Tumor neoantigen recognition is likely a major factor in the success of clinical immunotherapies1,2,3,4,5 and patients with high tumor mutational burden (TMB)6 and significant number of neoantigens benefit more from immune checkpoint blockade as well as adoptive cell therapy (ACT) using natural tumor-infiltrating lymphocytes (TILs)4, 7. TIL therapy also showed success in low TMB cancers8,9,10,11. Unlike T cell clones that recognize TAAs, which are self-antigens, those recognizing neoantigens are presumably not subject to negative thymic selection12. Consequently, neoantigen-specific T cells may be of higher functionality13 and their superior antigen sensitivity was recently demonstrated14. T cell functionality is partly determined by the avidity of TCRs for their cognate pMHC. Because this is dictated by structure, it is referred to as structural avidity and it is determined through the dissociation kinetic of monomeric pMHCs and TCRs15. Conversely, sensitivity to antigen reflects the properties of the TCR but also the functional state of the cells, and is thus referred to as functional avidity16. Studies in mice and humans indicate that structural and functional avidities of CD8 T cells correlate15 and determine T cells performance17.

In TIL-ACT, clinical efficacy has been correlated with the persistence of adoptively transferred TIL clones in vivo4,18, linked to unique gene expression patterns19. However, how avidity affects tumor engraftment of tumor-specific T cells is presently not well understood. Yet, this is a key parameter affecting the success of T cell-based immunotherapy.

In this work, we evaluate broadly the structural avidities and antigen sensitivities of neoantigen-specific T cells and ask whether these correlate with the cell aptitude for tumor infiltration and homing. We profile a large library of CD8 T cells specific for neoantigens, tumor-associated antigens and virus epitopes from tumors and peripheral blood from healthy donors and patients with melanoma, ovarian, lung or colorectal cancer. Although neoantigen-specific T cells exhibit superior avidity than TAA-specific cells as expected, a wide range of avidities is observed. High clonotype avidity is specifically associated with tumor residence at steady state, higher CXCR3 expression and tumor engraftment following ACT in mice. Finally, we show that high-avidity TCRs share biophysicochemical properties and this allows us to generate an in silico predictor of TCR avidity. We imply a direct relationship between the strength of antigen recognition, CXCR3 expression and tumor infiltration, and provide a functional parameter for screening neoantigen-specific T cells for ACT.


Neoantigen-specific CD8 T cells are structurally and functionally heterogeneous

Neoepitopes are generally considered as prototypical tumor rejection antigens. Yet, it remains unclear whether their clinical relevance stems from their tumor specificity alone or whether they truly drive better effector T cells relative to TAAs. To learn more, we generated a library of 371 CD8 T cell clones recognizing 19 neoantigens, TAAs and virus epitopes (Supplementary Table 1) from 16 patients with melanoma, ovarian, lung or colorectal cancer and 6 healthy donors (Supplementary Table 2), and investigated the functional and structural profiles of their TCRs in 190 and 338 clones, respectively (Fig. 1a). Antigen-specific cells were sorted using double-fluorescent reversible pMHC multimers (i.e. NTAmers), which avoid the selective loss of high-avidity cells20.

Fig. 1: Structural avidity of neoantigen-, TAA- and virus-specific CD8 T cells.
figure 1

