Tumor-Associated Antigen Burden Correlates with Immune Checkpoint Blockade Benefit in Tumors with Low Levels of T-cell Exhaustion

Genomic datasets and clinical trial datasets

TCGA Pan-cancer gene expression and non-synonymous somatic mutation data (Pan-Cancer GDC version) were retrieved from UCSC Xena browser (https://xenabrowser.net). The Genotype-Tissue Expression (GTEx) bulk normal somatic tissue expression data were retrieved from GTEx portal (https://gtexportal.org/). The clinical trial datasets were retrieved from Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo), the European Genome-phenome Archive (EGA, https://ega-archive.org), or database of Genotypes and Phenotypes (dbGaP, https://dbgap.ncbi.nlm.nih.gov/), and the details of these datasets were summarized in Supplementary Table S1.

The CTA burden algorithm

To calculate CTA burden, a known CTA gene set was compiled from the CT antigen database (18), a database that documents experimentally validated and curated CTAs. The list of known CTA genes used in this study is provided in Supplementary Table S2. To calculate CTA burden, we compared the tumor expression data from the clinical trial datasets with the GTEx expression data for normal somatic tissues (n=13280). Immune privileged tissues such as brain, ovary, and testis as well as cell lines were excluded from the analysis. The expression data in the tumor dataset and GTEx dataset were normalized using quantile normalization. To mitigate batch effects between the clinical trial dataset (Supplementary Table S1), comprising tumor samples, and the GTEX somatic normal tissue dataset, which gathers normal tissue data from healthy individuals, we employed the "ComBat" R package. Although it was not feasible to match the outcome variable across these datasets, the application of ComBat addressed the baseline variations of TAAs that show outlier expressions in both tumor and normal samples.

We then determined the differential overexpression level ($\Delta exp$) between a selected tumor from a clinical trial dataset and non–immune privileged normal somatic tissues profiled by GTEx. For this calculation, we used the higher of two thresholds: 95% percentile of expressions in normal somatic tissues, or 98% percentile of the normal tissue counterpart where the tumor originates from, to compute level of overexpression for a CTA.

Then CTA burden was calculated using the following formula:

This algorithm allows quantitation of the cumulative overexpression pattern of CTAs with consideration of their overexpression levels. This method to some degree incorporated the level of $\Delta exp$ in the model via logistic sigmoid transformation and the $\lambda$ in the formula determines the slope of the sigmoid curve. The higher the $\lambda$ , the deeper is the Hill Slope, and the output is closer to counting the number of overexpressed CTAs.

Heterogenous expression profile (HEPA) analysis and calculation of TAA burden

HEPA analysis was performed using the TCGA Pan-cancer dataset, and the GTEx Pan-normal tissue dataset. For a given gene i in a given cancer type k, an outlier expression score was calculated based on an adjusted upper quantile mean of its expression in TCGA cancer type k, which is defined as the mean level of the 75th percentile to 95th percentile expression values of gene i. To evaluate the expression profile of candidate antigens in normal tissues, for the same gene i, its average expression in normal somatic tissue types was calculated based on trimmed mean using the GTEx dataset, which was used to calculated a normal expression depreciation penalty for gene i as previous described (19). The immune privileged organs such as brain, testis, and ovary were excluded. Then the HEPA score was calculated by subtracting the log2 transformed outlier expression score in cancer type k and the normal expression depreciation penalty. Genes were then ranked by their max HEPA scores across different cancer types, known protein coding genes with and a max HEPA cut off of 0.2 was considered as putative TAA encoding genes based on the optimal enrichment and detection of known CTAs. Genes encoding cytokines, or genes detected in normal blood (based on Human Protein Atlas database: https://www.proteinatlas.org) were filtered out. The final TAA gene set was generated by combining the CTA gene set with the putative TAA gene set, and TAA burden was calculated based on the same algorithm as the CTA burden detailed above.

Calculation of immune signatures and cell states

To calculate immune-related scores, we collected cancer immune-related gene sets as described by Trujillo (20), and accessed a T-cell exhaustion signature from a previous study that compared genes upregulated in exhausted CD8⁺ T cells compared to effector CD8⁺ T cells during virus infection (21). We then used a statistical method called Singscore (22) to calculate gene set scores for each tumor sample. To systematically quantitate immune cell states based on RNAseq data, we leveraged a computational tool called Ecotyper that utilizes transcriptome sequencing to characterize cell states and ecosystems in tumors (23). High Ecotyper CD8 S3 state was defined based on the cutoff of median + median absolute deviation based on the outlier expression profile of the score across tumors.

