: 1.231, 95% CI: 1.227 – 1.235, likelihood ratio test (LRT) P = 2.4 × 10⁻³²⁴⁸). Percentile points are colored by T_reg:T_conv ratio ranging from blue (lowest) to purple (highest). (d) Validation of TiRP in scRNAseq of CD4⁺ tumor microenvironment T cells^, (n = 27,721 cells). Each unit increase in TiRP (corresponding to one SD for the scores in 5c) was associated with a 16% increase in the odds of T_reg status (OR: 1.16, 95% CI: 1.13–1.19, LRT P = 4.0 × 10⁻²⁵). (e) Validation of TiRP in human thymic T cells (n = 60,424 cells). Among developing thymocytes, each unit increase in TiRP was associated with a 9% increase in the odds of T_reg fate (OR: 1.09, 95% CI: 1.05 – 1.13, LRT P = 8.8 × 10⁻⁷). For (d) and (e), error bars outline 95% confidence intervals for T_reg/T_conv odds in each TiRP score decile, computed by bootstrap resampling (Methods). (f) Validation of TiRP in TCR-targeted gDNA sequencing from grafted human thymi of humanized mice (n = 466,551 TCRs). Each unit increase in TiRP was associated with a 12% increase in the odds of T_reg status (OR: 1.12, 95% CI: 1.11–1.12, LRT P = 3.1 × 10⁻¹⁷⁷).

Figure 6.. TiRP helps to explain clonal plasticity in the tumor microenvironment.

(a) Reference T cell dataset, colored by cell type clusters according to transcriptional and surface marker variation depicted in Extended Data Figure 7c−d. (b) Select gene expression (FOXP3, GZMB) and surface marker abundance (CD25, CD127) for cells in the reference T cell dataset (low = purple, high = light green). (c) Tumor microenvironment T cells of expanded clones mapped into the reference embedding by Symphony. Each cell is colored by the TiRP score of its paired TRB chain, with KNN smoothing for visualization (Methods). TiRP is scaled such that 0 corresponds to the mean score and one unit corresponds to one standard deviation of held-out bulk sequencing TCRs (Figure 5c). (d) Cell members of three example mixed clones are highlighted in color according to their cell type classification by Symphony (colors as in (a)). Within a given plot, each cell expresses the same CDR3β DNA sequence, the same CDR3α amino acid sequence, and was observed within the same donor (CDR3β amino acid sequence listed above CDR3⍺ amino acid sequence for each). (e) Same as (c), with each cell colored according to clone type: purple for clones containing only T_reg cells, blue for clones containing only T_conv cells, and yellow for clones containing both T_reg and T_conv cells (“mixed” clones). (f) TiRP scores of T_conv, T_reg, and ”mixed” expanded clones from held-out bulk sequencing data. P = 2.0 × 10⁻⁴⁰ for mixed-T_conv difference, P = 9.1 × 10⁻¹⁶ for mixed-T_reg difference. (g) Scores as in (f) for tumor-infiltrating scRNAseq data. P = 3.0 × 10⁻⁴ for mixed-T_conv difference, P = 0.55 for mixed-T_reg difference. For (f) and (g), vertical bars denote mean and standard error of the mean per clone type. (h) Correspondence between TiRP score and the T_reg:T_conv ratio for each clone. Best fit line is shown in gray; clones are colored by T_reg:T_conv ratio and sized proportionally number of constituent cells. β corresponds to the slope of the regression line between the log-transform of the T_reg:T_conv ratio and TiRP score. For (f)-(h), P values are computed by the LRT between mixed effect logistic regression models (Methods).

Figure 7.. Two axes of TCR-driven cell states.

(a) 67 samples from the replication cohort colored by cell type and arranged by principal component space according to variation in TCR sequence feature frequencies (Methods). (b) Distribution of PC1 embeddings for each cell type; each vertical line corresponds to one sample. Naive T_convs have the highest PC1 embedding in 15 of the 16 donors with all three cell types available. P value is computed by the binomial test with n = 16 and k = 15. (c) Percent contribution of each type of TCR sequence feature to the first two principal components. (d) Loadings of each of the TCR sequence features on PC1 and PC2, depicted by arrows, separated by TCR region and colored by the same scheme as in (c). (e) Samples arranged in PC space as in (a), colored by mean TiRP in the V-region of the TCR (vTiRP). (f) Same as in (e), colored by mean TiRP in the CDR3βmr (mTiRP). P values for (e)-(f) are calculated by a two-sided t-test with Fischer transformation on Pearson’s R. jTiRP = TiRP (T_reg-intrinsic regulatory potential) of the J-region of the TCR (IMGT positions 113–118) mTiRP = TiRP (T_reg-intrinsic regulatory potential) of the middle region of the TCR (IMGT positions 108–112) vTiRP = TiRP (T_reg-intrinsic regulatory potential) of the V-region of the TCR (IMGT positions 1–107)

Figure 8.. Isolating the drivers of TiRP.

(a) We investigated the drivers of TiRP by separately examining the two elements of the human T_reg TCR ligand: the self-peptide and the human MHC II molecule. To do so, we scored 1) murine T_reg TCRs, which share an affinity to mammalian self-peptides but not to human MHC II molecules, and 2) human memory T_conv TCRs, which share an affinity to human MHC II molecules but not to self-peptides. (b) Left: mean increase in TiRP score of Helios-sorted T_regs compared to naive T_convs in Helios-GFP Foxp3-RFP reporter mice. Right: mean increase in TiRP score of memory T_convs compared to naive T_convs from held-out donors of the replication dataset. (c) Left: TiRP score increases in Helios-sorted murine T_regs broken down into TiRP score components by TCR region. Right: TiRP score increase in human memory T_convs broken down into TiRP score components by TCR region. (d) Correspondence between TCR feature odds ratios for T_reg-T_conv odds (x-axis, meta-analytic odds between discovery and replication cohort), and memory-naïve odds (y axis, replication cohort only) with their 95% confidence intervals. TRBV genes are highlighted in green; V06–01 indicates TRBV06–1; V25–01 indicates TRBV25–01. Pearson’s R is calculated with respect to TRBV gene odds ratios only. P values in (b)-(c) are calculated by the LRT between mixed effects models (Methods); P value in (d) is calculated by a two-sided t-test with Fischer transformation on Pearson’s R. jTiRP = TiRP (T_reg-intrinsic regulatory potential) of the J-region of the TCR (IMGT positions 113–118) mTiRP = TiRP (T_reg-intrinsic regulatory potential) of the middle region of the TCR (IMGT positions 105–112) vTiRP = TiRP (T_reg-intrinsic regulatory potential) of the V-region of the TCR (IMGT positions 1–104) Figure created with BioRender.com.