Blockade of LAG3 and PD1 leads to co-expression of cytotoxic and exhaustion gene modules in CD8+ T cells to promote antitumor immunity

Summary

Relatlimab (rela; anti-LAG3) plus nivolumab (nivo; anti-PD1) is safe and effective for treatment of advanced melanoma. We designed a trial (NCT03743766) where advanced melanoma patients received rela, nivo or rela+nivo to interrogate the immunologic mechanisms of rela+nivo. Analysis of biospecimens from this ongoing trial demonstrated that rela+nivo led to enhanced capacity for CD8+ T cell receptor signaling and altered CD8+ T cell differentiation, leading to heightened cytotoxicity despite retention of an exhaustion profile. Co-expression of cytotoxic/exhaustion signatures was driven by PRDM1, BATF, ETV7 and TOX. Effector function was upregulated in clonally expanded CD8+ T cells that emerged after rela+nivo. A rela+nivo intratumoral CD8+ T cell signature was associated with favorable prognosis. This intratumoral rela+nivo signature was validated in peripheral blood as an elevated frequency of CD38+TIM3+CD8+ T cells. Overall, we demonstrated that cytotoxicity can be enhanced despite the retention of exhaustion signatures, which will inform future therapeutic strategies.

Graphical Abstract

In Brief

Analysis of biospecimens from a phase II study of advanced melanoma shows that combined blockade of LAG3 and PD1 modulates differentiation of CD8+ T cells by enhancing responses to T cell receptor and IFN-γ signaling, leading to enhanced effector functions despite retention of an exhaustion profile.

Introduction

Immune responses to cancer and chronic viral infection are limited by CD8+ T cell exhaustion^1,2. Inhibitory receptors expressed on the surface of exhausted CD8+ T cells activate downstream molecular pathways leading to reduced proliferation, metabolic dysfunction, cytokine production and cytotoxic effector activity. Blockade of inhibitory receptors (IRs) cytotoxic T-lymphocyte associated protein 4 (CTLA4)³ or the programmed cell death protein 1 (PD1) pathway has led to remarkable successes in the treatment of cancer, but not all patients benefit^4,5. Based on the premise that different IRs contribute to CD8+ T cell exhaustion through non-redundant mechanisms, combination therapy targeting multiple IRs may improve CD8+ T cell function. Indeed, dual blockade of CTLA4 and PD1 leads to a long-term survival benefit in >50% of melanoma patients, but also leads to a higher frequency of severe immune-related adverse events^6-8. More effective and safer combinations of immunotherapy are needed to overcome CD8+ T cell exhaustion and improve clinical outcomes, both in melanoma and other solid tumors.

Many therapeutic combinations targeting IRs have been proposed to restore cellular immunity against cancer and chronic viral infections. Antibodies targeting CTLA4 and PD1 or its cognate ligand PD-L1 were the first to enter the clinic and the first to be used in combination. LAG3 is another IR expressed on CD8+ T cells in the setting of chronic antigen stimulation. Functionally, LAG3 inhibits the expansion of T cells and the number of cells entering the memory T cell subset⁹. Preclinical data in murine models of chronic viral infection and cancer revealed that simultaneous blockade of the IRs LAG3 and PD1 led to synergistic reinvigoration of CD8+ T cell mediated immunity^10,11, suggesting that these IRs function through distinct pathways. These preclinical data demonstrated that blockade of PD-(L)1 and LAG3 in combination could enhance antitumor immunity in patients with cancer.

Recently, this hypothesis has been validated in clinical trials. In patients with advanced melanoma, the fixed-dose combination of anti-LAG3 (relatlimab; rela) and anti-PD1 (nivolumab; nivo) led to superior progression-free survival versus nivo alone¹². These results led to the FDA approval of this combination in early 2022. Similar clinical efficacy was recently observed in patients with advanced melanoma that had progressed following prior PD-(L)1 therapy¹³. Further evidence for the efficacy of this combination strategy has recently come from the neo-adjuvant setting of resectable stage III melanoma, with pathological response in 70% of patients and recurrence-free survival at two years for 92% of patients who achieved a pathological response¹⁴. Overall, clinical evidence indicates that rela+nivo leads to improved outcomes in patients with melanoma across a spectrum of disease states and stages.

While clinically effective, the mechanism(s) underlying the enhanced antitumor immunity observed following rela+nivo combination treatment are unknown. Understanding the molecular consequences of combination therapy and whether these consequences are distinct from those of the respective monotherapies has important implications for future development of combinatorial immunotherapies. Following PD1/PDL1 blockade, a subset of CD8+ T cells expressing CXCR5 undergo a proliferative burst^15-17. Consistent with this, there is an increased frequency of cycling CD8+ T cells observed in melanoma patients following blockade of the PD1 pathway, with the ratio of CD8+ T cell reinvigoration to tumor burden associated with subsequent outcome¹⁸. Although the molecular mechanisms that contribute to reinvigoration of cellular antitumor immunity following blockade of the PD1 pathway are at least partially understood, insight into the mechanism(s) underlying either enhanced antitumor immunity or resistance following combinatorial therapies are required to inform future immunotherapeutic strategies and combinations.

To address this question, we designed a randomized, open label, phase 2 study where patients were randomly allocated to receive rela alone, nivo alone, or rela+nivo for a 4-week lead-in phase (NCT03743766). There were two major objectives of this trial: (1) mechanistically dissect the transcriptional changes across immune populations following dual blockade of PD1 and LAG3, and (2) determine the mechanism(s) that promote antitumor efficacy of LAG3 and PD1 blockade in advanced melanoma. The clinical efficacy of rela+nivo blockade has now been reported in a phase III trial¹² and in the neoadjuvant setting¹⁴. However, the mechanisms governing this efficacy at the cellular and molecular level in patients are unknown. Blood and tumor biopsies were obtained as part of this study at baseline and 4 weeks post-treatment (with subsequent blood draw and voluntary tumor biopsies at 16 weeks). After the 4-week lead-in phase, all patients commenced rela+nivo combination therapy. Whereas this trial design negated our ability to determine clinical efficacy, as all patients received the rela+nivo combination therapy early in their treatment course, it did allow us to gain unique mechanistic insight into both the single and combo modalities, especially and importantly including rela alone. This unique trial design permitted procurement of biospecimens for mechanistic dissection of the effects of LAG3 and PD1 blockade alone and in combination across immune populations while still allowing all patients to receive the clinical benefit of combination therapy. Understanding the changes in immune populations following combination immunotherapy remains an important topic that will inform pre-clinical and clinical combinational immunotherapy strategies. Assessment of antitumor activity remains an independent objective of this ongoing trial.

Our analysis of the biospecimens from this trial lead to several unique conclusions because we were able to assess the changes associated with rela alone, nivo alone and rela+nivo in combination. Single-cell RNAseq (scRNAseq) data from blood and tumor-resident immune populations demonstrated that rela+nivo treatment promoted T cell activation despite terminal exhaustion signatures remaining in the tumor infiltrating CD8+ T cells. Orthogonal analyses also showed that CD8+ T cells had enhanced responses to T cell receptor signaling following rela+nivo treatment. Unique to combination therapy, we observed a modulation of CD8+ T cell differentiation leading to the co-expression of cytotoxic and exhaustion modules. We also found CD8+ T cell receptor clones that were undetectable prior to therapy but “emerged” following therapy were enriched for these functional signatures including interferon gamma (IFN-γ) response and T cell activation despite retention of transcriptional hallmarks of exhaustion. Gene regulatory network inference revealed that the transcription factors PRDM1, BATF, ETV7, and TOX together governed gene modules associated with cytotoxicity and exhaustion within the same cells. We identified an intratumoral transcriptional signature of rela+nivo therapy, which was found to be associated with favorable prognosis in multiple external cohorts. CD38 and HAVCR2 (gene for TIM3) were found to be key components of this rela+nivo signature, and an elevated frequency of CD38+TIM3+CD8+ T cells was present in blood following rela+nivo treatment, which offered protein validation of our transcriptional signature.

Overall, our study demonstrates a complex molecular mechanism underlying combination blockade of LAG3 and PD1 distinct from what would be expected from either monotherapy alone. This mechanism relies on enhanced responses to T cell receptor signaling, modulation of CD8+ T cell differentiation, and co-expression of cytotoxic and exhaustion gene modules. These findings have implications for the understanding of combination immunotherapy and the design of future immunotherapy strategies that are both safer and more effective than those currently available.