a Neoantigen- and TAA-specific CD8 T cells were purified from in vitro expanded CD8 T cells of melanoma, ovarian, lung or colorectal cancer patients and virus-specific CD8 were isolated from healthy donors and cancer patients. After single-cell cloning and expansion, individual clones were subjected to antigen sensitivity and structural avidity measurements as well as TCR sequencing. Molecular modeling of pMHC-TCR interactions were also performed. b Representative examples of structural avidity of virus-, TAA- and neoantigen-specific CD8 T cells. Structural avidity was determined with reversible pMHC multimers to measure monomeric pMHC-TCR T1/2. c Structural avidity of individual virus-, TAA- or neoantigen-specific CD8 T cells (mean ± SEM). The number of clones is indicated in brackets (each clone was tested individually). d Cumulative structural avidities per classes of antigen- (virus, TAAs and neoantigen)-specific CD8 T cells. The number of clones is indicated in brackets. P values are provided when significant at 95% confidence interval and using two-sided Mann–Whitney test. e Coefficient of determination R2 of the regression analyses (Supplementary Fig. 4a–f) between pMHC binding/stability and immunogenicity predictors values and the medians of T1/2 (s) or EC50 (M) of antigen-specific CD8 T cells. Pearson coefficients (two-sided test) were calculated and mentioned when significant. Patients and clones are described in Supplementary Tables 1 and 2. Source data are provided as a Source Data file.

We first assessed T cell structural avidity, intended as the strength of TCR binding to cognate pMHC. This was determined through the dissociation kinetic (pMHC-TCR half-life, T1/2) of monomeric pMHCs and TCRs, as we described previously15. Briefly, rapid decay of reversible pMHC multimers to pMHC monomers allows dissociation rate measurements of fluorescent monomeric pMHC off CD8 T cells. We detected polyclonal responses against individual epitopes of any class in most patients or donors, with marked variance of T1/2 among clones recognizing the same epitope in each class of antigen specificity (Fig. 1b, c). Overall, the structural avidities of neoantigen-specific CD8 T cells were higher than that of TAA-specific CD8 T cells (Fig. 1d). Similar conclusions were drawn when we examined exclusively HLA-A*0201-restricted CD8 T cells (Supplementary Fig. 1a) or unique CDR3 sequences (Supplementary Fig. 1b-c and Supplementary Table 1). This supports the long-proposed hypothesis that neoepitope-specific TCRs are of higher structural avidity than T cells directed against “self” tumor antigens21,22.

We next assessed the antigen sensitivity of each clone by IFNγ ELISpot, measuring the peptide concentration required for half-maximal T cell activation (effect concentration 50%, EC50, Supplementary Fig. 2a, b). Similar to pMHC-TCR T1/2, we observed an important variance in EC50 among different clones recognizing the same peptide for each class of antigens, including for HLA-A*0201-restricted T cell responses (Supplementary Fig. 2c–e) or genetically unique clonotypes (Supplementary Fig. 2f, g). As expected, a positive correlation was observed between T1/2 and EC50, despite some variability (Supplementary Fig. 3a). We attribute the latter to more reproducible and reliable measurements obtained by pMHC-TCR dissociation kinetics when compared to functional assays, which also depend on T cell intrinsic regulatory mechanisms17,23. To further test the reproducibility of the structural avidity parameter, we cloned six TCRαβ chain pairs into healthy peripheral blood T cells; measurements of structural avidity remained more consistent between original and recipient T cells, maintaining similar ranking between clones, as opposed to antigen sensitivity (Supplementary Fig. 3b–e and Supplementary Material 1). This supports the robustness of structural avidity as a biophysical parameter to profile T cells.

We used an in vitro pMHC refolding assay24 to validate the predicted affinity of each peptide for the cognate HLA allele. The overall ranges of pMHC affinity ruled out any important bias in measurements of antigen sensitivity due to low peptide-MHC interactions. Highlighting the limitation of commonly used algorithms for predicting epitope immunogenicity, we found poor correlations between measured structural avidity (or antigen sensitivity) with in silico predictors of pMHC affinity, stability or processing, mainly relying on the determination of antigen presentation (Fig. 1e and Supplementary Fig. 4)25,26,27,28. However, structural avidity was significantly correlated with immunogenicity predicted by PRIME29 and with pMHC Dissimilarity-to-Self (DisToSelf)30 (Fig. 1e), also significant when viral epitopes were excluded (Supplementary Fig. 4a–f) and when only genetically unique clonotypes were considered (Supplementary Fig. 4g). PRIME not only considers the binding capacity of a peptide to a given MHC but also integrates its propensity to be recognized by TCRs29. DisToSelf determines the similarity (or dissimilarity) of a given peptide with the human proteome30. Peptides with high DisToSelf scores are recognized by higher avidity T cells (Fig. 1e and Supplementary Fig. 4g).

