Specific mechanisms by which tumor-infiltrating lymphocytes (TIL) become dysfunctional remain poorly understood. Here, we employed a two-pronged approach using single-cell mass cytometry and tissue imaging technologies to dissect TILs from 25 patients with resectable and 35 patients with advanced non–small cell lung cancer (NSCLC). We identified a burned-out CD8+ TIL subset (Ebo) that specifically accumulated within the tumor microenvironment (TME) but not in adjacent nontumoral tissues. Ebo showed the highest expression of proliferation and activation markers but produced the lowest amount of IFNγ and were the most apoptotic CD8+ TIL subset. Using a humanized patient-derived tumor xenograft model, we demonstrated that Ebo expansion occurred within the TME in a PD-1/B7-H1 pathway-dependent manner. Ebo abundance in baseline tumor tissues was associated with resistance to anti-PD therapy in patients with NSCLC. Our study identifies a dysfunctional TIL subset, with distinct features from previously described exhausted T cells, and implies strategies to overcome immunotherapy resistance.
We identified a highly proliferative, overactivated, and apoptotic dysfunctional CD8+ tumor-infiltrating subpopulation that is functionally distinct from previously described exhausted T cells. This population is expanded in lung cancer tissues in a PD-1/B7-H1-dependent manner, and its abundance is associated with resistance to cancer immunotherapy, thus becoming a potential tissue biomarker.
This article is highlighted in the In This Issue feature, p. 1601
Progressing tumors develop multiple cellular and molecular mechanisms to evade and/or resist immune attack. Collectively, these mechanisms are known as adaptive immune resistance. The archetype of these mechanisms is the upregulation of B7-H1 (PD-L1) in the tumor microenvironment (TME), which is primarily induced by IFNγ secretion from tumor-infiltrating lymphocytes (TIL; refs. 1, 2). These mechanisms typically result in the exclusion or functional impairment of TILs, thereby limiting endogenous antitumor immune response. Strategies such as blockade of the PD-1/B7-H1 pathway (hereafter referred to as anti-PD therapy) seek to correct and reinvigorate the impaired T-cell response against cancer (3), an approach described as normalization cancer immunotherapy (4). However, fundamental questions such as the specific mechanisms by which CD8+ T cells become dysfunctional or how anti-PD therapy overcomes this dysfunctionality remain unanswered.
One major obstacle to understanding the mechanism of T-cell reinvigorating therapies is the cellular heterogeneity within the TME. Upon antigen presentation, T cells sequentially differentiate from naïve to effector and then to T memory cells in acute inflammatory conditions (5). However, in a chronic inflammatory environment such as the TME, CD8+ T-cell differentiation could be altered or deviated, resulting in a greater diversity of subsets. Using chronic viral infection mouse models, a CD8+ subtype referred to as exhausted T cells has been described (6). Exhausted CD8+ T cells are characterized by the loss of IL2, TNFα, and IFNγ secretion; impaired proliferation; and high expression of coinhibitory receptors such as PD-1, LAG3, and TIM3 (7). A similar exhausted CD8+ T-cell subset was subsequently identified within the TME of mouse tumors and patients with melanoma that was also characterized by high PD-1, LAG3, and TIM3 expression (8). However, recent studies indicate that patient CD8+ TIL phenotype and functional profiles do not completely overlap with murine exhausted CD8+ T cells observed in chronic infection models, and neither do the factors contributing to their differentiation (9, 10). The complex and unique ecosystem of the human TME that is composed of genetic tumor diversity, spatial heterogeneity, and asymmetric alteration by anticancer therapies likely contributes to CD8+ TIL diversity and may determine unique pathways of CD8+ TIL differentiation. In this regard, detailed characterization of dysfunctional human CD8+ TILs and gaining a deeper understanding of the mechanisms leading to dysfunction are unmet needs that could be used to improve antitumor T-cell responses in patients.
Advances in high-resolution genomic and proteomic single-cell technologies allow for further dissection of the cellular diversity within TME, improving our ability to interrogate and characterize the patient's tumor–immune context. Specific advances in single-cell transcriptomics are a very powerful approach by which multiple aspects of cellular processes can be assessed (11, 12). Nevertheless, protein isoforms, posttranslational modifications, proteolytic products, and cell-surface receptor/ligand protein expression are missed by this approach. In contrast, multiparametric proteomic single-cell analysis tools such as flow cytometry (13) and more recently cytometry by time of flight (CyTOF; ref. 14–16) or Imaging Mass Cytometry (IMC; ref. 17, 18), uniquely enable investigation of cell identity and behavior at the protein level, which are largely the key executors of biological processes. Clear understanding of the human immune TME has also been hampered by the lack of suitable experimental models. Patient-derived organoids (19, 20) and patient-derived explants (21) have been developed to recreate the complexity of the human immune TME. Although these models are increasingly utilized, they are limited by the short viability of immune cells (typically 2–3 days) and distinct drug pharmacodynamics and pharmacokinetics profiles when compared with in vivo settings. Alternatively, humanized patient-derived xenograft (PDX) models enriched with an autologous human immune system derived from peripheral blood mononuclear cells (PBMC; ref. 22), hematopoietic stem-progenitor cells (CD34+; refs. 23, 24), or TILs (25, 26) have been proposed as an in vivo approach to recapitulate the tumor–immune cell interactions within the TME, offering longer experimental windows and more accurate pharmacokinetic/pharmacodynamic conditions.
In this study, we progressively dissected the CD8+ TIL compartment in patients with non–small cell lung cancer (NSCLC), using a multilayered approach with orthogonal and complementary analysis platforms, including single-cell mass cytometry (CyTOF), quantitative immunofluorescence (QIF), IMC, and RNA sequencing (RNA-seq), to interrogate the nature and features of dysfunctional human T cells. Furthermore, to analyze the effect of PD-1/B7-H1 blockade on different human TIL subsets, we developed a humanized PDX model (herein referred to as immune PDX) in which intact patient-derived TILs can be experimentally modulated and analyzed. Our studies reveal a distinct dysfunctional CD8+ TIL subset that expands within the TME in a PD-pathway–dependent manner. This CD8+ TIL subset displays markers of end-stage differentiation, produces lower IFNγ upon stimulation, and shows increased apoptosis but is also highly proliferative. Finally, we demonstrate that this CD8+ TIL subset accumulates in patients with more advanced NSCLC, and its higher abundance is associated with worse clinical response to anti-PD therapy.