Statistical analysis

For interaction analysis, multiple logistic regression models were generated for pair-wise interactions of the TAA burden with each of the potential interacting factors and the multivariate regression models with or without the interaction term for each of the potential interacting variables were compared via Chi-Square test implemented in the ‘anova’ function, as we previously described (24). Two way and three-way interactions between TAB and confounding variables were analyzed by Johnson-Neyman plot using R package “interactions”. P-values of the box or dot plots were analyzed based on Welch’s T test. Two-tailed tests were used during the hypothesis generating process using the IMvigor210 dataset, and one-tailed tests were used when directional hypothesis was made to provide better statistical power on clinical trial datasets of small sample sizes. The correlation analyses of various ICB markers were performed based on Pearson’s statistics. The composite model for TMB and TAB were generated based on logistic regression modeling binary response. The cutoffs were determined based on mean or median plus median absolute deviation in accordance with the normality of the data. For data that showed strong outliers, such as Ecotyper CD8 S3 state, median plus median absolute deviation (MAD, default constant) was used as cutoff, for data that show normality, such as TAB or CTB, mean plus MAD was used as cutoff. For survival analysis, the P-values were calculated based on log-rank tests. Receiver operating characteristic (ROC) curves were used for analyzing the predictive values of biomarkers on binary responses. Multivariate logistic regression or Cox proportional hazard regressions were utilized to model binary response or overall survival respectively, while considering the presence of other covariates. For models consisting solely of continuous variables, p values of z statistics (Pr(>∣z∣)) were computed. In cases involving categorical covariates, a type-II ANOVA was employed on fitted logistic regression or Cox proportional hazard regression models to assess the independent significance of each covariate.

Data availability

The TCGA Pan-cancer gene expression datasets can be retrieved from https://xenabrowser.net, and GTEx normal tissue expression data can be retrieved from https://gtexportal.org/. The clinical trial datasets can be retrieved from GEO (https://www.ncbi.nlm.nih.gov/geo), dbGaP (https://dbgap.ncbi.nlm.nih.gov/), or EGA (https://ega-archive.org). The accession numbers of these clinical trial datasets are summarized in Supplementary Table S1. The CTA and TAA burdens, CD8 S03 and T-cell exhaustion scores for all datasets, putative TAAs predicted by HEPA, and immune gene sets used in this study are provided in Supplementary Data File S1. The R package for calculating CTA and TAA burdens is available on GitHub at: https://github.com/wangxlab/TAA-burden.

Development of a TAA burden algorithm based on recognized CTAs and putative TAAs.

In this study, we developed an algorithm to calculate the cumulative overexpression pattern of TAAs in tumors. Instead of simply counting the overexpressed TAAs, this algorithm transforms the overexpression levels using logistic sigmoid function and quantitates TAA cumulative overexpression pattern with minor consideration of the overexpression levels. This, to some degree, can attenuate the batch effects of clinical trial datasets with unmatched normal tissues. To systematically study the cumulative expression pattern of TAAs, we first compiled a known CTA gene list (n=243) from the CTD database (18), a database that documents experimentally validated cancer-testis (CT) antigens identified thus far. In addition, we sought to leverage the HEPA algorithm we developed (19), to identify putative TAAs based on the Pan-cancer RNAseq datasets from TCGA, and normal somatic tissue datasets from the GTEx Project. HEPA analysis utilizes specific algorithms designed to account for the unique expression profiles of TAAs: 1) prototype TAAs demonstrate remarkable overexpression in a limited subset of tumors, and 2) their expression in somatic normal tissues is restricted to immunologically privileged locations, such as germ cells (Fig. 1A). The efficacy of the HEPA algorithm has been previously validated through studies of a large panel of patient blood from multiple tumor entities (19). As in our previous study (19), ranking the human genes by their HEPA scores substantially enriched the known CTAs from the genome (Fig. 1B). Finally, we calculated a CTA burden that represents the overexpression pattern of known CTAs, and a TAA burden that represents the overexpression pattern of both known CTAs and putative TAAs.