High structural avidity neoantigen-specific CD8 T cells reside in tumors

Given the unveiled heterogeneity of TCR avidities for given tumor epitopes, we asked whether TCR strength discriminates cells with a propensity for tumor infiltration. Indeed, if higher avidity cells were to carry an antitumor response, they would be expected to be rather enriched in the tumor microenvironment31,32. Strikingly, TILs recognizing neoantigen- or TAA-epitope exhibited significantly superior antigen sensitivity relative to cognate peripheral blood lymphocytes (PBLs) recognizing the same epitope across melanoma, ovarian, colorectal and lung cancer patients (Supplementary Figs. 5 and 6).

To assess whether differences in antigen sensitivity could be attributed to structural avidity attributes of TIL vs. PBL clones (Fig. 2a), we analyzed seven pairs of tumor-specific T cells originating from TILs or PBLs. We found that the structural avidity of TILs was significantly higher than that of cognate PBLs across all studied cancers (Fig. 2b, c and Supplementary Fig. 6a, b). Thus, antigen-specific T cells infiltrating tumors, particularly neoantigen-specific clones, display stronger structural avidity than their blood counterparts, including when genetically unique clonotypes are considered (Fig. 2d and Supplementary Fig. 6c, d).

Fig. 2: Association between structural avidity and tumor tropism.
figure 2

a Correlation between the structural profile of antigen-specific CD8 T cells and tropism at steady state. b Representative examples and cumulative analyses (Mean ± SEM) of monomeric pMHC-TCR dissociation kinetics of MMP9-specific PBLs and TILs assessed with NTAmers (reversible pMHC multimers). The number of clones is indicated in brackets (each clone was tested individually). P values are provided at 95% confidence interval and using two-sided Mann–Whitney test. c Comparison of the structural avidity of seven pairs of PBLs and TILs recognizing the same pMHCs. The number of clones is indicated in brackets (each clone was tested individually). Wilcoxon two-sided test was used to determine the P value. d Structural avidity of TAA- and neoantigen-specific PBLs and TILs. The number of clones is indicated in brackets. P values are provided at 95% confidence interval and using two-sided Mann–Whitney test when significant. e UTP20-specific CD8 T cells from patient Lung1 were sorted from TILs and PBLs using NTAmers, bulk TCR sequenced and cloned by limiting dilution. The Manhattan plots of TCRα repertoires are shown and only clonotypes identified in both PBLs and TILs repertoires are color-coded. f Monomeric pMHC-TCR dissociation kinetics of three UTP20-specific clones of patient Lung1 assessed with reversible pMHC multimers (NTAmers). g Relative frequency of clones 1, 3 and 5 among UTP20-specific CD8 TILs (left) and PBLs (right). Structural avidity for each clone is also plotted. h Superimposition of in silico analyses of the pMHC-TCRs molecular interactions for UTP20-specific clones 5 and 1. Pink and green are used to color TCR ribbons, MHC (shaded color) and peptides (ball and stick) for clones 5 and 1, respectively. Source data are provided as a Source Data file.