Results

Study design, patient demographics and identification of cell types

The objective of this study was to perform mechanistic dissection of the effects of rela, nivo, and rela+nivo on immune populations in peripheral blood and within the tumor microenvironment in patients with advanced (unresectable or metastatic) melanoma naïve to prior immunotherapy and targeted therapy. To address this question, we utilized biospecimens from an ongoing trial (NCT03743766) in which patients were assigned to one of 3 arms for a 4-week lead-in period: (a) rela alone, (b) nivo alone, and (c) rela+nivo in combination. Details regarding the study are presented in a CONSORT diagram (Figure S1). In total, 61 patients were screened to identify 43 eligible patients who were allocated to one of the 3 lead-in arms (15 patients to nivo monotherapy, 14 patients to rela monotherapy, 14 patients with rela + nivo combination therapy). Therapy was discontinued due to toxicity or disease progression in 9 patients from the nivo arm, 12 patients from the rela arm, and 6 patients from the rela + nivo arm. Following 4 weeks of lead-in therapy, all patients received rela+nivo for all subsequent treatment cycles. Peripheral blood was drawn, and tumor core biopsies were obtained prior to therapy and following the 4-week lead-in period and single-cell RNAseq (scRNAseq) was performed on these specimens (Figure 1A; Methods). Additional blood samples were obtained at week 16 following treatment and optional third tumor core biopsies were collected. Tumor biopsies were always collected from the same pre-specified tumor target site pre- and post-treatment to limit sampling bias. Importantly, this study design permitted the assessment of changes in transcriptional signatures between baseline and four weeks of therapy for each immunotherapeutic agent alone or in combination.

Figure 1. Nivolumab, relatlimab, and nivolumab + relatimab in patients with metastatic melanoma is most likely to have a cell-intrinsic effect on CD8+ T cells.

A) Patients with advanced melanoma naïve to prior immunotherapy were enrolled and stratified 1:1:1 to an initial lead-in period of treatment with either relatlimab (rela) alone, nivolumab (nivo) alone, or rela + nivo. Blood and tumor biopsies were obtained at baseline and following 4 weeks of therapy. Additional blood draws and optional tumor biopsies were also performed at week 16. B) UMAP showing the Identification of immune cell subsets across 293,732 cells in PBMC (left panel) and TIL (right panel) across all timepoints and samples. C) Dot plots where the size of the dot represents the frequency of cells positive for a given gene and the color represents the magnitude of expression or the same gene. CD8+ T cells express the highest levels of LAG3 and PDCD1 in both PBMC (top panel) and TIL (bottom panel). D) UMAPs showing expression levels of PDCD1 in PBMC (left panel) and TIL (right panel) across all immune cell subsets. E) UMAPs showing expression levels of LAG3 in PBMC (left panel) and TIL (right panel) across all immune cell subsets.

Patient demographics across the three treatment arms included a median age of 68 years, 33% female, 87% stage IV and 51% with BRAF mutations (Table S1; Data S1). A total of 13 patients were evaluated from the rela lead-in arm, 12 from the nivo lead-in arm, and 14 patients from the rela+nivo lead-in arm. scRNAseq was performed in all instances where sufficient cells were available (Table S1; Data S1). After quality control, we recovered a total of 293,732 cells from all patients and conditions, including 199,950 from PBMC and 93,782 from TIL (Methods). Major immune lineages were identified based on expression of canonical immune markers across clusters (Figure 1B; Methods). We next evaluated the expression of LAG3 and PDCD1 across all immune cell subsets at baseline (Figure 1C), which revealed that these genes were most highly expressed in CD8+ T cells (Figure 1D-E). As such, CD8+ T cells are most likely to be intrinsically influenced by blockade of LAG3 and/or PD1 and are the major focus of downstream analyses.

Rela, nivo and rela+nivo lead to distinct changes within the CD8+ T cell compartment

We next identified clusters corresponding to distinct states of CD8+ T cells. We isolated 75,160 CD8+ T cells in silico across all patients and conditions and performed Leiden-based clustering (Figure 2A). This led to the identification of 9 unique clusters of CD8+ T cells (Table S2), with differential enrichment between PBMC and TIL (Figure 2B). We next related transcriptional signatures across CD8+ T cell clusters to known CD8+ T cell states and biological processes using curated gene sets from the Molecular Signatures Database and published sources (Methods). This analysis allowed us to classify clusters by canonical CD8+ T cell states such as naïve, central memory, effector, effector memory, tissue resident, and progenitor and terminal exhaustion (Figure 2C; Table S2). Additionally, we also classified clusters by biological states such as IFN-γ response, enrichment of genes downstream of TCR signaling, and cell cycle. Cluster 5 had relatively low enrichment for all gene sets tested, but had upregulation of CXCR5 and TCF7 plus genes associated with early activation, suggesting that it is a transitional cluster that bears some homology to the previously described stem-like CD8+ T cells (Data S1)¹⁹. Overall, we were able to relate our CD8+ T cell clusters to previously described states using a curated battery of gene sets.

Figure 2. Immunotherapy with relatlimab alone, nivolumab alone, or relatlimab+nivolumab in combination in PBMC and TIL leads to distinct transcriptional phenotypes with combination therapy promoting enhanced T cell receptor signaling.

A) A total of 75,160 CD8+ T cells were bioinformatically isolated from all immune cells and re-clustered for downstream analysis. Leiden clustering identified 9 distinct clusters of CD8+ T cells in PBMC (36,900 cells; left panel) and TIL (38,260 cells; right panel). B) Same UMAPs of CD8+ T cells as (A) but visualized using 2-dimensional density distribution. Clusters 1, 3 and 4 were enriched with cells from PBMC, while clusters 6, 7, 8 and 9 consisted of cells derived from TIL. Clusters 3 and 5 were present in both PBMC and TIL. C) Gene set enrichment analysis across CD8+ T cells clusters revealed association between clusters and biological states of CD8+ T cells. D) Confidence intervals (95%) and enrichment scores from linear mixed effects models between baseline and treatment groups at week 4 in PBMC and TIL for gene sets associated with CD8+ T cell function. Notably, rela+nivo treatment led to higher levels of enrichment for gene sets associated with TCR signaling in PBMC and TIL. P values were calculated using post-hoc comparisons for linear mixed effects models and controlling for multiple comparisons using a two-sided false discovery rate of 5%. ***, p<0.001; ** p<0.01; * p<0.05. E) Schematic depicting the experimental approach for quantifying phospho-protein levels downstream of T cell receptor signaling by flow cytometry. F) Rela+nivo treatment led to a statistically significant increase in phospho-SLP76 (i.e. early T cell receptor signaling) at week 4 compared to baseline in PBMC from MEL patients in this study. P values were calculated using paired T tests.

We next assessed whether CD8+ T cell states were related to treatment conditions. Using linear mixed effects models (Methods), we assessed whether gene set scores were related to lead-in therapy at 4 weeks post-treatment across all CD8+ T cells. In PBMC, T cell activation and IFN-γ response gene sets were among the most statistically significant across treatment groups (Table S2). In TIL, terminal exhaustion, IFN-γ response, and TCR signaling were among the most significantly differential gene sets (Table S2). Further dissection of these signatures revealed that nivo alone reduced terminal exhaustion in TIL, but rela alone and rela+nivo had no effect on terminal exhaustion (Figure 2D). Surprisingly, despite the absence of any change in terminal exhaustion, we found that a gene signature associated with TCR signaling was higher at 4 weeks following rela and rela+nivo in TIL (Figure 2D). In PBMC, TCR signaling was also elevated following rela+nivo (Figure 2D). Importantly, there were no significant differences in the distributions of CD8+ T cells across UMAPs and similar trends were observed in terms of enrichment for these key transcriptional signatures regardless of BRAF disease status (Data S1). Further corroborating these results, we also found that cross-over to rela+nivo (from either rela monotherapy or nivo monotherapy) led to increases in TCR signaling at 16 weeks for rela and nivo lead-in therapies (Data S1). In summary, rela and rela+nivo lead to increased TCR signaling despite the retention of exhaustion.

Our observation that gene sets associated with TCR signaling were elevated following rela+nivo led us to perform flow cytometry for phosphorylated proteins associated with the T cell receptor signaling pathway (Methods; Figure 2E). To evaluate TCR signaling ex vivo, we thawed cryopreserved PBMC from patients at baseline and week 4 following lead-in treatment and stimulated T cells in culture by crosslinking the TCR and CD28 using biotinylated anti-CD3 and anti-CD28 antibodies plus streptavidin (Methods; Figure S2A-B). Comparing baseline to 4 weeks samples revealed that pSLP76 levels were elevated in CD8+ T cells from PBMC only following rela+nivo treatment (p=0.019; Figure 2F and Figure S2C). Thus, rela+nivo treatment appears to cooperatively enable better transduction of TCR signals compared to rela or nivo alone.

Rela+Nivo modulates CD8+ T cells differentiation, promoting overlapping cytotoxic and exhaustion states

We next evaluated whether differentiation trajectories differed across the treatment groups. To achieve this, we leveraged RNA velocity²⁰ (Methods) to identify vector fields of differentiation and a pseudotemporal ordering of CD8+ T cells (Figure 3A). Briefly, RNA velocity leverages the ratio of unspliced (that is, nascent) and spliced transcripts to infer differentiation trajectories from scRNAseq data. This permits the construction of a relative ordering of cells along a differentiation trajectory by assigning a pseudotime to each cell. Importantly, pseudotime corresponds to relative point along a trajectory and not to an actual temporal scale. Our pseudotemporal ordering was consistent with known biology, beginning with naïve/memory CD8+ T cells and ending with terminally differentiated CD8+ T cells (Figure 3B). Thus, our global pseudotemporal ordering could be dissected to understand how each therapy influences CD8+ T cell differentiation.