Cumulative overexpression pattern of TAAs is associated with clinical response to PD-L1 blockade in mUC.

To examine the effect of TAB on clinical outcomes, we used one of the largest transcriptomic datasets of ICB clinical trials for non-melanoma cancer patients as the discovery set. This trial profiled a large cohort of metastatic UC (mUC) patients treated with atezolizumab in the IMvigor210 clinical trial by transcriptome sequencing (25). UC, also known as transitional cell carcinoma (TCC), originates from the urothelial cells of the urinary system, and is one of the highest TAA expressers among non-melanoma cancer entities. In the IMvigor210 trial, histology pure or predominant TCC patients were enrolled, and all patients were treated with 1,200 mg i.v. every 3 weeks until loss of clinical benefit (26,27). In this study, we defined patients with complete or partial responses (CR or PR) as responders, and those with stable or progressive disease (SD or PD) as non-responders. When applying our model to the RNA-seq data of pre-treatment mUC samples, we observed a significant increase in both CTA and TAA burdens in responders compared with non-responders. In particular, TAB exhibited a more pronounced and statistically significant association with treatment response (Fig. 1C). Next, we investigated the levels of CTA and TAA burdens across various molecular subtypes of bladder cancer within the IMVigor210 cohort. The basal/SCC-like subtype of Lund2, and the TCGA III/IV subtypes (recognized as basal subtypes), exhibited significantly elevated CTA and TAA burdens (Fig. 1D).

Clinicopathological variables that modify the association of CTAs with ICB treatment.

To investigate how clinicopathological variables influence the relationship between TAB and the effectiveness of ICB therapy, we conducted a comparative analysis of multivariate logistic regression models by incorporating versus excluding an interaction term for each clinicopathological variable. This revealed that PD-L1 staining on immune cells displayed the most pronounced interaction effect with the association of TAB and ICB benefits (Fig. 2A). PD-L1 expression in immune cells is a better predictive marker for ICB response than PD-L1 expression in tumor cells in certain cancer types such as UC and head and neck cancer (28,29). PD-L1 expression levels in immune cells (referred to as IC levels) can be classified into IC0, IC1, or IC2+ based on PD-L1 expression in <1%, ≥1% and <5%, or ≥5% tumor-infiltrated immune cells. In the IMvigor210 Trial, a favorable response to ICB treatment was observed in 16% (13/83) of IC0 patients (Supplementary Fig. S1). In our analysis, a significant difference in TAB was observed between responders and non-responders within the IC0 group, but not in the IC1 or IC2+ groups (Fig. 2B). The differences of TAB associated with ICB response within IC level groups were more pronounced compared to stratification by tumor cell PD-L1 levels. Stratified survival analysis based on IC levels indicated a significant survival difference between tumors categorized as TAB-high and TAB-low within the IC0 group (Fig. 1C, Supplementary Fig. S2). This finding suggests a potential variation in the immunogenicity of TAAs in accordance with TIME.

TAA burden is the most influential covariate of ICB benefit in the PD-L1 IC0 group.

We then compared TAB with other known predictive markers of ICB response, including TMB, intragenic rearrangement (IGR) burden (30), CD274 (PD-L1), T effector signature (31), CXCL13 (32), and pan-fibroblast TGF-β signature (25). TAB showed minimal to no correlation with these biomarkers (Fig. 2D). In a multivariate logistic regression model including all these markers, TAB was the only significant predictor of binary response in the IC0 group (Fig. 2E). To examine if this association could be driven by molecular subtypes or other clinical variables, we performed multivariate logistic regression modeling binary responses and Cox proportional hazard regression analyses modeling overall survival. Both models incorporated biomarkers such as TAB, CTB, TMB, molecular subtypes such as Lund2 and TCGA subtypes, and clinicopathological variables such as metastasis status, Bacillus Calmette–Guérin (BCG) vaccine administration, and platinum exposure. These analyses revealed that TAA and CTA burdens held the greatest influence on treatment outcomes in the IC0 group (Supplementary Fig. S3A), whereas TMB was the most influential antigen covariate on treatment outcome in the IC1-2+ groups (Supplementary Fig. S3B). In addition, TAA burden demonstrated a higher predictive value than CTA burden in our analysis, underscoring the advantage of including putative TAAs in our predictive model for better estimation of TAA burden (Supplementary Fig. S3A, S4).