To better understand the relative enrichment of TILs in high-avidity cells, we sequenced the TCRs of sorted primary CD8 PBLs and TILs recognizing the same neoepitope from the UTP20 protein from patient Lung1. Neoantigen-specific T cells were oligoclonal, but only three TCRs were shared between PBLs and TILs (Fig. 2e). Remarkably, clonotype 5, which was dominant in TILs (58.8% of neoepitope-specific cells), was only contributing to 1.4% of the PBL repertoire, while clonotype 1 was less frequent in TILs (9.6%) but dominant in PBLs (26.1%), and clonotype 3 showed similar frequency in TILs (17.2%) and in PBLs (16.2%). Interestingly, the structural avidity of UTP20-specific TCRs (Fig. 2f) correlated with their frequency in the tumor compartment, and was the highest for clonotype 5, indicating that tumor-resident clones have higher structural avidity (Fig. 2g). We previously showed31 that molecular modeling of TCR and pMHC can accurately infer the strength of their interaction. Here we confirmed that clone 5 TCR established significantly more favorable interactions with UTP20 pMHC than clone 1 TCR (Fig. 2h and Supplementary Table 3). Similar results were obtained in a second example in patient CRC1 for PHLPP2-specific TCRs confirming the association between structural avidity and tumor residence (Supplementary Fig. 7a–d). These observations indicate that the preferential accumulation of high and low-avidity clones in tumors and blood, respectively, is also true among clonotypes from the same antigen-specific repertoires.

To experimentally validate the preferential tumor infiltration by high structural avidity T cells (Fig. 3a), we took advantage of a well-characterized panel of NY-ESO-1157165-specific TCRs with high (DMβ), intermediate (WT) and low (V49I) structural avidity. Their avidity covers the range of viral-, neoantigen- and TAA-specific T cells33,34,35. We stably transduced CD8 T cells of an HLA-A*0201 donor with DMβ, WT or V49I TCRs and profiled their structural and functional avidities (Fig. 3b). Unlike V49I-transduced T cells, both WT and DMβ variants showed equivalent in vitro responsiveness to HLA-matched Me275 melanoma tumor expressing NY-ESO-1 (Fig. 3b). ACT of 5 × 106 T cells in interleukin-2 (IL-2) NOG mice bearing Me275 tumors indicated a correlation between the in vivo efficacy and the structural but not the functional avidity of TCR-transduced T cells (Fig. 3c). Following ACT, DMβ-transduced CD8 T cells significantly better infiltrated tumors as compared to V49I- and WT-transduced cells (Fig. 3d), confirming higher engraftment propensity of high-avidity clones.

Fig. 3: Tumor infiltration of high-avidity clones is associated to CXCR3 expression.
figure 3

a Preferential tumor infiltration by high-avidity clones and CXCR3-mediated tumor homing was validated by in vivo ACT in mice and in four melanoma patients receiving T cell therapy. b Reactivity of V49I, WT and DMβ-transduced T cell was measured through IFN-γ secretion upon coculture with Me275 tumor cells (right, n = 2 independent experiments, Mean ± SEM) and monomeric pMHC-TCR dissociation kinetics of the three mutants were determined using reversible pMHC multimers (left, n = 2 independent experiments). c Me275 tumor growth in IL-2 NOG mice adoptively transferred at day 7 post-tumor engraftment with 5 × 106 primary CD8 T cells transduced with V49I, WT or DMβ NY-ESO-I157–165-specific TCRs (n = 1 independent experiment, Mean ± SEM). Log-rank two-sided tests were used to determine P values. d Representative examples of Me275 tumor sections, harvested from engrafted IL-2 NOG mice at day 8 post-ACT with 5 × 106 V49I, WT or DMβ-transduced T cells (n = 2 independent experiments). Tumors were stained for SOX10, PD-1 and CD8. DAPI was used to stain nuclei. For tumor infiltration by CD8 T cells (cells/mm2), bounds of box are 25th to 75th percentiles with median, whiskers are min to max. Mann–Whitney two-sided test was used to calculate P values. Analyses were performed using Inform v2.3.062. e Me275 tumor growth in IL-2 NOG mice adoptively transferred with 2 × 106 DMβ-transduced primary CD8 T cells at day 5 and co-injected or not with anti-CXCR3 blocking antibody (100 μg at day 5 and day 10) (n = 2 independent experiments, Mean ± SEM). Log-rank test was used to determine P value. f Quantitative measurement of tumor infiltration by CD8 T cells (cells/mm2) 10 days post-ACT of 2 × 106 DMβ-transduced primary CD8 T cells co-injected or not with anti-CXCR3 blocking antibody (n = 2 independent experiments). Bounds of box are 25th to 75th percentiles with median, whiskers are min to max. Mann–Whitney two-sided test was used to calculate the P value. Source data are provided as a Source Data file.