Figure 3. Differentiation of CD8+ T cells in PBMC and TIL is uniquely affected by relatlimab, nivolumab, and relatlimab+nivolumab, with the combination therapy leading to co-expression of exhaustion and effector gene modules.

A) UMAPs showing RNA velocity streams (left panel) and inferred pseudotime (right panel) for all CD8+ T cells from PBMC and TIL and baseline and week 4. B) Ordering of Leiden clusters (from Figure 2A) by pseudotime reveals a differentiation trajectory associated with activation and subsequent terminal differentiation. C) Pseudotemporal inference of gene dynamics over time revealed 10 distinct gene modules with discrete temporal dynamics. Cells are ordered by increasing pseudotime and hierarchically clustered to identify gene modules. D) Pseudotime was also broken down into bins to smooth gene expression patterns. Proportion of PBMC and TIL within each pseudotemporal bin (top panel). Median gene module score across baseline and treatment groups for exhaustion (middle panel) and cytotoxic (bottom panel) temporal gene modules. E) Rela+nivo treatment led to substantial co-expression of exhaustion and cytotoxic temporal gene modules in the same cells versus baseline and rela or nivo monotherapy.

We next asked which genes differed significantly in their expression levels during differentiation (Methods; Figure 3C). Hierarchical clustering of the top 50 genes related to pseudotime across baseline and each treatment group yielded 10 distinct gene modules that were associated with differentiation. To assess differences in gene module temporal dynamics across treatment groups, we evaluated the median enrichment score for each of the 10 gene modules across pseudotime, binning pseudotime into 10 bins to smooth the expression patterns. To better understand the relative contribution of PBMC and TIL to each pseudotime bin, we visualized the frequency of cells from each site across pseudotime bins (Figure 3D, top panel). We found that several of these gene modules had differential temporal dynamics across treatment groups, particularly at later pseudotime (Figure S3). This temporal analysis demonstrated that rela and rela+nivo led to elevated levels of terminal exhaustion (Figure 3D, middle panel). Despite these high levels of terminal exhaustion, rela+nivo treatment led to elevated cytotoxic gene module scores particular at the later stages of differentiation (Figure 3D, bottom panel). Interestingly, these changes began at early stages of differentiation and were maintained throughout the subsequent stages of differentiation.

Finally, we evaluated whether the elevation in both terminal exhaustion and cytotoxic gene modules occurred in the same cells following rela+nivo treatment. To interrogate this question, we used density plots to evaluate the co-expression of the exhaustion module and cytotoxic module in the same cells (Figure 3E; Methods). We found that 38% of CD8+ T cells had high levels of co-expresion for the cytotoxic and exhaustion modules following rela+nivo treatment, which was substantially higher versus all other groups. Nivo alone led to lower levels of exhaustion, and rela alone had similar levels of terminal exhaustion to rela+nivo but did not upregulate cytotoxicity. Overall, rela+nivo modulated CD8+ T cell differentiation towards a unique terminal cell state that simultaneously co-expressed genes associated with cytotoxicity and exhaustion.

Emergent CD8+ T cell clones are enriched for functional signatures following immunotherapy

We next evaluated the transcriptional signatures of clonally expanded CD8+ T cells in PBMC and TIL. First, identified quartiles of CD8+ T cells with clonally expanded T cell receptors (TCRs) based on TCRβ complementarity determining region 3 sequences (CDR3; Figure 4A; Methods). TCRs were considered clonally expanded if present with a count of at least three. Next, we visualized the distribution of clonally expanded CD8+ T cells from each quartile across CD8+ T cell UMAPs and found that the most clonally expanded cells were present in the intermediate exhaustion, activated transitional, and terminally exhausted / T cell activated clusters (Figure 4B). We also assessed overlap of clones across clusters with respect to baseline and the 3 lead-in groups. At a baseline, there were two sets of clusters with overlapping TCRβ CDR3 clones: (i) terminal exhaustion / T cell activation, the terminal exhaustion / progenitor exhaustion 1 cluster and the cell cycle cluster; and (ii) the activated transitional and intermediate exhaustion cluster (Figure S4A). Nivo alone had substantial overlap between the activated transitional cluster and the intermediate exhaustion and terminal exhaustion /T cell activation clusters (Figure S4B). Rela alone showed a similar pattern to baseline (Figure S4C). Rela+nivo led to a high degree of overlap between the terminal exhaustion / T cell activation and cell cycle clusters (Figure S4D). Overall, we found significant overlap of TCRβ CDR3 CD8+ T cell clones across clusters and that this overlap differed by treatment.

Figure 4. Rela, nivo, and rela+nivo lead to emergent CD8+ T cell clones in PBMC and TIL, with rela+nivo emergent clones exhibiting higher exhaustion and functional gene signatures.

A) Distribution of counts from CD8+ T cell TCRβ CDR3 sequences and separation of the degree of clonal expansion of TCRβ complementarity determining region 3 (CDR3) sequences into quartiles. B) Same UMAPs of CD8+ T cells as Figure 2A but showing the distribution of clonally expanded CD8+ T cells by TCRβ CDR3 quartiles. C) Density plots showing the distribution of TCR clones that were present at baseline and following 4 weeks of therapy (i.e., shared clones) in PBMC and TIL. D) Density plots showing the distribution of TCR clones only present following 4 weeks of therapy (i.e., emergent clones) in PBMC and TIL. E) Quantification of the frequency of CD8+ TCR clones in a cluster that are from emergent clones. No comparisons were statistically significant, but nivo trended towards a higher frequency of emergent clones in intermediate exhaustion and activated transitional clusters, while rela and rela+nivo lead to a high fraction of emergent clones in the terminal exhaustion / T cell activated cluster. F) Emergent clones were enriched for exhaustion and functional gene signatures compared to shared clones. Emergent clones following rela+nivo had substantially higher levels of terminal exhaustion (top panel), interferon gamma responses (middle panel), and T cell activation (bottom panel) compared to emergent clones following rela or nivo therapy. P values are derived from linear mixed effects models.

We next evaluated whether CD8+ T cells with shared CDR3 sequences between baseline and treatment had distinct transcriptional signatures. Here, we defined shared clones to be those that are detectable in either PBMC, TIL or both at baseline and following four weeks of treatment. To achieve this, we visualized the distribution of cells with shared CDR3 sequences in UMAPs across treatment groups at 4 weeks (Figure 4C). In PBMC, shared clones were mostly localized in the intermediate exhaustion and activated transitional clusters whereas in TIL shared clone localization differed by treatment status.

In addition to shared clones, we also found emergent clones that were not present at baseline but were detected following treatment with each therapeutic arm, consistent with previous observations in squamous and basal cell carcinoma²¹. Surprisingly, we found that the emergent clones were present in different regions of the UMAP versus the shared clones (Figure 4D). In PBMC, emergent clones were found predominantly in early activated and activated transitional clusters. In TIL, rela and nivo monotherapy led to diffuse patterns of emergent clones, while rela+nivo led to high enrichment of emergent clones in the terminally exhausted / T cell activation cluster. We next quantified the fraction of clones within a cluster derived from emergent clones from individual patients for each cluster in which at least 25% of the clones were emergent on average for at least one treatment group. TIL resident emergent clones following nivo were an elevated fraction of the total clones within the intermediate exhausted and activation transitional clusters, while rela and rela+nivo constituted an elevated fraction of clones in the terminally exhausted / T cell activation cluster (Figure 4E). Although these results were not statistically significant, emergent clones constitute a significant fraction of the TCR clones in several clusters.

We next identified differentially expressed genes and quantified enrichment across key functional gene sets from emergent clones at 4 weeks post-treatment versus shared clones. First, we found distinct transcriptional signatures from the clonally expanded cells from each treatment group (Figure S4E-G; Table S3). Comparisons of terminal exhaustion and gene signatures associated with IFN-γ response and T cell activated in shared versus emergent clones showed trends towards elevation across signatures in emergent clones, but these changes were most striking following rela+nivo treatment (Figure 4F). Importantly, the emergent clones following rela+nivo treatment showed enrichment for signatures associated with function while maintaining higher levels of terminal exhaustion. We also interrogated IFN-γ response signatures across other immune populations in PBMC and TIL (Figure S5), which showed significant upregulation of IFN- γ response across B cells, NK cells, T cells and myeloid cells in PBMC and elevated IFN-γ response in CD14-CD16+ monocytes plus conventional CD4+ T cells in TIL following rela+nivo. Taken together, we found that emergent CD8+ T cell clones exhibit enrichment for functional genes following each treatment, but most significantly following rela+nivo treatment.

A suite of transcription factors including PRDM1, BATF, ETV7 and TOX drive cytotoxicity and exhaustion gene modules following combination treatment

We next sought to assess whether there were distinct transcription factors that contributed to the gene expression profiles observed following each treatment. To perform this analysis, we used the miloR framework to relate treatment groups to cellular neighborhoods derived from graph-based clustering (Methods)²². In the context of our study, miloR uses an underlying graph network based on gene expression profiles to identify groups of cells (cellular neighborhoods on the graph network) that are associated with lead-in treatment groups in TIL. First, we identified 8 cellular neighborhoods across cells from all samples (Figure 5A, top left panel). This revealed enrichment in neighborhood 6 for rela alone, neighborhoods 5 and 8 for nivo alone, and neighborhoods 5, 7 and 8 for rela+nivo (Figure 5A). Differential gene expression between the cellular neighborhoods associated with each treatment group revealed distinct gene expression profiles associated with each lead-in group (Figure 5B; Table S4). Interestingly, rela+nivo shared genes with both nivo and rela monotherapy groups but also had a distinct set of upregulated genes. Amongst the top 100 differentially expressed genes from each treatment group, the overlap was minimal (Data S1). Consistent with our analysis of CD8+ T cell differentiation, CD8+ T cells expressed higher levels of cytotoxic molecules in conjunction with inhibitory receptors following rela+nivo therapy.