PD-L1 IC level is associated with a cell ecosystem that reflects pre-existing antitumor immunity.

We investigated the impact of PD-L1 IC levels on the functional states of tumor-infiltrated immune cells and their potential contribution to tolerance towards TAAs. We leveraged a computational tool called Ecotyper that utilize transcriptome sequencing to characterize the states of a variety of tumor, stroma, and immune cell types in tumors (23). We correlated the resulting 71 cell states with PD-L1 IC levels across the entire IMVigor210 patient cohort. This analysis identified CD8 T cell S3, epithelial cell (EC) S4, and dendritic cell (DC) S03 as the most correlated cell states (Fig 3A). Among the bladder cancer molecular subtypes, this cell ecosystem is more enriched in basal-like tumors (Supplementary Fig. S5), which aligns with the enrichment of PD-L1⁺ immune cells in this subtype (33). EC S4 and DC S03 represent pro-inflammatory epithelial cells and mature inflammatory DC cells respectively (23). CD8 T cell S03 state expresses both markers of effector cells (i.e., IFNG, GZMB) and exhausted cells (i.e., LAG3) (23). Among different PD-L1 levels, the CD8 S3 state displayed a stronger correlation with immune-cell PD-L1 levels compared to tumor cell levels (Fig. 3B).

Next, we analyzed the correlation between CD8 S3 state and a T-cell exhaustion signature (21), and an effector T-cell signature (31). CD8 S3 state positively correlated with the T-cell exhaustion signature but negatively correlated with the effector T-cell signature (Supplementary Fig. S6). These findings suggest that PD-L1 IC stains may be indicative of a pro-inflammatory cell environment compromised by T-cell exhaustion, potentially resulting from pre-existing anti-tumor immunity. To ensure consistency in our analyses, we employed CD8 T cell S3 state as a stratification factor, which incorporates markers for both CD8⁺ T cells and the T-cell exhaustion state. In contrast, T-cell exhaustion signatures were derived from differentially expressed genes by comparing exhausted CD8⁺ T cells with effector CD8⁺ T cells. Practically, unlike the limited availability of PD-L1 data, CD8 T cell S3 state can be determined through deconvolution analysis of RNAseq data, offering a more accessible stratification method.

TAA burden correlates with ICB benefit in urothelial tumors of low CD8 T cell S3 state.

We speculated that the relationship between TAB and the benefits of ICB therapy may be weakened in an exhausted tumor immune context, as supported by the diminished predictive value of TAB observed with higher CD8 S3 state and higher T-cell exhaustion signature (Fig. 3C). In the subgroup analyses based on CD8 T cell S3 state or T-cell exhaustion signature, we observed significantly higher TAB in responders within the patient group exhibiting a non-elevated CD8 T cell S3 state or T-cell exhaustion signature (Fig. 3D). In the IMVigor210 cohort, a total of 10.1% of mUC patients exhibited high TAB levels and a non-elevated CD8 T cell S3 state. Within the TCGA primary bladder cancer cohort, the incidence of tumors with high TAB and low CD8 T cell S3 state increased as tumor stages progressed: 9.2% in stage II (n=130), 14.8% in stage III (n=135), and 15.2% in stage IV (n=132). These data suggest a growing proportion of patients with high TAB and low CD8 T cell S3 state as the tumor stage advances.

Further stratified interaction analysis revealed that in tumors with low TMB, TAB exhibited a strong predictive association only when CD8 S3 state was diminished to a certain level (Supplementary Fig. S7A). Conversely, in tumors with a non-elevated CD8 S3 state, the predictive value of TAB was evident at lower levels of TMB (Supplementary Fig. S7B). This suggests that TMB may serve as a distinct modifying factor, which aligns with the lack of correlation between TMB and PD-L1 levels (Supplementary Fig. S8) (34). We thus stratified the patients with unelevated CD8 S3 state into TMB high and low groups based on median TMB levels. Among the TMB low group, the AUROC of TAB, CTB, and TMB for predicting ICB benefit were 0.87, 0.78, and 0.57 respectively, and combining TAB with TMB did not result in better predictive values (Fig. 3E, left), while in TMB high tumors, the AUROCs for TAB and CTB decreased to 0.70 and 0.62 respectively (Fig. 3E, right).