Immune Profiling Reveals That CD8+ TILs Are Actively Expanding in NSCLC TME
Surgically resected tumoral (Tu) and nontumoral (non-Tu) paired lung tissues were obtained from 25 patients with resectable NSCLC. Specimens were divided into two mirroring halves and subjected to proteomic-based multiparametric CyTOF and QIF analysis (Supplementary Fig. S1A). We standardized a 39-plex CyTOF panel (Supplementary Table S1) including phenotypical and functional markers to interrogate lung tissue immune cell infiltration. Comparative analysis of T-cell abundance within the total non-Tu lung cell suspension by CyTOF or intact tissues by QIF showed a high correlation for both CD8+ (r = 0.74, P < 0.001; Supplementary Fig. S1B) and CD4+ (r = 0.70, P = 0.0002; Supplementary Fig. S1C) T-cell compartments, supporting the consistency of measurements and limited impact of our tissue digestion protocol. Rigorous validation of antibodies to recognize inducible markers was conducted using unstimulated versus stimulated human PBMCs (Supplementary Fig. S2). Next, we applied the aforementioned strategy to interrogate the immune composition of Tu and non-Tu lung tissues (Supplementary Fig. S1A). We defined the main immune cell populations based on the differential expression of specific-lineage markers (Fig. 1A; Supplementary Fig. S3A and S3B) and found that all lymphoid subsets were significantly increased in Tu tissues by CyTOF (Fig. 1A and B), and also by QIF (Fig. 1C) analysis. In contrast, non-lymphoid subsets were significantly less prominent in the TME, and the overall ratio of lymphoid cells to non-lymphoid cells was significantly higher in Tu tissues (Fig. 1D; Supplementary Fig. S4). Additionally, CD8+ TILs were significantly more proliferative (Ki-67+) than non-Tu CD8+ T cells (P < 0.0003; Fig. 1E). Other lymphoid populations, including conventional CD4+ T cells, T regulatory cells (Tregs), and B cells did not show a significant change in proliferation levels (Fig. 1E). These data indicate that CD8+ TILs expand within the TME of patients with NSCLC and are the most proliferative TIL fraction.
A Proliferative, Terminally Differentiated CD8+ TIL Subset Selectively Expands in TME
To further understand why proliferative CD8+ TILs are unable to control tumor growth in patients with NSCLC, we conducted a more comprehensive phenotypic and functional analysis of TILs by comparing them with lymphocytes from paired non-Tu lung tissues. Employing an unsupervised hierarchical clustering method named CITRUS (27), 2.1 × 104 CD3+ T cells from Tu and non-Tu lung tissues were stratified according to the expression profile of phenotypic markers. In the first iteration, CITRUS automatically stratified tissue-infiltrating CD3+ T cells in CD4+ and CD8+ branches (Fig. 2A; Supplementary Fig. S5A). Subsequently, CD8+ T cells were segmented into two main populations of effector (Eff; CCR7−CD45RA+CD45RO−) and effector memory (EM; CCR7−CD45RA−CD45RO+) cells, showing that these two CD8+ subgroups dominate the lung tissue infiltration (Fig. 2B and C; Supplementary Fig. S5B and S5C). The EM population was further subdivided into two additional subsets. One subset showed a conventional effector memory phenotype, referred to here as effector memory (Em) cells (Fig. 2B and C; Supplementary Fig. S5B and S5C). The other subset, in contrast, was characterized by high CD45RO, EOMES, FAS, CD27, CD28, KLRG1, PD-1, LAG3, and TIM3 expression. This profile is consistent with an activated, terminally differentiated, and proapoptotic phenotype, referred to here as burned-out effector (Ebo) cells (Fig. 2B–D; Supplementary Fig. S5B and S5C).
Comparison of functional markers across all three CD8+ TIL subsets showed that Ebo cells had the highest levels of proliferation (Ki-67). In contrast, granzyme B (GzmB), TBET, and KLRG1 expression was more prominent in classic Eff cells. Interestingly, EOMES was highly expressed by both Ebo and Eff cells, but the ratio of EOMES to TBET was significantly higher in Ebo (median = 4.8) than Eff (median = 2.8) cells, consistent with an exhausted phenotype previously described in murine models, EOMEShiTBETlo (Fig. 2D; ref. 8). Additionally, these two effector CD8+ T-cell populations were asymmetrically represented between TME and surrounding non-Tu lung tissues. While Ebo clusters were predominantly localized within the TME (false discovery rate [FDR] < 0.01; Fig. 2E), Eff clusters were more abundant in non-Tu lung tissues (FDR < 0.001; Fig. 2E). When we looked at PBMCs from patients with NSCLC, we observed that Ebo CD8+ T cells were almost absent, similar to non-Tu lung tissues (Supplementary Fig. S5D). Paired Tu and non-Tu analysis revealed that Eff CD8+ TILs express less GzmB and KLRG1 when compared with the paired non-Tu subsets, indicating a loss of function in addition to loss of cell numbers of Eff CD8+ T cells within the TME (Fig. 2F). Finally, although expression of PD-1 and TIM3 was significantly increased in the three CD8+ TIL subsets compared with non-Tu tissues, LAG3 expression was exclusively increased in Ebo, indicating that the LAG3 increase is a distinctive molecular feature of this subset (Fig. 2F). When we further stratified Ebo CD8 TIL cluster in high versus low Ki-67 expression, we observed that the high Ki-67 subpopulation displayed higher PD-1, LAG3, TIM3, and FAS expression levels (Supplementary Fig. S6A), and this subcluster appeared separately stratified in a t-distributed stochastic neighbor embedding (t-SNE) analysis (Supplementary Fig. S6B). These data reveal the existence of three major phenotypically distinct CD8+ TIL subsets, among which Ebo stands out, as it is highly proliferative (Ki-67+, CD28+), fully activated (PD-1+, LAG3+, TIM3+), terminally differentiated (EOMEShiTBETlo), and selectively expanded in the TME.