CXCR3-mediated tumor infiltration and control by high-avidity T cells

Having established a relationship between T cell avidity and tumor homing, we hypothesized that high-avidity cells may be endowed with a superior ability for tumor infiltration and retention (Fig. 3a). Several studies reported that key chemokine receptors, especially CXCR3, may be required for tumor homing36. We analyzed the expression of a panel of chemokine receptors on seven pairs of low and high-avidity antigen-specific CD8 T cells. CXCR3 was more strongly expressed and upregulated after short-term stimulation by high as compared to low-avidity T cell clones (Supplementary Fig. 8a). This observation was specific to tumor homing-related molecules since no significant difference was found for chemokine receptors that are not specifically involved in tumor infiltration (e.g. CCR7). In addition to CXCR3, CD103 and CD49a (VLA-1), two major integrins associated with a tissue-residency phenotype37,38,39, were both upregulated in high-avidity clones (Supplementary Fig. 8b). Therefore, T cell structural avidity is associated to CXCR3 expression and, to a lower extend, to CD103 and CD49a expression.

Higher CXCR3 expression was also observed on DMβ- relative to V49I- or WT-transduced T cells upon pMHC stimulation in vitro (Supplementary Fig. 8c). Of interest, addition of an anti-CXCR3 antibody after ACT in IL-2 NOG mice bearing the Me275 melanoma tumor, known to express CXCR3 ligands i.e. CXCL9/10/1140, with DMβ-transduced T cells (Fig. 3a) significantly impaired tumor control (Fig. 3e). Consistently, lower densities of CD8 T cells were observed in animals treated with an anti-CXCR3 blocking antibody post-ACT (Fig. 3f). The inhibition of tumor control after ACT by blocking CXCR3 was further demonstrated in two additional models using neoantigen-specific TCRs (Supplementary Fig. 8d, e). This confirms the contribution of CXCR3 in tumor homing and mechanistically links CXCR3 expression with high-avidity clones and tumor infiltration.

Biophysicochemical inference of tumor-specific T cells that engraft in tumors

The above findings collectively suggest that tumor-infiltrating lymphocytes are enriched in tumor-specific T cell clones endowed with high-avidity TCRs. Inspired by prior demonstration that differences in the antigen sensitivity of T cell clones targeting a given pMHC correlates with the strength of pMHC-TCR binding, specifically the number of atomic contacts between TCR and pMHC inferred by molecular modeling31, we sought to develop further methods to infer the avidity of clones for a given epitope (Fig. 4a). We used homology modeling (see methods) to compare TCRs recognizing the same pMHC with high or low structural avidity, applied to five distinct antigens. The number of favorable interactions (bonds) of each TCR with its cognate pMHC, inferred based on the modeled structures of its α and β chains and the cognate pMHC, was consistently higher for high structural avidity TCRs (Supplementary Fig. 9a and Supplementary Table 3), and significantly correlated with pMHC-TCR T1/2 (Supplementary Fig. 9b).