Figure 5. Combination therapy leads to co-expression of terminal exhaustion and effector function in TIL governed by a suite of transcription factors including PRDM1, BATF, ETV7 and TOX.

A) MiloR analysis was performed to identify cellular neighborhoods associated with CD8+ T cells in TIL following each lead-in therapy. Cellular neighborhoods are outlined (top left) and those that are positively associated with each treatment group are colored in yellow on the remaining panels. Unique groups of cellular neighborhoods were significantly associated with rela lead-in therapy, nivo lead-in therapy and rela+nivo lead-in therapy in TIL. B) Heatmap showing differentially expressed genes derived from the cellular neighborhood associated with each group. Notably, combination therapy led to higher levels of effector molecules (GZMB, IFNG), trafficking molecules (CXCR6), inhibitory receptors (PDCD1, LAG3, TIGIT, and CTLA4) and interferon responses (IFNG, ISG15 and IFI27). C) Bar plots showing the number upregulated genes that are putatively controlled by a given transcription factor (TF) inferred from SCENIC across the lead-in treatment groups. D) Network of the top 10 transcription factors governing the expression of significantly upregulated genes following rela+nivo treatment. Both exhaustion-associated genes and effector associated genes are governed by this suite of top transcription factors.

After identifying unique transcriptional signatures across the treatment groups at 4 weeks post-therapy, we next sought to infer the transcription factors that drive expression of these differentially expressed genes. We utilized the SCENIC framework^23,24 to derive transcription factor-driven gene modules, using all CD8+ T cells in PBMC and TIL across treatment groups as input (Methods; Table S5). We then evaluated the number of differentially expressed genes that are driven by a given TF across the treatment groups (Figure 5C) and constructed networks linking TFs to the genes they govern. While nivo and rela treatment each led to different suites of TFs that control distinct downstream gene expression patterns, it is notable that there were minimal genes associated with cytotoxicity governed by these TFs (Figure S6).

For the CD8+ T cells from rela+nivo treated patients, we found that the TFs PRDM1 (gene for BLIMP1), BATF, ETV7, ID2 and ZBED2 were among the top TFs responsible for differentially upregulated genes (Figure 5C). Interestingly, we found that the TFs PRDM1, BATF, and ZBED2 were responsible for driving genes associated with exhaustion and effector function following rela+nivo treatment (Figure 5D). Specifically, IFNG, PRF1 and HAVCR2were among the top 10 genes driven by PRDM1. Similarly, ZBED2 also drove expression of IFNG and HAVCR2. BATF, on the other hand, drove GZMB, IFNG and LAG3 expression. Conversely, the TFs ID2 and ETV7 had different roles, with ID2 driving expression of GZMB and IFNG in the absence of any inhibitory receptors and ETV7 driving genes associated with interferon responses. It was also interesting to find TOX in the top 10 TFs, given its role in promoting an exhausted lineage^25,26. Thus, in the context of combination treatment, a suite of TFs cooperatively govern effector function, interferon response, and terminal exhaustion.

Rela+nivo on-therapy signature is associated with favorable outcome and CD38/TIM3 expression

We next asked whether the on-therapy signatures we identified were related to clinical outcomes across multiple external cohorts. We first leveraged publicly available scRNAseq from PBMC in cohort of melanoma patients treated with rela+nivo (Figure 6A)²⁷ to interrogate whether our on-therapy rela+nivo signature was associated with response. After identifying T cell subsets (Figure 6B), we created gene set enrichment scores for all CD8+ T cells based on the top 100 differentially expressed genes from our on-therapy rela+nivo signature. We found that our TIL on-therapy rela+nivo signature was significantly elevated in CD8+ T cells at 1-month post-treatment in patients that achieved a clinical benefit (Figure 6C). This finding demonstrates that our on-therapy signature is associated with better outcomes following rela+nivo treatment in melanoma.

Figure 6. An on-therapy rela+nivo genes signature from TIL is associated with favorable outcome across cohorts and is present as a higher frequency of CD8+CD38+TIM3+ cells in PBMC.

A) Schema depicting the assessment of the on-therapy rela+nivo signature in an external cohort of MEL patients that received immunotherapy. B) scRNAseq of 46,679 CD8+ T cells, 54,928 CD4+ conventional T cells and 3,985 CD4+ regulatory T cells from PBMC of MEL patients at baseline, 1 month post-treatment and 3 months post-treatment with rela+nivo. C) Our TIL on-therapy rela+nivo signature was significantly elevated in CD8+ T cells from PBMC at 1 month post-treatment in patients achieving a clinical benefit to rela+nivo. P value is derived from a linear mixed effect model. D) Schema showing the derivation of 4 cohorts of MEL patients that received immunotherapy (MEL IO Cohorts) and derivation of on-treatment CD8+ T cell signatures from this study to evaluate associations with overall survival following treatment with immunotherapy. E) Overall survival analysis based on immune infiltration, rela signature, nivo signature, and rela+nivo signature. Each of these signatures was significantly associated with better overall survival in MEL patients that received immunotherapy across 4 independent studies. P values are derived from log-rank tests. F) Example flow cytometry results at baseline (top panel) and 4 weeks following treatment (bottom panel) across treatment groups for CD38 and TIM3 in CD8+ T cells from PBMC. G) The frequency of CD8+ CD38+TIM3+ T cells are shown across treatment groups, with rela+nivo leading to a significant increase in this population at week 4 versus baseline. Horizonal lines indicate the mean frequency of CD38+TIM3+CD8+ T cells in each timepoint/group. P values are from paired T tests.

We next evaluated whether our on-therapy signatures were associated with better prognosis in two other external cohorts. We first interrogate The Cancer Genome Atlas Skin Cutaneous Melanoma (TCGA-SKCM) by integrating bulk RNAseq outcome data (Methods; Data S1)²⁸. Using the top 100 differentially expressed genes from our treatment-derived CD8+ T cell gene signatures, we derived enrichment scores for each TCGA-SKCM patient (Methods; Data S1). We also utilized ESTIMATE²⁹ to evaluate the degree of immune infiltration across samples (Data S1). Unadjusted Cox regression analyses revealed that immune infiltration, and the rela and rela+nivo signatures were significantly associated with better overall survival in TCGA-SKCM (Data S1). Adjusted analysis showed that the rela+nivo signature was independently associated with better overall survival (hazard ratio [HR]=0.69, p=0.035; Data S1), while none of the other signatures were independently associated with outcome.

Next, we assessed whether these signatures were related to outcome following checkpoint blockade in melanoma patients using a set of 4 external datasets containing clinical and bulk RNAseq data (Methods; Figure 6D)^30-34. We evaluated the relationships between overall survival and immune infiltration, rela, nivo and rela+nivo enrichment scores. Further corroborating that the on-therapy CD8+ T cell signatures are reflective of effective antitumor immunity, unadjusted Cox regression analyses revealed that immune infiltration, rela, nivo and rela+nivo signatures were significantly associated with better overall survival across these cohorts of immunotherapy treated melanoma patients (Figure 6E). Taken together, these analyses suggest that on-therapy transcriptional signatures associated with monotherapy and combination therapy were all associated with favorable outcome, demonstrating that these CD8+ T cell signatures are likely reflective of antitumor immunity.

We next sought to evaluate whether a limited set of genes could retain the predictive power to discriminate high versus low enrichment for each of the on-treatment signatures. To identify the most important genes that drive enrichment across the gene sets, we performed 10-fold nested cross-validation with elastic net regression to identify patients with high- versus low- enrichment across each of the gene sets in the TCGA-SKCM dataset (Methods). This cross-validation approach demonstrated that (i) a small number of genes can accurately predict results from the larger gene sets (Data S1) and (ii) that top genes from the rela+nivo group contained CD38 and HAVCR2 (Data S1).

Finally, we wanted to validate that CD38 and HAVCR2 are reflective of the larger on-therapy rela+nivo signature. To achieve this, we utilized cryopreserved PBMC from our study for flow cytometric analysis of these two markers on CD8+ T cells (Figure 6F). There were no statistically significant differences between baseline and week 4 samples following rela and nivo monotherapy, but as anticipated there was a significant increase in the frequency of CD38+TIM3+ (encoded by HAVCR2) CD8+ T cells following rela+nivo (p=0.049; Figure 6G). Overall, our flow cytometry data are consistent with the transcriptomic analyses, suggesting that co-expression of CD38 and TIM3 on CD8+ T cells in PBMC may serve as an on-therapy biomarker of rela+nivo combination therapy.