Ebo CD8+ TILs Are a Transcriptionally Distinct and Dysfunctional Subset
Given that Ebo was the only subset selectively expanded in the TME, we further explored the transcriptional and functional characteristics of these cells in order to understand what was unique about them. TILs were isolated from freshly resected NSCLC specimens; the three CD8+ TIL subsets were sorted by flow cytometry and processed for whole transcriptome RNA-seq analysis. Ebo showed greater transcriptional similarities with Em than with Eff CD8+ TILs, consistent with our proteomic CITRUS clustering results (Supplementary Fig. S7A). Nevertheless, when we compared Ebo directly with the Em CD8+ TIL population, we found a clearly distinct gene signature. Specifically, a total of 888 genes were significantly upregulated in Ebo over Em, and 881 genes were upregulated in Em over Ebo (FDR < 1%; adjusted P < 0.01), confirming that, although they are clustered together, these two cell populations are transcriptionally different (Fig. 3A). Consistent with our proteomic-based CyTOF analysis, Ki-67 gene expression was upregulated in Ebo compared with Em (Fig. 3B). Additionally, Ebo had higher expression of other transcripts associated with cell proliferation such as POLR3K and DNA2. In contrast, Em showed upregulated expression of genes related to tissue development and cellular differentiation, including ARMCX2 and FLIP1L (Fig. 3A and B). In line with these data, gene set enrichment analysis (GSEA) showed proliferative signatures enriched in the Ebo population (Fig. 3C and D, left panel), whereas the Em population showed an enrichment for stem cell–type signatures (Fig. 3C and D, right panel). Furthermore, we analyzed T-cell receptor (TCR) sequencing transcripts from the three subsets and observed that Eff CD8+ TILs displayed the highest clonality, whereas Ebo and Em showed a higher clonotype diversity (Fig. 3E and F; Supplementary Fig. S7B and S7C). No significant differential expression of the associated exhaustion markers ENTPD1, TOX, NR4A2, or CD38 was found across the three subsets (Supplementary Fig. S7D). Next, we interrogated the functional capacity of these three CD8+ TIL subsets to produce IFNγ. CD8+ TIL subsets were sorted by flow cytometry from nine primary NSCLC tissues, and intracellular IFNγ expression was analyzed upon ex vivo PMA–ionomycin stimulation. Intriguingly, although Ebo cells showed the highest baseline expression of proliferation and activation markers, they produced the least amount of IFNγ when compared with the Em and Eff populations (Fig. 3G). Because Ebo showed the highest expression of FAS, a proapoptotic marker, we hypothesized that the lack of IFNγ production may be explained because these cells are engaged in cell death programs. To avoid induction of cleaved caspases that may occur during tissue digestion to harvest TILs, we optimized the use of QIF to measure cleaved caspases within intact tumor tissues (Supplementary Fig. S8A and S8B). Among the three CD8+ TIL subsets, Ebo displayed the highest levels of cleaved caspase-3 expression as compared with the Em and Eff subsets (Fig. 3H; Supplementary Fig. S8C). Additionally, GSEA showed an apoptotic signature enriched in the Ebo subset. This signature is associated with cyclin-dependent kinase inhibitor 1A, which is considered instrumental in the execution of apoptosis following caspase activation (Fig. 3I). Together, these data show that the three CD8+ TIL subsets are transcriptionally distinct, consistent with the clustering observed in CyTOF analysis, and indicate that the Ebo subset is enriched in apoptotic and dysfunctional CD8+ T cells.
The B7-H1/PD-1 Pathway Mediates Ebo CD8+ TIL Expansion within TME
To explore the dynamic changes of CD8+ TIL subsets within the TME and their response to immunotherapy, we developed an immune-proficient PDX model. This alternative PDX model (immune-PDX), uses 4 to 5 mm3 fresh primary NSCLC surgical specimens engrafted subcutaneously into NOD/SCID IL2Rγ−/− (NSG) mice. Xenografts contain all human components of the patient's unique TME, including hematopoietic, stromal, and tumor cells, and support their modulation and study (Fig. 4A). We observed that CD8+, CD4+, B, natural killer T, and myeloid mononuclear phagocyte cellular composition remained constant for at least 2 weeks after engraftment (Fig. 4B and C). In contrast, natural killer cells and granulocytes significantly declined during the same period of time (Fig. 4B and C). We monitored the abundance of the three CD8+ TIL subsets within the TME at baseline and 2 weeks after engraftment by CyTOF and found that Ebo significantly increased (27.6 ± 5.6%; P = 0.0012), but Em (–22.8 ± 4.6%; P = 0.0010) and Eff (–5.62 ± 2.75%; P = 0.0755) CD8+ TILs progressively decreased over time (Fig. 4D and E). Interestingly, tumors expressing high B7-H1 levels maintained their expression for 2 weeks after engraftment (Supplementary Fig. S9), indicating that this pathway remains intact in this model and is amenable to modulation. To test the effect of anti-PD therapy on CD8+ TIL populations, we treated immune-PDX engrafted with B7-H1+ tumors with avelumab (anti-human B7-H1 mAb) or isotype control. CD8+ TIL subsets were evaluated by CyTOF after two doses of treatment. As shown in Fig. 4F and G, avelumab treatment reduced the expansion of Ebo (–13.96 ± 4.44%; P = 0.012) and prevented the reduction of Eff (+3.978 ± 1.58%; P = 0.0311), resulting in a favorable Eff/Ebo ratio in avelumab-treated mice (Fig. 4F and G). Additionally, avelumab decreased the expression of the proapoptotic marker (BIM) in the Eff cells and increased the levels of the activation markers CD27 and CD69. These results suggest that B7-H1/PD-1 blockade may play a protective role for this population during activation (Fig. 4H). In both Eff and Em, TBET was significantly increased after B7-H1 blockade (Fig. 4H), in line with the observed ex vivo increased IFNγ production upon restimulation (Fig. 3E). These data demonstrate that the Ebo CD8+ TIL subset is expanded within the TME over time and that this expansion can be reduced by B7-H1 blockade.
Ebo Abundance Is Associated with Tumor Progression and Resistance to Anti-PD Therapy
Because we observed that Ebo expands within human TME over time, we speculated that expansion of this dysfunctional CD8+ T-cell population may correlate with advancing disease and contribute to immune escape and/or resistance to immunotherapy. To test this hypothesis, we standardized a 37-plex IMC panel (Supplementary Table S2) to identify the Ebo CD8+ TIL subset in 35 patients with early and advanced NSCLC (Fig. 5A). Patients with NSCLC with clinical stage IV disease showed a significantly higher abundance of Ebo CD8+ TILs than patients with clinical stage I disease (23.03 ± 7.42%; P = 0.006), supporting the notion that this population expands as tumors continue to progress and evolve over time (Fig. 5B). Age, histologysubtypes, and CD8 TIL density were studied across the 35 NSCLC cases and were homogeneously distributed (Supplementary Fig. S10A). Additionally, we did not observe any association between these variables and Ebo abundance (Supplementary Fig. S10B and S10C). We did observe that, in advanced NSCLC cases, the presence of tertiary lymphoid structures was associated with an ∼2.5-fold lower Ebo CD8+ TIL proportion, but this difference did not reach statistical significance (Supplementary Fig. S10D and S10E). Finally, consistent with previous observations by QIF, we identified among advanced NSCLC cases a CD8 TIL cluster coexpressing cleaved caspase-3, CD45RO, EOMES, Ki-67, PD-1, LAG3, and TIM3 (Supplementary Fig. S11). Importantly, patients with non-durable clinical benefit (NDB) to anti-PD therapy showed a significantly higher proportion of the Ebo CD8+ TIL subset than patients with durable clinical benefit (DCB) in baseline tumor samples (P = 0.004; Fig. 5C). Alternative Ebo CD8+ TIL gating strategies including Ki-67 protein expression were used to confirm this result (Supplementary Fig. S12A–S12C). Consistent with this finding, patients with NDB also showed CD8+ TILs with higher PD-1, LAG3, and Ki-67 expression, markers associated with the Ebo CD8+ TIL phenotype (Fig. 5D). Finally, Ebo CD8+ TIL subset abundance was associated with a worse overall survival (hazard ratio = 6.08; 95% confidence interval, 2.57–5.25; P < 0.05) in patients with NSCLC treated with anti-PD therapy (Fig. 5E). A multivariate analysis including PD-L1 expression as a covariate identified Ebo CD8+ TIL abundance as an independent predictor of outcome (Supplementary Table S3). Taken together, our findings support the suggestion that Ebo is expanded in the TME of patients with advanced NSCLC, and its expansion is independently associated with reduced efficacy of anti-PD therapy.