Association of TAB with clinical benefit of UC patients treated with various anti-PD1/PD-L1 agents.

To test if this finding could be generalizable to various anti-PD1/PD-L1 agents, we examined an independent dataset for mUC as validation set; in this dataset, patients were treated with either anti-PD1 such as Pembrolizumab or Nivolumab, or anti-PD-L1, such as Atezolizumab, Avelumab, or Durvalumab (35). We defined clinical benefits as complete or partial responses following ICB and histology pure TCC are included in this analysis to match the pathological characteristics of the IMVigor210 cohort. Consistent with our previous findings, TAB was only significantly higher in the responders compared to non-responders within the patient group with unelevated Ecotyper CD8 S3 state or T-cell exhaustion signature (Fig. 4A). The effect was diminished in T-cell exhausted tumors. The AUROC for TAB and CTB were 1.0 and 0.9 in the TMB low group, and 0.86 and 0.8 in the TMB high group (although such high AUROC could be biased due to the small sample size). TMB did not add additional predictive effect when combined with TAB (Fig. 4B). Among the tumors with unelevated exhausted CD8⁺ T-cell state, TAB high tumors (7/41) showed significantly better progression-free survival (PFS) and overall survival (OS) compared to the rest of the tumors (Fig. 4C). In fact, only 2 out of 7 patients with high TAB progressed, whereas all patients with unelevated TAB progressed. About 86% (23/34) patients with high TAB survived at the end of follow-up, compared to 14% (1/7) patients with unelevated TAB.

Taken together, our data suggest the association of TAB with ICB benefits in UC with unelevated T-cell exhaustion signature: in the IMvigor210 cohort, TAB showed a positive predictive value of 44% at a sensitivity of 84% for predicting ICB benefits (binary responses); in the UNC cohort, TAB showed a positive predictive value of 60% at a sensitivity of 86% (Supplementary Fig. S9). Next, we sought to examine if such associations were predictive rather than prognostic. As neither the IMvigor210 cohort nor the UNC cohort have a non-ICB treated control arm, we sought to test the prognostic effect of TAA and CTA burdens in the TCGA patient cohort and examined their associations with PFS and OS in the bladder cancer patients stratified based on T-cell exhaustion state. Our results showed that neither high TAA burden nor high CTA burden were significantly associated with PFS and OS in tumors from the TCGA bladder cancer patient cohorts similarly stratified by CD8⁺ T-cell exhaustion states (Supplementary Fig. S10). This data suggests that the associations of TAB with patient outcomes are more likely predictive of ICB benefit rather than prognostic.

TAA burden associates with clinical benefit in head and neck cancer patients treated with anti-PD1.

Next, we sought to explore whether this finding could be generalized to other cancer types with overall high CTA burden but moderate to low TMB. To identify suitable cancer types for this analysis, we plotted the median CTA burden against median TMB levels for major solid tumors profiled by TCGA. Processed gene expression and mutation data from TCGA were used for this analysis. Among the solid tumor types, head and neck squamous carcinoma (HNSC), liver hepatocellular carcinoma (LIHC), uterine corpus endometrial carcinoma (UCEC) and adenoid cystic carcinoma (ACC), were the cancer types that showed highest TAB levels but moderate to low TMB (Fig. 5A). Among these cancer types, our literature investigations revealed two independent clinical trial datasets for HNSC, which we then used as additional validation sets. HNSC was observed to have relatively lower TMB level than bladder cancer, and the TAB levels were higher in Human Papillomavirus (HPV) negative tumors than positive tumors. In both datasets, clinical or pathologic response data following ICB treatment were available, but TMB or patient survival data were not provided.

We first accessed a neoadjuvant trial of advanced HNSC patients treated with 240 mg nivolumab every 2 weeks for two doses prior to surgery, in which most of the patients were HPV-negative (36). Tadalafil was administered in half of the patients in addition to nivolumab, which did not result in difference in response rates (36). Thus, patients from both treatment arms were included in our analysis. In pretreatment samples, TAA burden was significantly higher in responders than non-responders (p=0.045), the difference was not significant for CTA burden (p=0.16) (Fig. 5B). In posttreatment samples, both TAA burden and CTA burden showed more obvious reduction following ICB treatment in responders compared to non-responders. When stratified by exhausted CD8⁺ T-cell state, HNSC tumors with unelevated CD8 S3 signature showed a trend of more significant difference in TAB levels comparing responders to non-responders (Supplementary Fig. S11). Among the HNSC tumors with unelevated exhausted CD8⁺ T-cell state, the AUROCs of TAB and CTB for predicting ICB response were 0.71 and 0.6 respectively (Fig. 5C, left panel), and among all subjects, the AUROC of TAB and CTB were 0.68 and 0.66 respectively (Supplementary Fig. S12, left panel).