Discussion

In this study, we leveraged biospecimens from a phase 2 study of metastatic melanoma in which treatment naïve patients received a 4 week lead-in treatment of rela alone, nivo alone, or the combination or rela+nivo. First, we found that rela+nivo led to enhanced responses to T cell receptor signaling. Recently, we showed that LAG3 led to suppression of TCR signaling in a cell-intrinsic manner by promoting co-receptor/Lck dissociation³⁵. However, we did not find enhanced TCR signaling following rela or nivo alone, meaning that blocking both LAG3 and PD1 is required for this phenotype. PD1 ligation inhibits T cell function by modulation of CD28 co-stimulatory signaling^36,37, but may also potentiate TCR signaling by inducing sustained Lck activation³⁸. The downstream pathways need to be further elucidated, but rela+nivo are both required for this elevated ability to respond to TCR stimulation.

Second, blockade of both LAG3 and PD1 modulated CD8+ T cell differentiation, leading to co-expression of cytotoxic and exhaustion gene modules. Some reports have found that activation and exhaustion modules are intertwined in exhausted CD8+ T cells^39-41, while another study reported cytotoxicity and exhaustion modules to be mutually exclusively expressed in intratumoral CD8+ T cells⁴². Rela+nivo treatment leads to a unique state in which cytotoxicity is enhanced despite transcriptional hallmarks of exhaustion. This phenotype was especially pronounced in emergent clones of CD8+ T cells in the tumor microenvironment. In contrast to previously described emergent clones that were enriched for exhaustion signatures following monotherapy targeting the PD1/PDL1 axis²¹, our study revealed enhanced functional signatures in these clonally expanded cells. Combination therapy likely promotes emergence of these clones due to either a reduction in the TCR activation threshold, rescue of TCR signaling, or both. Recent evidence modeling clonal differentiation of CD8+ T cell exhaustion in mice and humans has revealed two distinct populations: a terminally differentiated population⁴³ and a killer cell lectin-like receptor (KLR) expressing population⁴⁴. A key determinant of the progression towards terminal differentiation was the avidity of the TCR for antigen⁴⁴, with higher avidity promoting terminal exhaustion. In our study, TCR signaling was enhanced in CD8+ T cells following rela+nivo, which suggests emergence of a functionally resilient terminally differentiated state.

We found that distinct sets of TFs controlled the differentially expressed genes associated with each treatment group. Following rela+nivo treatment, specific TFs governed genes involved in both cytotoxicity and exhaustion. PRDM1 is a central regulator of CD8+ T cell effector function and terminal versus memory T cell differentiation^45-47, indicating that PRDM1 expression promotes skewing towards effector responses in our study. ID2 controls CD8+ T cell survival, leading to the accumulation of CD8+ T cells and effective immunity⁴⁸. A role for ETV7 has recently been described in suppressing type I interferon responses during viral infection⁴⁹. ZBED2 has been similarly implicated in suppressing IRF1-mediated interferon responses⁵⁰, although in our study it appears to also drive genes associated with cytotoxicity and exhaustion. BATF promotes the enhancement of effector functions and reduces CD8+ T cell exhaustion in chimeric antigen receptor T cells via interaction with IRF4⁵¹, and may play a similar role following rela+nivo. Importantly, our data demonstrate that CD8+ T cell cytotoxic programs can be promoted in TOX-expressing CD8+ T cells. Recent studies have suggested that TOX-mediated programs may play a greater role in promoting CD8+ T cell longevity and survival over exhaustion^52,53. Collectively, our study demonstrates that a suite of TFs is responsible for the co-expression of cytotoxicity and exhaustion following rela+nivo.

Interestingly, it appears that CD8+ T cells do not significantly upregulate multiple cytotoxic effector molecules after nivo or rela monotherapy. Instead, rela treatment leads to interferon responses and expression of CXCL13, which is associated with dysfunctional, antigen-specific CD8+ T cells^42,54. Nivo treatment, conversely, leads to expression of GZMH, which is associated with favorable CD8+ T cell function^42,55. However, IKFZ3 (gene for Aiolos) inhibits IL-2 production by T cells, limiting their autocrine and paracrine signaling potential⁵⁶. Notably, knockout of IKZF3 in CAR T cells promotes killing of solid tumors⁵⁷. Thus, although CD8+ T cells can partially upregulate effector functions following rela or nivo monotherapy, they do not upregulate cytotoxic functions to the degree observed follow rela+nivo.

The unique transcriptional signature derived from CD8+ T cells following treatment with rela+nivo was associated with favorable outcomes across multiple external cohorts. We further found that enrichment of the rela+nivo signature was reflected by a small set of genes, including CD38 and HAVCR2. Our findings that the frequency of CD8+ T cells that co-express CD38 and TIM3 proteins in peripheral blood goes up significantly following rela+nivo treatment suggest that this may be an on-therapy biomarker of rela+nivo treatment as early as 4-weeks after therapy. On-treatment biomarkers would provide a means to identify patients achieving a pharmacodynamic response and help to identify patients likely to be resistant to rela+nivo therapy. Such biomarkers (CD38+TIM3+CD8+ T cells in PBMC) should be rigorously evaluated in future studies.

We have shown here that rela+nivo leads to distinct and unique cell intrinsic changes in the CD8+ T cell compartment, and that these changes were not simply the sum of the changes observed following rela or nivo alone. Our results agree with analysis in chronic viral infection models, where PD1 and LAG3 contribute to distinct aspects of CD8+ T cell exhaustion (Ngiow et al co-submitted manuscript). Our results are also mechanistically supported by observations in murine tumor models, where autocrine IFN-γ signaling from CD8+ T cells contributes to superior tumor control (Andrews and Butler et al co-submitted manuscript). Collectively, these findings demonstrate that combinatorial disruption of LAG3 and PD1 supports superior antitumor immunity driven by transcriptional and functional changes that are distinct from the impact of disrupting either LAG3 or PD1 alone.

Many combination immunotherapies are currently being evaluated for the treatment of solid tumors⁵⁸. Prioritization of combinations that are safer and more effective than monotherapies is urgently needed. Our study provides insights into the synergy underlying combination blockade with rela and nivo and establishes a paradigm wherein CD8+ T cell cytotoxicity can be promoted in cells that otherwise appear exhausted. We also show that rela is not ‘inert’ despite its apparent lack of monotherapy efficacy in murine models, raising the possibility that other combinations that include LAG3 blocking therapies might be efficacious, even without the inclusion of PD1/PDL1 blockers. These findings will aid in the selection of additional combination strategies for the treatment of cancer and chronic viral infections.

Limitations of the study

The cohort of patients utilized for this trial is small, which limits the power. Outcome data cannot be fully reported from the ongoing trial from which these biospecimens are derived until the trial has formally concluded. Although the inclusion of a rela only arm was a critical component of this trial, the crossover design limits our ability to draw outcome inferences related to lead-in therapies. The peak timing of maximum immunological response to rela+nivo is unknown and may not be week 4 post treatment initiation, and thus larger effects are possible at either earlier or later timepoints. Inference of genes governed by transcription factors should be validated in future studies using either CRISPR or CHIP-seq mediated approaches. Other cell-intrinsic or cell-extrinsic changes in peripheral and intratumoral immune and non-immune populations beyond IFN-γ response should be evaluated in future studies. Future studies should also rigorously evaluate biomarkers of immunological and clinical response to rela+nivo as well as identifying baseline signatures that would predict response to either nivo monotherapy or combination therapy with rela+nivo.

STAR Methods

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dario Vignali (dvignali@pitt.edu).

Materials availability

This study did not generate any new reagents.

Data and code availability

The raw and processed single-cell RNAseq data is available on the Gene Expression Omnibus under the accession number GSE225063. The data necessary to recreate the figures is available on Zendo (https://doi.org/10.5281/zenodo.10221413). The code necessary to recreate the figures is available on Github (https://github.com/CilloLaboratory/rela_nivo_melanoma) with the following DOI: https://zenodo.org/10.5281/zenodo.10223848. Any other data required to re-analyze the data reported in this manuscript is available from the lead contact upon request.

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Study design and rationale

We designed an open-label randomized phase 2 trial to study the mechanisms underlying the efficacy of nivolumab in combination with relatlimab in patients with advanced melanoma (NCT03743766). In this trial, patients aged 18 years or older with advanced (inoperable or metastatic) melanoma prior to immunotherapy were enrolled and stratified 1:1:1 to receive two cycles of either anti-PD1 alone, anti-LAG3 alone or anti-PD1 plus anti-LAG3 in combination. After the initial two cycles, all patients subsequently received combination therapy with anti-PD1 and anti-LAG3. Exclusion criteria included: uncontrolled central nervous system metastases; active autoimmune disease requiring immunosuppressive treatment except for type 1 diabetes, vitiligo, resolved asthma, controlled hyper- or hypo-thyroidism, and hypoadrenalism or hypopituitarism; prior systemic treatment with immunomodulatory agents or targeted therapies. For anti-PD1 treatment, patients received 480 mg nivolumab intravenously. For anti-LAG3 treatment, patients received 160 mg relatlimab intravenously. For anti-PD1+anti-LAG3, patients received sequential intravenous infusion with 160 mg relatilmab followed by 480 mg nivolumab. Study drug was provided by Bristol-Myers Squibb. We rationalized that this unique study design would enable collection of blood and tumor biospecimens to enable mechanistic dissection of the roles of PD1 and LAG3 blockade alone and in combination in the setting of advanced melanoma.