The basis of T-cell dysfunction in cancer and the best way to overcome it remain incompletely understood. One of the main hurdles to achieving greater understanding has been the use of low-resolution phenotypical technologies and the lack of suitable in vivo experimental models to investigate the human immune TME. Here, we used multiparametric single-cell mass cytometry and tissue imaging technology tools combined with a lung cancer humanized immune-PDX model to explore CD8+ TIL diversity and its response to immunotherapy. We identified three phenotypically, transcriptionally, and functionally distinct CD8+ TIL subsets: Eff, Em, and Ebo. The Ebo subset is characterized by a burned-out dysfunctional phenotype and specifically expands within the TME in a PD-pathway–dependent manner. Furthermore, Ebo progressively accumulates in patients with advanced NSCLC and is associated with primary resistance to anti-PD therapies. Our findings suggest that this CD8+ TIL population may contribute to a less responsive immune TME and provide insights into the development of future T cell–based cancer immunotherapies.
To phenotypically characterize CD8+ TILs, we employed single-cell mass cytometry technology, which allowed us to multiplex 39 markers per cell. Different methods of analysis have been designed over the past few years to maximize this multidimensional single-cell data (28). Here, we used CITRUS, a method specifically designed to identify and compare hierarchical clusters between groups of samples in an unsupervised manner. Using this method, we compared paired Tu and non-Tu lung tissue for immune cell composition. This approach resulted in fundamental findings: (i) lymphoid cells, but not myeloid cells, are preferentially expanded in lung tumor tissues; (ii) CD8+ TILs are highly proliferative in contrast to non-Tu tissue-infiltrating CD8+ T cells; and (iii) three major CD8+ TIL subsets were identified, but only one of them, Ebo, was expanded in the TME (Fig. 2E). In contrast, the other two CD8+ TIL subsets were either diminished (Eff) or unchanged (Em) in the TME compared with adjacent non-Tu lung tissue. More importantly, we found that, whereas Ebo represents 10% to 15% of the total CD8+ TILs in early-stage NSCLC, its abundance rises to 30% to 40% in advanced NSCLC, indicating a higher relevance of this compartment in late-stage disease (Fig. 5B).
To demonstrate the dynamic evolution of Ebo within the TME and its response to immunotherapy, we developed a previously undescribed immune-proficient PDX model, referred to as the immune-PDX model. In contrast to conventional immune-deficient PDX models, immune-PDX retains an intact patient-derived tumor immune microenvironment (Fig. 4A and B). Fresh patient-derived lung tumor xenografts are engrafted into severely immunodeficient mice and contain all the original components of the growing tumor in the patient, including hematopoietic, stromal, and tumor cells. Although fibroblasts, neutrophils, and natural killer cells are progressively replaced by murine stroma cells, all human lymphoid populations remain stable during the first 2 weeks after engraftment, supporting their modulation and study. When TILs are studied in patient samples, a challenging aspect is that new T cells are continuously arriving at the TME site, making it difficult to know which CD8+ subsets have just arrived and which have been differentiated within the TME. Because we recreated the patient TME in a T cell–deficient host, one of the main advantages of this model is that the evolution and reactions of TIL subsets can be analyzed without the interference of “newcomers.” In this context, we demonstrated that the Ebo CD8+ TILs were progressively expanded within TME over time. More interestingly, we found that Ebo expansion was minimized when mice were treated with avelumab (clinical grade anti–B7-H1 mAb), indicating that the PD pathway is required for the expansion of this population. Although the contribution of the PD pathway to a CD8 subset expansion seems to be contradictory with prior reports suggesting that PD-1 transmits a signal to suppress T-cell proliferation (29, 30), it is important to consider that these studies employed polyclonal and heterogeneous T cells. It is still possible that the PD pathway provides a proliferative signal for a subset of T cells, which is consistent with our earlier work showing that B7-H1 could be costimulatory for human T cells (31). Supporting this hypothesis, it has been recently reported that PD-1 silencing inhibits the proliferative activity of chimeric antigen receptor modified T cells (32). It remains to be clarified, however, whether the observed expansion of Ebo associated with the PD pathway is due to a selective proliferative stimulus or if, instead, the B7-H1/PD-1 blockade is preventing the conversion of other populations into Ebo. The fact that clonality was low and clonotype diversity high in the Ebo compartment further supports the idea that the expansion of this compartment is more dependent on the conversion of other populations into Ebo.
Dysfunctional CD8+ T-cell subsets have been previously described in the TME (33). The most frequently used term to designate these cells has been “exhausted” T cells, imported from chronic viral infections lymphocytic choriomeningitis virus. The Ebo CD8+ TILs in our study are also dysfunctional and share several features with exhausted T cells, including coexpression of coinhibitory receptors (PD-1, LAG3, and TIM3), loss of IFNγ production upon restimulation, and a terminally differentiated phenotype (EOMEShiTBETlo). However, in contrast with the classic exhausted phenotype, Ebo is highly proliferative. This intriguing result is supported by two recent studies showing a highly proliferative dysfunctional CD8+ TIL subset in the TME of patients with melanoma and NSCLC (9, 10). In our case, this dysfunctional population is characterized not only by a highly proliferative signature but also by the highest expression of activation markers such as CD45RO, CD27, PD-1, LAG3, TIM3, CD28, and FAS, suggesting that these cells may be pushed beyond the threshold of optimal activation. Supporting this interpretation, we previously demonstrated that patients with NSCLC with KI-67hi and/or GzmBhi CD8+ TILs are associated with worse response to cancer immunotherapy (34). Other groups have suggested that an overactivation of CD8+ TILs may be counterproductive, due to an excess of IFNγ (35) or TNFα (36) that may lead to T-cell death programs, thereby compromising the antitumor response. In this regard, we have observed that the Ebo population is characterized by the highest FAS and cleaved caspase-3 levels compared with the other CD8+ TIL subsets, reinforcing the interpretation of apoptotic hyperactivated CD8+ TILs. Given these distinctions from previously described exhausted TILs, we have proposed the term “burned-out” T cells to more accurately reflect the overactive and apoptotic phenotype of these cells. Importantly, blockade of the PD pathway reduced BIM and cleaved caspase-3 in activated Eff CD8+ TILs (Fig. 4H). Collectively, our results suggest that anti-PD therapy may be preventing activated T cells from entering into an apoptotic death program rather than restoring the function of exhausted/dysfunctional T cells. Although these findings require further confirmation, they are consistent with previous studies reporting that anti-PD therapies may be promoting the expansion and preventing the deterioration of a stem-like CD8+ TIL subset, identified as PD-1+Tcf1+, rather than restoring exhaustion (37–40).