To further verify this result, we accessed another phase II trial dataset for neoadjuvant Pembrolizumab in advanced HPV-unrelated HNSC patients (37). In baseline tumors, the AUROC of TAB and CTB for pathologic responses were 0.68 and 0.60 in all patient subjects (Supplementary Fig. S12, right panel), and 0.85 and 0.88 respectively in patients with unelevated CD8 T cell S3 signature (Fig. 5C, right panel). These data suggest the association of TAB with ICB benefit in HNSC patients, particularly in tumors with low levels of exhausted CD8⁺ T cells.

TAA burden may be reduced with pre-existing T-cell immunity in cancer.

To assess the influence of pre-established T-cell immunity on TAB, we evaluated the association of TAB with IC levels in the IMVigor210 cohort. This revealed a significantly higher TAB in the ICB responders characterized by an IC0 status compared to those in the IC1-2+ group (Fig. 2B). To validate these findings, we stratified patients with UC into three distinct groups according to the CD8 T cell S3 state tertiles across the IMVigor210 and Rose datasets. We consistently found that both CTA and TAA burdens were elevated in ICB responders with low levels of CD8 T cell S3 signature compared to intermediate or high levels, while TMB persistently exhibited higher levels in responders across all IC categories (Fig. S13). This pattern suggests that TAAs overexpressed in tumors without pre-existing antitumor immunity, may act as primary antigenic targets following an ICB-induced immune response.

To expand our investigation into the association of CTA, TAA, and TMB with the CD8 T cell S3 state, we utilized TCGA pan-cancer data for a broader perspective. Our analysis reinforced the above findings: with increases in CD8 S3 state, both CTA and TAA demonstrated a substantial reduction, in contrast to TMB, which maintained high levels in tumors with increasing CD8 S3 state (Fig. 6A). Moreover, in most tumor types, heightened CD8 S3 states correlated inversely with TAB (Fig. 6B). These data collectively suggest that TAB may be diminished in tumors with pre-existing T-cell immunity, possibly resulting from immunoediting processes (38). In addition, we observed that melanomas with high T-cell exhaustion displayed a greater TAB, suggesting a deeper state of TAA tolerance in melanoma compared to other cancer types (Fig. 6B).

Assessing the influence of CD8 T cell S3 state on the association of TAB with ICB response in melanoma

Considering that previous studies indicated TAAs lack predictive value for ICB effectiveness in melanoma (8), we aimed to assess whether stratification by CD8 T cell S3 state might identify predictive patterns. To this end, a large-scale whole-transcriptomic dataset was necessary. We managed to identify a single suitable dataset that included transcriptomic profiles from over 100 melanoma patients who had undergone ICB treatment, along with corresponding TMB data (Liu et al. (39)). This stratification revealed that in CD8 S3 high melanoma, TMB was a significant predictor of ICB response (AUC=0.71). Whereas among melanomas with lower CD8 S3 scores, CTB and TAB provided moderate predictive utility for ICB success in cases where TMB exceeded the median level (Supplementary Fig. S14), differing from our findings in TCC and HNSC. More interestingly, discounting patients who received steroids from the analysis enhanced the predictive values of TMB, CTB, and TAB across all groups, aligning with evidence that steroids can impair ICB effectiveness (40). Our result is consistent with the recent report that in mutation-rich melanoma, anti-PD1 can expand TAA-reacting TILs which correlates with treatment efficacy (41), whereas in mutation poor tumors immune-stimulating agents will be required. This suggests that mutations may provide the co-stimulation signal required for ICB-induced T-cell activation in melanoma, whereas urothelial carcinoma may be subjected to other immune induction mechanisms such as platinum or BCG commonly administered prior to ICB. Larger-scale transcriptome sequencing studies of melanomas with low T cell exhaustion levels are needed to better understand the relationship between TAB and immunotherapy response.