After the trial is complete, the reporting strategy will consist of the pre-specified formal statistical comparisons between treatment arms. Specifically, the primary outcome measures are the following:

• Change in LAG3 expression at baseline and 4 weeks;

• Change in PD1 expression at baseline and 4 weeks;

• Change in tumor size RECIST v1.1 where tumor lesions must be accurately measures in at least one dimension with a minimum size of 10 mm by CT scan, 10 mm caliper measurement by clinical examination, or 20 mm by chest X-ray;

• Overall response rate as quantified by the number of patients experience either complete response or partial response by RECIST v1.1, the number of patients with stable disease, and the number of patients with progressive disease. Complete response is defined as the disappearance of all target lesions and reduction of pathological lymph node size to <10 mm on the short axis. Partial response is defined as at least a 30% decrease in the sum of the diameters of a target lesion versus baseline diameters. Progressive disease is defined as at least a 20% increase in the sum of diameters of target lesions using the smallest sum of diameters at baseline as reference. Stable disease is defined as a decrease less than that required for partial response and an increase less than that required for progressive disease.

Secondary outcome measures are the following:

• Clinical benefit rate defined as the sum of the number of patients experience either complete response, partial response or stable disease divided by the total number of patients assessed by RECIST v1.1;

• Duration of response defined as the time from complete or partial response to disease progression;

• Progression-free Survival defined as the date from randomization to the first date of disease progression or death due to any cause;

• Overall survival up to four years, defined as the time between date of randomization and date of death due to any cause;

• LAG3 expression at week 16;

• PD1 expression at week 16;

• Change in CD4+ tumor infiltrating lymphocytes;

• Change in CD8+ tumor infiltrating lymphocytes;

• Change in serum granzyme B levels;

• Change in cell effector/memory status;

• Change in activation and maturation of dendritic cells;

• Change in T cell count from baseline to 4 weeks and 12 weeks post-treatment;

• T cell count at time of disease progression up to 4 years;

• Change in soluble LAG3 levels at baseline, 4 weeks and 12 weeks post-treatment;

• Soluble LAG3 levels at time of disease progression up to 4 years;

• Change in Regulatory T cell levels at baseline, 4 weeks and 12 weeks post-treatment;

• Regulatory T cell levels at time of disease progression up to 4 years.

Patients and specimens

A CONSORT diagram describing the study is provided as Figure S1. Patient demographic information and additional details are provided in Table S1 and summarized in Data S1. Peripheral blood and tumor biopsies were collected prior to the initiation of therapy, at 4 weeks following the initial treatment stratified into one of the three treatment arms, and at 16 weeks of therapy. Follow-up assessment was made at subsequent intervals of 16 weeks or upon clinical evidence of tumor progression. Peripheral blood was obtained at subsequent follow-up timepoints, but tumor biopsies were only obtained when feasible. Specimens obtained for analysis are tabulated in Table S1 and Data S1.

METHOD DETAILS

Sample processing

Whole blood was collected in EDTA tubes and peripheral blood mononuclear cells were isolated by Ficoll-Hypaque density gradient centrifugation. Whole blood was diluted with Hanks Balanced Salt Solution (HBSS) and 30 mL of diluted blood was layered onto 15 mL of Ficoll. Blood samples were then centrifuged at 400xg for 20 minutes with the break set to off. The PBMC layer was then removed and washed with HBSS, followed by centrifugation for 5 minutes at 400xg. Red blood cells were then lysed using BD PharmLyse for 5 minutes at room temperature, and PBMC were again centrifuged at 400xg for 5 minutes. Tumor biopsies were collected into RPMI supplemented with 10% fetal bovine serum. Biopsies were mechanically disrupted and passed over a 100 μm filter and then spun down at 400xg for 5 minutes. Red blood cell lysis was performed as described above for PBMC.

Generation of single-cell RNAseq data

Following isolation of PBMC and generation of single-cell suspensions from tumor biopsies, we next sorted all live immune (CD45+) cells from PBMC and tumors using either a MoFlo Astrios or a Sony MA900 at the Hillman Cancer Center Flow Cytometry Core. Cells were stained with anti-CD45::PE (clone: 2D1; Biolegend) in PBS with 10% FBS a 4C. For samples that were multiplexed using TotalSeq antibodies (Biolegend), staining with TotalSeq antibodies was performed concomitantly to CD45 staining. Following 15 minutes at 4C, cells were washed with PBS + 10% FBS followed by centrifugation at 400xg for 5 minutes. Next, viability staining was performed using eFluor 780 viability dye (eBioscience) diluted 1:4000 in PBS at 4C for 15 minutes, followed by a wash with PBS and centrifugation as described above. Sorted cells were then resuspended in 0.04% bovine serum albumin in PBS and were counted using a Cellometer Auto2000 (Nexcelom). Cells were then loaded into the 10X Controller for 5’-based single-cell RNAseq library generation. Both NextGEM and 5’-v1 kits were used (10X Genomics). Gene expression, TotalSeq, and TCR libraries were generated using the manufactures’ protocol (10X Genomics). Final libraries were pooled and sequenced on a NextSeq550 at the University of Pittsburgh Genomics Core or a NovaSeq6000 at the University of Pittsburgh Medical Center Genomics Core.

Processing of single-cell RNAseq data

After sequencing, raw sequencing files were download to the University of Pittsburgh Center for Research Computing High Performance Computing Cluster. Files were demultiplexed using bcl2fastq (v2.20.0) and were aligned to the reference genome GRCh38 using CellRanger (v6.0.1). Single-cell TCRs were reconstructed using CellRanger vdj. For samples that were multiplexed by TotalSeq, CiteSeqCount (v1.4.3) was used to identify uniquely labeled samples. Following alignment, feature/barcode matrices were read into Seurat (v4.0.5)⁵⁹ and a unified analysis object was created in R (v4.0.0). Samples that were multiplexed with TotalSeq were demultiplexed using an automated approach wherein positive and negative cutpoints for each Totalseq marker were identified in each multiplexed sample using k-means clustering. Cells that passed both positive and negative thresholds were attributed to each individual sample. Quality control was performed by evaluating the number of genes and the percent of mitochondrial genes per cell. Cells with less than 200 genes or with >10% reads aligning to mitochondrial genes were excluded. Genes were excluded that were expressed in fewer than 3 cells.

Dimensionality reduction and clustering

Single-cell variation inference (scVI) from scvi-tools was performed to normalized data across individual samples and to perform dimensionality reduction^60,61. scRNAseq data was processed using scanpy (v1.8.2) and anndata (v0.7.8)⁶². The top 2000 highly variable genes were first identified, and a variational autoencoder from scVI (v0.14.5)⁶¹ was modeled to reduce the dimensionality to a 10 dimensional latent space using 128 nodes per hidden layer and 1 layer with a dropout rate of 10% using individual samples as a potential confounding variable. This resulting 10-dimensional latent space was then used for generation of UMAPs and clustering using the Leiden algorithm. Scvi-tools and scanpy were implemented in Python (v3.8). Variational autoencoders were fit to identify the overall cell types, T cell subsets, and CD8+ T cells using the approach described above.

Identification of cell types

To identify cell types, we related Leiden-derived clusters to expression of canonical genes as previously described^63,64. Briefly, non-immune cells were filtered out based on lack of PTPRC (gene for CD45) expression or the presence of expression of PMEL (melanoma cells) or HBB (red blood cells). Immune cell subsets were then identified by first identifying the major immune lineages (i.e. T cells, B cells, plasma cells, myeloid cells, neutrophils, etc). To achieve these, we evaluated expression of CD3D, CD14, FCGR3A, MS4A1, FUT4, and CD1C across clusters. After identifying the major immune lineages, we next performed in silico isolation of T cells and re-clustered these cells to identify CD8+ T cells, conventional CD4+ T cells, and regulatory CD4+ T cells using expression of CD4, CD8A, and FOXP3 across clusters. CD8+ T cells were then isolated in silico for downstream analysis.

Analysis of CD8+ TCRs and clonal replacement

We used the TCRB CDR3 counts from individual CD8+ T cells to assess shared and emergent clones. We selected TCRs present with at least 3 counts to be clonally expanded. To evaluate clonal replacement following immunotherapy, we first identified all clones that were present at baseline in individual patients. Next, we identified clones that were uniquely present post-therapy and determined the percentage of each emergent clone across clusters of CD8+ T cells to infer the transcriptional state of the emergent TCR clones.

Gene set curation and gene set enrichment analysis

Gene set enrichment analysis was performed to characterized subsets of CD8+ T cells in PBMC and derived from tumors. Gene sets were curated from the Molecular Signatures Database⁶⁵ using the msigdbr package⁶⁶ and a murine study of LCMV infection that identified precursor and terminally exhausted CD8+ T cell subsets⁴³. The gene sets from LCMV were further refined by identifying genes that were restricted to each of the unique progenitor or terminally exhausted cell states. The AddModuleScore function from the R package Seurat was used to generate enrichment scores for individual cells.