T cell–inflamed tumors with abundant Ebo CD8+ TILs may be refractory to cancer immunotherapy due to an excess of irrecoverable apoptotic TILs, as previously suggested (41, 42), potentially competing for space and resources with other CD8+ TIL populations. Along these lines, a recently published study showed an association between higher T-cell infiltration and higher terminally exhausted CD8+ T cells, indicating that the greater the T cell accumulation in tumors, the greater the likelihood of becoming dysfunctional (40). One interpretation of these results is that a lack of T-cell homeostasis and an excess of activated ineffective TILs may be responsible for impaired T-cell responses in some patients. If this is the case, T-cell enhancer strategies (e.g., bispecific T-cell engagers) may be detrimental, and strategies designed to protect activated T cells from apoptosis may be needed instead. An alternative interpretation of these results may be that the Ebo compartment must be depleted to provide space and resources to the cells with greater efficacy such as Eff CD8+ TIL subsets. In this regard, a potential benefit of concomitant chemotherapy or radiotherapy may be the clearance of the Ebo compartment.
Taken together, our data illustrate a previously unrecognized category of anti-PD–resistant T-cell inflamed tumors. Contrary to expectations, the underlying mechanism of resistance may be an excess of dysfunctional activated TILs. This work may contribute to a deeper understanding of T cell–inflamed tumor diversity, and Ebo may serve as a useful tissue biomarker with implications for the development of future T cell–based cancer immunotherapies.
Patients and Samples
Fresh tumor samples were prospectively collected from individuals with NSCLC undergoing primary surgical treatment between August 2015 and February 2018 at the Yale Cancer Center, New Haven, CT. After surgery, tumor and tumor-free sections (identified and supervised by a pathologist) were collected from each patient and processed in the first 2 hours after being surgically removed. Tumor-free fragments were collected at least 5 cm away from the macroscopic tumor border. Tumor and tumor-free fragments were reviewed by a licensed pathologist (K.A. Schalper) to confirm the diagnosis and quality of the samples. One tumor sample was identified as sarcoidosis, and one tumor-free sample was found to include areas of active pneumonia. Both of these cases were processed but not included in the analysis. For each specimen, a fragment was fixed in 10% neutral-buffered formalin and paraffin-embedded for histopathologic analysis. The remainder of the tissue was directly processed for single-cell CyTOF analysis (Supplementary Fig. S1). When tumor tissues were bigger than 10 mm3, an additional fragment of the tumor tissue was engrafted in immunodeficient NSG mice to establish a short-term PDX. The clinical and pathologic characteristics of these patients are described in Supplementary Table S4. Additionally, we included retrospectively collected formalin-fixed, paraffin-embedded tumor samples from patients with stage IV NSCLCs treated with single-therapy anti-PD agents at the Yale Cancer Center between 2011 and 2017. These cases were represented in a tissue microarray (TMA) format, each containing two independent 0.6-mm tumor cores. The TMAs were constructed by selecting areas containing viable tumor cells and stromal elements on hematoxylin and eosin–stained preparations (as assessed by a licensed pathologist) and without enriching for specific tumor regions, tissue structures, or immune-related features. The clinical and pathologic characteristics of these patients are described in Supplementary Table S4 and treatment characteristics in Supplementary Table S5. All tissues were used with each patient's written informed consent or, in some cases, a waiver of consent, after approval from the Yale Human Investigation Committee (protocols 1412015109, 9505008219, and 1608018220) and conducted in accordance with the tenets of the Declaration of Helsinki.
All metal-conjugated antibodies were either purchased pre-conjugated from Fluidigm or purchased purified and conjugated in-house using mass cytometry antibody conjugation kits from Fluidigm according to the manufacturer's instructions. For details, see Supplementary Tables S1 and S2. The following antibodies were developed, produced, and quality controlled at Pfizer facilities: avelumab (a fully human IgG1 anti-human PD-L1) and irrelevant human IgG1 as an isotype-matched control.
Female NOD/SCID IL2Rγ−/− (NSG) mice 6 to 8 weeks old were purchased from The Jackson Laboratory. All mouse protocols were in accordance with NIH guidelines and approved by the Animal Care and Use Committee of Yale University School of Medicine (reference 2013-11387 approval).
CyTOF Sample Preparation
A total of 1 × 106 to 3 × 106 suspension cells from each Tu or non-Tu lung tissue were incubated with antibodies against CD16/32 at 50 mg/mL for 10 minutes at room temperature to block Fc receptors. Subsequently, metal-conjugated mAb cocktails against cell surface molecules were added, and further incubated for 20 minutes on ice (see details in Supplementary Table S1). Cells were stained for viability with 5 mmol/L cisplatin in FBS (Fluidigm) for 1 minutes at 4°C. After the treatment with the Fixation/Permeabilization Buffer (Thermo Fisher), cells were further incubated with the metal-conjugated mAb cocktails against intracellular proteins (see details in Supplementary Table S1). Then, cells were washed and stained with 1 mL of 1:4000 191/193Ir DNA intercalator (Fluidigm) diluted in phosphate-buffered saline (PBS) containing 1.6% paraformaldehyde (Electron Microscopy Sciences) and stored at 4°C until acquisition.
Single-Cell Mass Cytometry Sample Acquisition
Samples were acquired in a Helios machine (Fluidigm). At the moment of acquisition, cells were washed once with PBS with 0.5% bovine serum albumin (BSA) and 0.02% NaN3 and once with double-distilled water (ddH2O); they were then resuspended in ddH2O containing bead standards (Fluidigm) to approximately 106 cells/mL. Subsequently, samples were acquired at an event rate of <400 events per second.