Fig. 4: Tumor infiltration after ACT correlates with predicted structural avidity inferred from TCR clustering analyses.
figure 4

a Computational analysis of TCR features led to the establishment of a predictor of TCR avidity and its application on patients’ TIL-ACT products allowed tracking of predicted low and high-avidity TCRs in post-ACT tumor samples. b Hierarchical clustering of 58 TCR sequences provided based on a biophysical approach43. TCRs sharing the closest 4-mer features are next to each other and TCRs recognizing the same pMHC have the same color code. TCR model numbers are presented as labels and further details about TRAV, TRAJ, TRBV, TRBJ, HLA and peptide are found in Supplementary Table 4. The structural avidity of each cognate TCRs is represented below (mean of n = 3 independent experiments). The black dashed box highlights a region where high-avidity TCRs recognizing multiple pMHC specificities are clustering. c Cumulative analysis for four melanoma patients of the percentage of predicted high-avidity CD8 T cells in blood and tumor samples. Values for individual patients are plotted (gray and black) as well as the cumulative analysis (in red) for which the P value was calculated as described in the method section. The number of clones is indicated below for each patient individually. d Monomeric pMHC-TCR dissociation kinetics of Jurkat cells transfected with neoantigen KIF1BS918F-specific TCR#1 and TCR#2, respectively predicted as high and low-avidity TCRs. e Autologous (Mel8) tumor growth in IL-2 NOG mice adoptively transferred at day 22 with 5 × 106 primary T cells transduced with KIF1BS918F-specific TCRs (n = 1 independent experiment, Mean ± SEM). Log-rank two-sided test was used to determine the exact P value. Source data are provided as a Source Data file.

A major limitation in identifying clinically relevant T cells is the lack of knowledge of possible cognate antigens. To solve this, we hypothesized that high-avidity TCRs may share common sequence features (Fig. 4a). Furthermore, it has been reported that highly frequent clones among TILs may not be tumor-specific41. To overcome these limitations, and driven by the above molecular modeling results, we asked whether we could infer specifically high structural avidity TCRs based on their sequence analysis and without prior knowledge of their specificity. We selected 58 individual TCRs recognizing 12 distinct pMHC, for which TCR α and β sequences as well as structural avidities (Supplementary Table 4) were known, and looked for structural patterns. We used biophysical features of k-mers encoded based on the Atchley factors and a generic hierarchical clustering algorithm42,43. We found that CDR3β sequences in high-avidity TCRs (T1/2 > 60 s) were significantly enriched in specific amino acid residues (i.e. N, E, I, K, T, Y, V; all P < 0.0001 compared to low-avidity TCRs). Conversely, A, R, D, L, M and P were more frequent in low-avidity TCRs (all P < 0.0001) (Supplementary Fig. 10 and Supplementary Table 5).

We next developed hierarchical clustering based on CDR3β motifs and, interestingly, we identified a hotspot enriched in TCRs with high structural avidity, irrespectively of their target (Fig. 4b, dashed black box). This comprised 62% of all TCRs of intermediate or high avidity (T1/2 > 10 s), while outside of this cluster, 65% of TCRs had structural avidity <10 s. Such enrichment was not observed in a control analysis with 1000 random clustering, illustrating the significance of this observation (P < 0.001) and indicating that some shared common CDR3β features were preferentially associated with higher structural avidity (Supplementary Material 2).

Guided by this observation, we derived a structure-based logistic regression model to predict the structural avidity of TCRs of unknown specificity (Fig. 4a, “Methods” and Supplementary Material 3). We applied it to our panel of 58 TCRs and were able to accurately discriminate between high and low-avidity TCRs with an AUC of 0.96, with only one false-positive and one false-negative high avidity among the full set of TCRs (sensitivity of 0.91 and specificity of 0.97) (see “Methods”). Three cross-validations schemes were successfully performed on viral peptides, TAAs and neo-antigens and on different HLA alleles to assess the robustness of the predictor, following the standard leave-20%-out protocol (cross-validations 1 and 2) as well as a more challenging leave-one-epitope-out cross-validation (cross-validation 3). We achieved a success higher than 70% in all cases, which leaves room for improvement but is significantly better than random and gives us confidence in the algorithm for the purpose and data described herein (Supplementary Material 3). The relevance and the domain of applicability of this model will require further investigations and will most probably increase in time by incorporating additional experimental data. In the future, thanks to the additional experimental data that will be collected by us and the community, more CDRs and residues will be included in our predictor.