T cell receptor signaling assay for CD8+ T cells

To assess whether TCR signaling is different across the treatment groups, we performed flow cytometry to evaluate early (pSLP76 and pZAP70) and late (pAKT and pS6) phosphorylation events downstream of TCR cross-linking and co-stimulatory signaling (pPI3K). Briefly, cryopreserved PBMC were thawed and rested overnight at one million cells per milliliter of complete RPMI supplementary with 10% FBS at 37°C and 5% CO₂. The following morning, cells were counted and resuspended into 96 well plates at 100,000 to 200,000 cells per well. Cells were then washed with PBS supplemented with 10% FBS followed by live/dead staining with Zombie Near-IR dye. Next, cells were surfaced stained with all antibodies including anti-CD3 and anti-CD28 conjugated to biotin and were keep on ice. For week 4 samples, functional anti-LAG3 and/or anti-PD1 antibodies were added (based on which treatment the patient received) to ensure maximum receptor occupancy. Then, 37°C pre-warmed complete RPMI with 10% FBS and streptavidin was added to all wells and cells were cultured at 37°C for 15 minutes. Paraformaldehyde fixative was added directly to the wells to stop the reaction and samples were incubated at room temperature for 15 minutes. Cells were then washed twice with Perm was buffer (Biolegend), followed by a 15-minute incubation in PBS + 10% FBS at 4°C. Finally, intracellular antibodies were added for 30 minutes are room temperature and cells were washed and resuspended in PBS + 10% FBS for flow cytometry analyses on the Cytek Aurora.

Flow cytometry staining and analysis

Flow cytometry staining was performed on cryopreserved PBMCs as described above for TCR signaling, with the following exceptions: cells were stained immediately following thawing rather than resting overnight followed by staining; and cells were only washed once with Perm Wash. Flow cytometry was performed using a Cytek Aurora.

Pseudotemporal analysis of differentiation

RNA velocity was performed to infer differentiation across CD8+ T cells, specifically using samples where both pre- and post-treatment samples from PBMC and TIL were available to reduce the probability of inferring trajectories from temporally unpaired samples. Velocyto (v0.17)²⁰ was used to align and quantify unspliced transcriptions from gene expression FASTQ files. scVelo (v0.2.4)⁶⁷ was then used to recover splicing dynamics from the top 10,000 highly variable genes with a minimum count of 10 UMIs across all CD8+ T cells. First-order moments were then identified using the scvi reduced space (see Dimensionality reduction and clustering) using the 50 nearest neighbors. Next, RNA velocity was quantified using the deterministic mode with a minimum r² of 1x10⁻⁶. Pseudotemporal ordering was then computed from the velocity graph and used for downstream inference.

Inference of drivers of differentiation

Following identification of an overall pseudotemporal differentiation process, we subset CD8+ T cells between baseline and each of the individual treatment arms. We then fit a random forest regression model for each individual group to identify the genes that were most strongly associated with differentiation in each group. To quantify the importance of individual genes, we employed random forest regression to fit each gene as a function of pseudotime using 10-fold cross validation. We then leveraged the importance scores for each gene within each treatment arm to identify those genes that contributed differently to differentiation across the treatment groups. The random forest was fit using ranger (v0.12.1)⁶⁸ through the R package caret (v6.0-86)⁶⁹.

Identification of distinct CD8+ transcriptional profiles associated with treatment

To identify associations between treatments and gene expression profiles, we used the R package miloR (v1.1.1)²². Milo identifies regions of a graph-based network that are associated with groups of cells derived from different conditions (e.g. treatment groups). Briefly, it accomplishes this by first building a graph-based network and identifying neighborhoods of cells that are transcriptionally similar. It then looks for enrichment of different conditions across the cellular neighborhoods and identifies regions of the graph that are enriched for specific conditions. Using this information about enrichment on the graph, differentially expressed genes associated with neighborhoods that are enriched can be identified. We implemented miloR by first performing dimensionality reduction using principal component analysis on all CD8+ T cells using the top 2000 highly variable gene across all CD8+ T cells. We then built the milo graph using 20 nearest neighbors across the top 10 principal components. Cellular neighborhoods were then identified by repeatedly sampling 10% of the vertices across the network, once again using 20 nearest neighbors across the top 10 principal components with refinement of the vertex sampling. We then identified cellular neighborhoods that differed across the treatment groups by calculating the Euclidean distance between neighborhoods using the top 10 principal components and then using a design matrix including individual replicate samples across treatment conditions (i.e. patients) as a covariate. We then identified neighborhood associated with each treatment group defined by replicate conditions and evaluate differentially expressed genes associated with these neighborhoods.

Defining transcription factor-driven gene regulatory networks

We leveraged a fast python-based implementation of single-cell regulatory network inference and clustering (pySCENIC; v1.1.2-2)^23,24 to infer transcription factor-driven expression of genes within CD8+ T cells. To do this, we first filtered out genes that were expressed with fewer counts than 3 in 1% of CD8+ T cells and required the gene to be expressed in at least 1% of CD8+ T cells. Next, we leveraged a database containing genes that are putative regulated by transcription factors in conjunction with GRNBoost2 to generate co-expression modules between transcription factors and the genes they potentially regulate. To ascertain which transcription factors regulate each other, we next evaluated the resultant adjacency matrix from GRNBoost2 for transcription factors. To relate transcription factor driven gene module expression to treatment conditions, we evaluated the ability of transcription factors to drive expression of the genes that were identified as upregulated in specific treatment conditions from miloR analyses. We filtered genes putative governed by transcription factors to the top 20 genes by importance from the adjacency matrix (i.e., the differentially expressed genes from a given treatment group had to be in the top 20 genes governed by the transcription factor to be considered driven by that transcription factor). Networks were visualized using the ggraph (v2.0.5)⁷⁰ and tidygraph (v1.2.0)⁷¹ packages.

Evaluation of treatment signatures in external cohorts

We utilized several external cohorts of bulk RNAseq data to determine if signatures derived from our datasets were associated with reflective of CD8+ T cell functionality, using either response to immunotherapy or overall survival to infer superior CD8+ T cell function based on gene sets derived from our study. The first cohort was skin cutaneous melanoma patients from The Cancer Genome Atlas. Bulk RNAseq count data aligned using STAR was download in R via TCGAbiolinks (v2.14.1)⁷². Counts were normalized using a variance stabilizing transformation implemented in DESeq2 (v1.26.0)⁷³. GSVA (v1.36.3)⁷⁴ was used to score each patient for gene signatures derived from the top 100 differentially expressed genes from anti-PD1, anti-LAG3 or anti-PD1+anti-LAG3 from miloR analysis. To assess whether our treatment-derived CD8+ T cell signatures were related to response to immunotherapy and overall survival following immunotherapy (in contrast to the TCGA patients), we next utilized 4 datasets that had bulk RNAseq data from prior to checkpoint blockade, response data, and overall survival data from melanoma patients^30-34. These data were likewise normalized using DESeq2 as described above. Elastic net regression as implemented in R package caret (v6.0-86)⁶⁹ was used to reduce the number of genes from 100 to 10. Briefly, we used 10-fold nested cross-validation from the TCGA dataset to identify combinations of genes that were able to predict whether a patient had high or low enrichment for a given gene set.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical assessments were performed across replicate samples with two-sided type I errors of less than 5% being considered statistically significant. For linear mixed effects models, we utilized the R package lme4 (v1.1-27.1)⁷⁵ in conjunction with the multcomp (v1.4-17)⁷⁵ and afex (v1.0-1)⁷⁶ packages. Models were specified using a random intercept that corresponded to individual patients and a fixed effect of treatment group (i.e., baseline, anti-PD1, anti-LAG3 oranti-PD1+anti-LAG3). Linear mixed effects models were only considered for pairwise comparisons is the omnibus test pass a false discovery rate-controlled type I error of less than 5%. P values derived from pairwise comparisons were conservatively corrected using the Holm correction. Simple linear models (i.e. those without fixed effects) were treated in the same manner. Correlations were assessed using Spearman’s correlation. Differentially expressed genes were identified using a Wilcoxon rank sum test comparing cells within a given cluster to all other cells as implemented in Seurat. Genes were considered significantly differentially expressed if the false discovery rate-controlled type I error was less than 5%. Likewise, gene sets were similarly evaluated for statistical significance using singleseqgset. Survival analyses were performed using the survival (v3.2-11)^77,78, survminer (vO.4.9)⁷⁹, survMisc (vO.5.5)⁸⁰ packages. Multiple variable survival analyses were performed to identify variables that were independently associated with overall survival across cohorts.

Supplementary Material

1

Data S1/Methods S1. Additional supporting information regarding the trial, clinical characteristics of study participants, and supplementary analyses, related to STAR Methods.

2

Figure S1. CONSORT diagram outlining the construction of the patient cohort, allocation to lead-in treatment arms, discontinuation of treatment and reasons for discontinuation, and patients including for clinical analyses, related to Figure 1.