Antibody Validation and Titration for Single-Cell Mass Cytometry
Isolated human PBMCs were stimulated with a concentration of 0.5 μg/mL plate-bound anti-CD3 (OKT3) and 2 μg/mL soluble anti-CD28 (28.2; eBioscience) for 72 hours or cultured in medium—RPMI 1640 with 10% FBS (PAN-Biotech), 1 mmol/L sodium pyruvate (Sigma), and 100 ng/mL penicillin/streptomycin—for the same period of time. Unstimulated PBMCs were used as a negative control for antibody titration and staining protocol optimization. The optimal antibody concentration was assayed by serial dilutions and determined by the obtained median mass intensity signal-to-noise ratio from internal positive and negative controls (see details in Supplementary Fig. S2). Additionally, in each CyTOF experiment unstimulated and stimulated PBMC samples were included as internal reference.
Using previously validated and standardized multiplexed QIF panels and serial TMA sections (43), we measured the levels of cytokeratin (clone AE1/AE3; eBioscience), PD-L1 (B7-H1, clone 405.9A11; Cell Signaling Technology), CD4 (clone SP35; Spring Bioscience), CD8 (clone C8/144B; Dako), CD20 (clone L26; Dako), CD68 (clone KP1; Dako), CD45RO (clone UCHL1; BioLegend), and CD45RA (clone HI100; BioLegend). Cleaved caspase-3 (clone D3E9; Cell Signaling Technology) was validated and standardized using Jurkat cells treated with dimethyl sulfoxide (untreated) or 1 μmol/L of camptothecin (treated) for 18 hours to induce apoptosis (see details in Supplementary Fig. S8A and S8B). Freshly cut TMA serial sections were deparaffinized, and antigen retrieval was performed using an EDTA buffer (Sigma-Aldrich), pH 8; they were then heated for 20 minutes at 97°C in a Lab Vision PT Module (Thermo Fisher Scientific). Endogenous peroxidase activity was inactivated by incubating the TMA section in a solution containing 0.3% hydrogen peroxide in methanol for 20 minutes, followed by incubation in a blocking solution containing 0.3% BSA in 0.05% Tween solution for 30 minutes. An antibody cocktail was used for the primary target detection, followed by a secondary detection using the specific horseradish peroxidase–conjugated IgG anti-isotype. Tyramide signal amplification with tyramide-based fluorescent reagents (PerkinElmer) was utilized for signal amplification and detection. Sections were washed twice with a 100 mmol/L benzoic hydrazide solution in between antibody incubation periods to eliminate residual horseradish peroxidase activity. Nuclei staining was performed with 4′,6-diamidino-2-phenylindole (DAPI). Fluorescence signal was measured quantitatively using the AQUA method for QIF. Samples with tissue or staining artifacts by visual examination were excluded.
IMC Sample Preparation
For IMC staining, antibodies were optimized using QIF as a gold standard and a control array including negative and positive controls, as well as cases with a range of expression for each marker. Fluorescence signal quantification was performed using the AQUA method of QIF. Each slide was visually examined to exclude samples with tissue/staining artifacts and those with less than 5% tumor content. Cases were considered to display detectable levels of each target when the QIF score was above the signal-to-noise threshold determined by measuring negative control preparations and by visual examination of the sample. For IMC staining, a comparable and compatible staining protocol was designed based on the QIF gold standard. Briefly, fresh TMA serial sections were deparaffinized, and antigen retrieval was performed using an EDTA buffer (pH 9) and heating for 20 minutes at 97°C in a Lab Vision PT Module. The EDTA solution was rinsed twice using Tris-buffered saline (TBS) solution, followed by incubation in a blocking solution containing 0.3% BSA in 0.05% Tween solution for 30 minutes. Tissue was covered with a cocktail of 26 metal-conjugated antibodies prepared in blocking solution. After 1 hour of primary incubation at room temperature, tissues were rinsed twice using a solution containing 0.025% Tween 20 in TBS solution to remove nonspecific antibody staining. DNA and membranes were costained using a solution containing 1 μmol/L pooled 191Ir and 193Ir, and 1 μg/mL 115In LipoR in TBS for 1 hour at room temperature. The costaining solution was rinsed using Tween 20 and the TBS solution followed by two washes for 1 minute each using ddH2O. The slide was finally air-dried for 60 minutes and stored until analysis.
IMC Sample Acquisition
For IMC analysis and instrument settings, the Hyperion machine from the Yale IMC core facility was used for all the studies described here. The mass cytometer and scanner were tuned every day at 20 Hz and 0 dB of laser intensity using a commercial tuning slide provided by Fluidigm. The tuning reference values used as a quality control for a working day were 2.37% ± 0.34% and 0.42% ± 0.11% for transient cross-talk 1 and 2, respectively, and for Lu(175) mean duals above 900. All tissue samples and cells were ablated at 5 dB and 200 Hz; an area of 750 × 750 μm was always considered for TMA spots and a smaller area of 100 × 100 μm for titrations. TMA slides were scanned using the batch tool. A panoramic view of the whole slide was uploaded, and a macro region of interest was created covering the tissue spot grid. On the created working area, each spot was selected based on the map layout of each TMA. A protocol was assigned and carried out until 20 spots were completed for a 24-hour continuous scan. No more than 25 cases were programmed to be scanned per tuning event.
Antibody Validation and Titration for IMC
For IMC validation and optimization, specificity, and sensitivity, the optimal titer was calculated separately for metal-conjugated and unconjugated versions of each antibody. The optimal antibody concentration was assayed by serial dilutions and determined by the obtained QIF scores and the signal-to-noise ratio from internal positive and negative control tissues (percentile p90 to percentile p10). Cytokeratin and vimentin were assayed in combination to confirm selective exclusion staining patterns for further tissue segmentation. Antigen retrieval conditions were optimized, and QIF was performed using 1 mmol/L EDTA (pH 8) and 10 mmol/L citrate (pH 6) independently. The detection limit was optimized by comparing the dynamic range for each condition.
Immune-PDX Tumor Model
Surgical specimens from primary lung carcinomas were implanted subcutaneously into the flank of 8 to 10 female NSG mice at 6 to 8 weeks of age per experiment. Mice were treated intraperitoneally with anti-hPD-L1 mAbs (avelumab) or isotype control at days 5 and 10 post-implantation. At day 7 or 14, mice were sacrificed, and tumors were collected for analysis.
Human TIL and PBMC Isolation
Human TILs were derived from original patient tumors, as well as tumor tissue from each immune-PDX model after treatment. TILs were obtained by processing tumor tissue with a gentleMACS Dissociator (Miltenyi Biotec) in the presence of RPMI 1640 with 0.5% BSA and 5 mmol/L EDTA. Cell suspensions were filtered using a 70-μm cell strainer (BD Falcon), centrifuged at 600g for 7 minutes at 4°C, and resuspended in PBS with 0.5% BSA and 0.02% NaN3, and were then ready for analysis. Human PBMCs were purified from buffy coats of healthy donors using Ficoll-Hypaque (Sigma-Aldrich) gradient centrifugation, according to the manufacturer's instructions.