We then applied the structure-based logistic regression to identify high and low-avidity TCRs in blood and tumors of four additional melanoma patients (Supplementary Table 2). When analyzing total tumor and blood TCR repertoires, we consistently found an enrichment in TCRs predicted to be of high avidity in tumors relative to blood (Fig. 4c, P = 0.05), therefore validating the preferential tumor tropism for high-avidity TCRs. Furthermore, we also identified two TCRs directed against neoantigen KIF1BS918F in enriched TILs44 from patient Mel8 (Supplementary Fig. 11a and Supplementary Table 2). Among the two KIF1BS918F TCRs, and despite no differences in functional avidity (Supplementary Fig. 11b), TCR#1 was predicted by our logistic regression to be of high avidity, while TCR#2 was predicted to be of low avidity and these predictions were validated experimentally (Fig. 4d). We then took the opportunity of the availability of autologous tumor cells to assess the relative clinical efficacy of the two KIF1BS918F TCRs. Despite the fact that both TCRs were tumor-reactive in vitro (Supplementary Fig. 11c), tumor control in mice bearing the autologous tumor was exclusively achieved after ACT with T cells transduced with the predicted and validated high-avidity TCR but not with the low-avidity TCR (Fig. 4e). This prototypical example illustrates the superior tumor infiltration and control by high-avidity cells, even when they target the same neoepitope but also highlights the clinical relevance of the logistic regression to predict clinically relevant TCRs regardless of their specificity.


Success of TIL-based immunotherapies for solid tumors relies on strong antitumoral activity of adoptively transferred T cells. Major efforts are thus made to develop methods allowing to estimate a priori the functional potential of tumor antigen-specific T cells. The strength of T cell recognition is a key parameter, as it affects T cell activation, proliferation, infiltration and effector functions, as well as longevity of T cell responses17,45. Besides the structural avidity of the TCR, multiple coreceptors are implicated in determining T cell functional avidity46,47. Cellular assays (cytotoxicity or cytokine production) have been traditionally used to determine antigen sensitivity for which EC50 represents a widely accepted parameter. However, cellular assay results depend on the state of cellular activation or exhaustion, limiting their performance17. The dissociation kinetic measurement of pMHC from the TCR, a structural avidity parameter reflecting the binding strength of a TCR15, can be readily applied on viable T cells. Reversible pMHC multimers can be used to reliably determine such dissociation kinetics, showing T cell functionality independently of the cellular activation state17.

Here we comprehensively profiled tumor antigen-specific T cells in patients with solid tumors and compared them with virus-specific cells. To do so, we generated 371 T cell clones upon FACS-sorting with reversible pMHC multimers, precluding the loss of high-avidity T cells prone to TCR-induced cell death induced by conventional pMHC multimers20. As expected, we found that virus-specific T cells show higher avidity and function than TAA-specific cells48. Neoantigen-specific T cells displayed higher structural avidity than TAA-specific T cells. The superiority of neoantigen-specific T cells over TAA-specific ones was not recapitulated by IFNγ release. Functional avidity assays largely rely on T cell activation states and are more prone to intra- and inter-experimental fluctuations17. T cell responses of high-avidity clonotypes can also be inhibited by exhaustion mechanisms49. This legitimates structural avidity as a robust and reliable biomarker of T cell responsiveness. However, a broad heterogeneity was observed in the avidity of neoantigen-specific T cells, ranging from low avidities, comparable to TAAs-, to high avidities, comparable to virus-specific cells. Also seen by others50,51, we found a correlation between TCR avidity and T cell potency for the same antigen but also across antigen specificities, providing the rationale for developing prioritization algorithms in TCR discovery. Of note, the strength of interaction between effector cells and cognate antigen is becoming an attractive parameter to measure T cell activation and predict efficacy52. Interestingly, the importance of the binding strength between effector cells and tumors was also recently demonstrated with CAR T cells53.