For isolation of Ebo, Em, and Eff CD8+ TIL subsets, 1 × 106 to 2 × 106 tumor-infiltrating leukocytes were resuspended in 200 μL staining buffer containing fluorophore-conjugated mAb cocktail against human cell-surface molecules: CD45 BV510 (clone HI30), CD8 PerCP-Cy5.5 (clone RPA-T8), CCR7 BV421 (clone G043H7), CD45RO PE (clone UCHL1), and CD45RA PE-Cy7 (clone HI100). After a 20-minute incubation at 4°C, cells were washed twice, and cell sorting was performed using a BD FACS Aria III (BD Biosciences). We trained a gating strategy to sort the subsets most similar to those identified by CITRUS. We first trained this strategy using our current CyTOF data doing manual gating and comparing the expression profile of the three subsets identified by CITRUS and the three subsets defined by manual gating (Supplementary Fig. S13A and S13B). After we confirmed that our manual gating strategy was able to identify subsets similar to those identified by CITRUS, we designed a similar strategy for the FACS (Supplementary Fig. S13C). Furthermore, after we sorted the cells, we tested whether or not the expression profile of the sorted subsets was comparable to that of the clusters defined by CITRUS (Supplementary Fig. S13D). Approximately 1 × 104, 1.5 × 104 to 2 × 104, and 0.5 × 104 to 1 × 104 of Ebo, Em, and Eff cells, respectively, were sorted as an average across experiments. In three experiments cells were directly sorted in lysis buffer RLT (RNeasy Micro Kit, 74004; QIAGEN) which already contained beta-mercaptoethanol and dithiothreitol, and total RNA was extracted from the lysed cells using the RNeasy Micro Kit protocol. Total RNA samples were submitted to the Yale Center for Genome Analysis (YCGA) where they were analyzed for quality control and RNA-seq. In the remainder of experiments, cells were cultured in vitro for T-cell stimulation experiments.
T-cell Stimulation In Vitro
From nine lung tumor samples, Ebo, Em, and Eff CD8+ TIL subsets were stimulated with 50 ng/mL PMA and 1 μg/mL ionomycin (BD Biosciences) for 6 hours. Two hours before the end of the PMA–ionomycin stimulation, 1× monensin and 1× brefeldin A (eBioscience) were added for intracellular IFNγ staining. For analysis of IFNγ intracellular levels, cells were fixed and permeabilized using the Fix/Perm solution (eBioscience) for 30 minutes at room temperature. After they were washed twice, cells were resuspended in 1× Permeabilization Buffer (eBioscience) containing fluorescein isothiocyanate–conjugated anti-IFNγ antibodies (clone 25723.11) for 15 minutes at room temperature. The acquisition of samples was carried out on an Attune NxT Flow Cytometer (Thermo Fisher Scientific).
CyTOF Data Analysis
All mass cytometry files were normalized together using the mass cytometry data normalization algorithm (44). For analysis, FCS files were manually pregated on 193Ir DNA+CD45+ events, excluding cisplatin-positive dead cells, doublets, and DNA-negative debris, by Cytobank. The gated CD45+ population was then clustered based on all labeled phenotypic markers using visualization (viSNE; ref. 45) or hierarchical clustering (CITRUS; ref. 27) methods. CITRUS is comprised of several steps: (i) Cells from samples are combined and clustered in a semi-unsupervised manner to automatically identify clusters of related cells. (ii) Descriptive statistics characterizing various properties of each cluster (cluster features) are extracted on a per-sample basis. (iii) Extracted cluster features are used in conjunction with a user-specified endpoint of interest to train a supervised model. (iv) Internal cross-validation is used to evaluate model fit and select the appropriate regularization threshold for a final model. Finally, (vi) model features are plotted as a function of endpoint of interest, and cluster phenotypes are determined by density plots of markers used for clustering. An equal number of events (or a subset of events) from all samples are randomly selected and then combined and clustered using agglomerative hierarchical clustering, producing groups of phenotypically similar cells. The dissimilarity between any two cells was specified by the Euclidean distance between clustering markers, and Ward's linkage was used as the agglomeration method. Rather than cutting the dendrogram at a fixed height to identify clusters, all clusters in the hierarchy of merged clusters larger than a user-specified size were retained for subsequent analysis.
IMC Data Analysis
A supervised phenotyping and segmentation analysis was developed using the IMC viewer provided by Fluidigm. For imaging studies and staining pattern comparisons, each metal channel was visualized and the image quality was optimized using gain and thresholds range tools. Integrated dual histograms were used as a reference for signal normalization to avoid saturated pixels and for comparing with a pathologist's validated immunofluorescence pattern. After fixing all parameters for the visualization of each stained marker, a supervised identification of T lymphocytes, macrophages, and tumor cells was performed based on morphology and the specific cluster of differentiation for lymphocytes CD4 (156Gd), CD8 (162-Dy), CD20 (161-Dy), CD68 (159Tb), and pan-cytokeratin (148Nd) for tumor cells. All scanned regions of interest included a total of 20 cells per field of view from a TMA batch and were analyzed in the same manner. The selected positive pixels for a marker were indicated as single cells. The integrated counts were exported as a TXT file, and the composed images were saved for further studies. The integral counts for each marker were normalized by the mean level of 132Xe and 133Xe detected in each cell compartment. This allowed us to account for the different areas covered by cells of different shapes and sizes, as the Xe ion is uniformly detected in tissue due to impurities in the argon plasma source. The Ebo, Em, and Eff CD8 TIL compartments identified by IMC were based in the combination of CD8, CD45RO, CD45RA, and EOMES markers, using the following approach: Ebo (CD8+, CD45RO+, CD45RA−, and EOMES+), Em (CD8+, CD45RO+, CD45RA−, and EOMES−), and Eff (CD8+, CD45RO−, and CD45RA+). The expression profile of each compartment was determined to evaluate consistency with the clusters identified by CITRUS (Supplementary Fig. S14).
RNA-seq Data Analysis
RNA samples were submitted to the YCGA, and the libraries were constructed by YCGA protocols (https://medicine.yale.edu/keck/ycga/Images/9_tcm240-21599.pdf). The libraries underwent 101-bp paired-end sequencing using an Illumina HiSeq 2500 according to Illumina protocols. Low-quality reads were trimmed and adaptor contamination was removed using Trim Galore (v0.5.0). Trimmed reads were mapped to the human reference genome (hg38) using HISAT2 (v2.1.0) (46). Gene expression levels were quantified using StringTie (v1.3.3b) (47) with gene models (v27) from the GENCODE project. Differentially expressed genes were identified using DESeq2 (v1.22.1) (48).