The broad heterogeneity of TCR avidities supports the notion that neoantigens may strongly differ in their potential to mediate antitumor effects in vivo. Identification of neoantigens relies on in silico prediction of antigen binding avidity to MHC molecules, with a discovery rate <5%, arguing that only a minor fraction of presented peptides are immunogenic31,41,54. Indeed, we did not find any correlation between functional or structural parameters and prediction of peptide binding affinities, but did so with immunogenicity prediction through PRIME29. This presumably reflects the importance of the mutation occurring at MHC anchor residues or directly in those in contact with the TCR, which is ultimately captured by molecular modeling29,55.

It has been reported that clonally expanded T cells can reside in tumor tissue and adjacent normal tissue or blood56. Here, by analyzing T cells from blood and tumor targeting the same tumor antigen, we found that the latter consistently show higher antigen sensitivity and structural avidity. However, while TILs are enriched for high-avidity T cell populations, common TCR clonotypes were also identified in PBLs (albeit at lower frequencies), consistently with the presence of tumor-reactive TIL clonotypes in the circulation32. The association between structural avidity and tumor infiltration was seen across multiple epitopes and multiple patients, but also within antigen classes (TAAs and neoantigens) and within distinct clonotypic repertoires of neoantigen-specific T cells. Furthermore, despite the fact that T cells from both blood and tumors were systematically interrogated for each patient and each antigen, neoepitopes and TAAs were preferentially detected in TILs and PBLs, respectively, consistently with the superior structural avidity of neoepitope-specific T cells. The high frequency of neoepitope-specific TCRs in tumors was recently shown in patients with lung cancer57.

We also showed that structural avidity is associated with CXCR3 expression, known to promote tumor infiltration36,58, as well as CD103 (αEβ7) and CD49a (VLA-1) expression, both associated with tumor residency38,39. CXCR3 blocking after ACT prevented tumor infiltration of high-avidity tumor-specific T cells. Tumor infiltration and eradication probably rely on several other complementary parameters, like T cell expansion and persistence. A major issue in ACT therapy is the downregulation of antigen presentation, which could indeed favor high-avidity T cells that are less dependent on antigen concentration. This was recently suggested in melanoma patients where TAA-related antigen expression was higher than that of neoantigen and was inversely correlated with the functional avidity of the respective antigen-specific T cells14.

Our unique capacity to comprehensively profile CD8 T cells by measuring structural avidity, linked to TCR biophysicochemical and sequence features, as well as structure modeling, allowed us to build a structure-based logistic regression model of TCR avidity level prediction of TCRs with unknown antigen specificity. To prevent overfitting of the model, we limited the number of parameters entering the equation and performed successfully several cross-validation tests to assess the robustness of the approach, following the standard leave-20%-out protocol as well as a more challenging leave-one-epitope-out cross-validation. Although this predictor will require further optimization—e.g. by addition of other parameters—and validation using a larger external test set when more experimental data will become available, it allowed us to identify TCRs with high-avidity features in four melanoma patients. These were found more frequently within tumors than blood at steady state, supporting the notion that high-avidity TCRs preferentially home and reside in tumors.

Our data link neoantigen recognition, T cell functionality and ability to infiltrate and reside in tumors, suggesting that the clinical relevance of neoantigen-specific T cells is not only related to their tumor specificity but also to their higher functionality and their preferential ability to infiltrate tumors. Our data also indicate that tumor-specific CD8 T cells are highly heterogeneous and that measurements of structural avidity can be used for better selection of clinically relevant T cells, avoiding the use of poorly functional clonotypes, both for TAA- and neoantigen-specific T cells. High-avidity T cells (i.e. preferentially TILs) should therefore be prioritized for personalized therapies, including TCR-based immunotherapy.