TCR Sequencing Data Analysis
T-cell receptor β-chain clonality was assessed using the MiXCR tool using standard parameters of the RNA-seq workflow. TCR clonality was analyzed using the R package immunarch (http://doi.org/10.5281/zenodo.3367200).
Student t test, two-way analysis of variance, and log-rank tests on GraphPad Prism 8.0 for macOS were used for statistical analysis. CD8 and Ebo abundance cutoff points were determined using the median values across all samples. Multivariate analysis was performed using the Cox proportional hazards regression model on SPSS Statistics for Windows. Statistical significance was considered to be P < 0.05. P values were reported as follows: NS, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. The error bars in figures represent standard error of the mean (SEM).
The data have been archived at https://www.ncbi.nlm.nih.gov/geo and can be accessed under the study accession number GSE167235.
F. Villaroel-Espindola has received personal fees from Instituto Oncológico Fundación Arturo López Pérez outside the submitted work. T.K. Kim has received research funds from NextCure, American Cancer Society, American Society of Clinical Oncology, and the Edward P. Evans Foundation. I. Melero has received grants and personal fees from BMS, Roche, AstraZeneca, Bioncotech, and Alligator and personal fees from Pharmamar, Genmab, F-Star, Numab, Gossamer, and EMD outside the submitted work. R.S. Herbst receives consulting fees from the following: AbbVie Pharmaceuticals; ARMO Biosciences; AstraZeneca; Biodesix; Bolt Biotherapeutics; Bristol Myers Squibb; Candel Therapeutics, Inc.; Cybrexa Therapeutics; eFFECTOR Therapeutics, Inc.; Eli Lilly and Company; EMD Serono; Genentech/Roche; Genmab; Halozyme Therapeutics; Heat Biologics; I-Mab Biopharma; Immunocore; Infinity Pharmaceuticals; Loxo Oncology; Merck and Company; Mirati Therapeutics; Nektar; Neon Therapeutics; NextCure; Novartis; Ocean Biomedical, Inc.; Oncternal Therapeutics; Pfizer; Ribbon Therapeutics; Sanofi; Seattle Genetics; Shire PLC; Spectrum Pharmaceuticals; STCube Pharmaceuticals, Inc; Symphogen; Takeda; Tesaro; Tocagen; WindMIL Therapeutics; Xencor, Inc. R.S. Herbst received research support from AstraZeneca; Eli Lilly and Company; Genentech/Roche; Merck and Company. R.S. Herbst is a member of the board of directors (nonexecutive/independent) for Junshi Pharmaceuticals; Immunocore Holdings Limited. K.A. Schalper received grants from the National Cancer Institute–National Institutes of Health, Department of Defense, AACR/Stand Up To Cancer, AstraZeneca, and The Mark Foundation during the conduct of the study; has received personal fees from Clinica Alemana Santiago, Novartis, Moderna, Shattuck Labs, Pierre-Fabre, AstraZeneca, Ono Pharmaceuticals, AbbVie, Dynamo Therapeutics, EMD Serono, Torque Therapeutics, Agenus, Takeda, Genmab, Merck, Bristol Myers Squibb, PeerView, and Fludigm outside the submitted work; and has received grants from Navigate Biopharama, Tesaro/GSK, Moderna, Takeda Pharmaceuticals, Surface Oncology, Pierre-Fabre Research Institute, Merck, Bristol Myers Squibb, AstraZeneca, Ribon Therapeutics, and Eli Lilly outside the submitted work. L. Chen received grants from the National Cancer Institute–National Institutes of Health during the conduct of the study; has received grants from NextCure, DynamiCure, and Tayu outside the submitted work; and has received personal fees from Junshi, Zai Lab, Vcanbio, and GenomiCare outside the submitted work. No potential conflicts of interest were disclosed by the other authors.
M.F. Sanmamed: Conceptualization, data curation, formal analysis, supervision, investigation, visualization, methodology, writing–original draft, writing–review and editing. X. Nie: Data curation, investigation, writing–review and editing. S.S. Desai: Resources, data curation, formal analysis, writing–review and editing. F. Villaroel-Espindola: Resources, data curation, formal analysis, investigation, methodology, writing–review and editing. T. Badri: Resources, data curation, formal analysis, investigation, methodology, writing–review and editing. D. Zhao: Resources, data curation, formal analysis, writing–review and editing. A.W. Kim: Resources, investigation, writing–review and editing. L. Ji: Resources, investigation, writing–review and editing. T. Zhang: Writing–review and editing. E. Quinlan: Investigation, writing–review and editing. X. Cheng: Investigation, writing–review and editing. X. Han: Writing–review and editing. M.D. Vesely: Resources, methodology, writing–review and editing. A.F. Nassar: Resources, data curation, formal analysis, methodology, writing–review and editing. J. Sun: Writing–review and editing. Y. Zhang: Writing–review and editing. T.K. Kim: Writing–review and editing. J. Wang: Data curation, writing–review and editing. I. Melero: Writing–review and editing. R.S. Herbst: Funding acquisition, project administration, writing–review and editing. K.A. Schalper: Resources, data curation, formal analysis, investigation, visualization, writing–review and editing. L. Chen: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing.
We thank other members of L. Chen's Laboratory at Yale University for helpful discussions and technical assistance. We thank Maria Olguin for her support with figure design and Sarin Siriamnothep for her patience and support during preparation of this manuscript. This study was supported by the Mark Foundation for Cancer Research (19-029-MIA; R.S. Herbst and K.A. Schalper), Navigate Biopharma (K.A. Schalper), AstraZeneca (K.A. Schalper), Yale SPORE in Lung Cancer (P50CA196530; R.S. Herbst), Department of Defense–Lung Cancer Research Program (W81XWH-16-1-0160; K.A. Schalper), the National Institutes of Health (R03CA219603 and R37CA245154; K.A. Schalper), Stand Up To Cancer–American Cancer Society Lung Cancer Dream Team Translational Research Grant (SU2C-AACR-DT1715; R.S. Herbst and K.A. Schalper), and a Stand Up To Cancer–American Cancer Colorectal Cancer Dream Team Translational Research Grant (SU2C-AACR-DT22-17; K.A. Schalper). Stand Up To Cancer is a division of the Entertainment Industry Foundation. The indicated Stand Up To Cancer grants are administered by the American Association for Cancer Research, the scientific partner of Stand Up To Cancer. M.F. Sanmamed is supported by a Miguel Servet contract (MS17/00196) and a grant from Instituto de Salud Carlos III, Fondo de Investigacion Sanitaria (PI19/00668). M.D. Vesely is supported by a Physician-Scientist Career Development Award from the Dermatology Foundation, a Dermatology Fellow Award from the Melanoma Research Alliance, and a grant from the National Center for Advancing Translational Sciences (KL2 TR001862).