A total of 61 patients were assessed for eligibility and 43 were allocated to lead-in treatment arms and received interventions (15 to nivo monotherapy, 14 each to rela monotherapy and rela+nivo therapy, respectively). Reasons for discontinuation and final clinical cohort included are shown.

3

Figure S2. Representative flow gating schema for pSLP76, related to Figure 2.

The approach here was utilized to assess TCR signaling in CD8+ T cells from PBMC at baseline and following 4 weeks of lead-in treatment. A) Cells were as live lymphocytes that expressed effector CD8 (i.e. CCR7-CD45RA-CD8+ T cells). B) pSLP76 levels in unstimulated control effector CD8+ T cells from a healthy donor versus CD8+ T cells crosslinked for 15 minutes with anti-CD3 and anti-CD28. C) Represtaive pSLP76 levels in unstimulated and cross-linked cells from patients at baseline and following 4 weeks of lead-in therapy. Analysis was performed with FlowJo software.

4

Figure S3. Temporal dynamics for the gene modules associated with differentiation of CD8+ T cells, related to Figure 3.

Different dynamics are observed across 10 gene modules, with some linear increase with pseudotime, some decreasing with pseudotime, and some demonstrating more complex dynamics. The dynamics for the 8 gene modules not shown in Figure 3 are shown here.

5

Figure S4. Assessment of TCR overlap across clusters from each treatment group, and differentially expressed genes between shared and emergent TCR clones for each week 4 lead-in group, related to Figure 4.

Differential TCR overlap was observed across baseline and each treatment group, and distinct patterns of differentially expressed genes were found in shared versus emergent TCR clones for each treatment group. A) TCR overlap across clusters from CD8+ T cells at baseline. Extensive overlap is observed between clusters 7, 8 and 9 and clusters 4 and 5. B) TCR overlap across clusters from CD8+ T cells following rela treatment at 4 weeks was like that observed at baseline. C) TCR overlap across clusters from CD8+ T cells following nivo treatment at 4 weeks revealed overlap between cluster 5 and clusters 4 and 7. D) TCR overlap across clusters from CD8+ T cells following rela+nivo treatment at 4 weeks showed significant overlap between clusters 7 and 9, and to a lesser degree clusters 4 and 5. E) Differentially expressed genes between shared and emergent clones following rela treatment. F) Differentially expressed genes between shared and emergent clones following nivo treatment. G) Differentially expressed genes between shared and emergent clones following rela+nivo treatment.

6

Figure S5. Assessment of IFN-γ signaling across immune cell subsets in PBMC and TIL, related to Figure 4.

Gene set scores were calculated for interferon gamma response across all immune cell subsets in PBMC (left panel) and TIL (right panel) and were compared to baseline using linear mixed effects models. In PBMC, rela+nivo treatment led to elevated interferon gamma response in numerous cell subsets. In TIL, a most restricted subset of cells had elevated interferon response compared to baseline including CD14-CD16+ monocytes and conventional CD4+ T cells.

7

Figure S6. Comprehensive network plots of transcription factors at the genes they putatively drive based on inference from SCENIC, related to Figure 6.

Transcription factors and the genes that they putative drive from amongst the differentially expressed genes derived from miloR. The top 10 transcription factors from each group and the top 10 genes they drive are shown.

8

Table S1. Clinical data and samples for genomic analyses, related to Figure 1.

Patient-level clinical and demographic data and samples that were utilized for genomic analyses.

9

Table S2. CD8+ T cell cluster and treatment group differentially expressed genes and gene sets, related to Figure 2.

Differentially expressed genes and gene sets across CD8+ T cell clusters and associations between gene sets and treatment groups.

10

Table S3. Differentially expressed genes between shared and emergent CD8+ T cell clones, related to Figure 4.

Genes that are differentially expressed between shared and emergent clones at 4 weeks post-treatment.

11

Table S4. Differentially expressed genes across treatment groups from miloR analysis, related to Figure 5.

Genes that are differentially expressed between unique populations of cells associated with each treatment group at 4 weeks post-treatment as defined by miloR analysis.

12

Table S5. Inferred targets of transcription factors from pySCENIC analysis, related to Figure 6.

Transcription factors and the genes that they are inferred to drive as inferred from pySCENIC.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Fixable viability dye eFluor780	ThermoFisher	Catalog number: 65-0863-14
Fixable viability dye Zombie NIR	ThermoFisher	Catalog number: 423106
Monoclonal mouse anti-human CD45, clone 2D1, PE conjugated	Biolegend	Catalog number: 368510; RRID: AB 2566370
Monoclonal mouse anti-human CD14, clone 63D3, BUV395 conjugated	BD	Catalog number: 567964; RRID: AB_3096982
Monoclonal mouse anti-human CD20, clone L27, BUV395 conjugated	BD	Catalog number: 740204; RRID: AB 2739954
Monoclonal mouse anti-human CD56, clone NCAM16.2, BUV395 conjugated	BD	Catalog number: 563555; RRID: AB 3096986
Monoclonal mouse anti-human CD8, clone RPA-T8, BUV496 conjugated	BD	Catalog number: 612942; RRID: AB 2870223
Monoclonal mouse anti-human CD45RA, clone HI100, BUV615 conjugated	BD	Catalog number: 751555; RRID: AB_2875550
Monoclonal mouse anti-human CXCR6, clone 13B 1E5, BUV805 conjugated	BD	Catalog number: 748448; RRID AB 2872864
Monoclonal mouse anti-human pS6 (Ser235/Ser236), clone A17020B, BV421 conjugated	Biolegend	Catalog number: 608610; RRID: AB 2814451
Monoclonal mouse anti-human CD38, clone HB-7, BV421 conjugated	Biolegend	Catalog number: 356618; RRID: AB 2566230
Monoclonal mouse anti-human Granzyme B, clone GB11, Pacific Blue conjugated	Biolegend	Catalog number: 515408; RRID: AB 2562195
Monoclonal mouse anti-human HLA-DR, clone L243, BV510 conjugated	Biolegend	Catalog number: 307646; RRID: AB_2561948
Monoclonal mouse anti-human PD1, clone EH12.2H7, BV650 conjugated	Biolegend	Catalog number: 329950; RRID: AB 2566362
Monoclonal mouse anti-human LAG3, clone 11C3C65, BV785 conjugated	Biolegend	Catalog number: 369322; RRID: AB 2716127
Monoclonal mouse anti-human CD39, clone A1, FITC conjugated	Biolegend	Catalog number: 328206; RRID: AB 940423
Monoclonal mouse anti-human CCR7, clone G043H7, PerCP conjugated	Biolegend	Catalog number: 353241; RRID: AB 2564545
Monoclonal mouse anti-human TIM3, clone F38-2E2, PerCP-Cy5.5 conjugated	Biolegend	Catalog number: 345016; RRID: AB 2561933
Monoclonal rabbit anti-human pPI3K (Tyr458/Tyr199), clone PI3KY458-1A11, PE conjugated	ThermoFisher	Catalog number: MA5-36954; RRID: AB_2896889
Monoclonal mouse anti-human BATF, clone 9B5A13, PE conjugated	Biolegend	Catalog number: 654804; RRID: AB_2563517
Monoclonal mouse anti-human pAKT (Ser473), clone SDRNR, PE-Cy7 conjugated	Biolegend	Catalog number: 25-9715-42; RRID: AB 2688172
Monoclonal mouse anti-human Perforin, clone dG9, PE- Cy7 conjugated	Biolegend	Catalog number: 308126; RRID: AB 2572049
Monoclonal mouse anti-human pSLP76 (Tyr128), clone HNDZ55, APC conjugated	ThermoFisher	Catalog number: 17-9037-41; RRID: AB 2573283
Monoclonal mouse anti-human interferon-γ, clone 4S.B3, APC conjugated	Biolegend	Catalog number: 502512; RRID: AB_315237
Biological samples
Blood and tumor biopsies from melanoma patients	UPMC Hillman Cancer Center, Pittsburgh, PA USA	https://clinicaltrials.gov/study/NCT03743766
Chemicals, peptides, and recombinant proteins
Foxp3 / Transcription Factor Staining Buffer Set	ThermoFisher	Catalog number: 00-5523-00
Critical commercial assays
Chromium Single Cell V(D)J Reagent Kits (v1.1 Chemistry)	10X Genomics	Catalog number: PN-1000165
Chromium Next GEM Single Cell 5’ Reagent Kits v2 (Dual Index)	10X Genomics	Catalog number: PA-1000263
Deposited data
Processed single-cell RNAseq data is available on the Gene Expression Omnibus. Raw sequencing data is available from the Sequence Read Archive.	This study	Gene Expression Omnibus: GSE225063; Sequence Read Archive: PRJNA933637
Single-cell characterization of anti-LAG3 and anti-PD-1 combination treatment in patients with melanoma	Huuhtanen et al23	Zenodo: 10.5281/zenodo.5747250
Genomic and Transcriptomic Features of Response to Anti-PD-1 Therapy in Metastatic Melanoma	Hugo et al30	Gene Expression Omnibus: GSE78220
Tumor and Microenvironment Evolution during Immunotherapy with Nivolumab	Riaz et al31	Gene Expression Omnibus: GSE91061
Distinct Immune Cell Populations Define Response to Anti-PD1 Monotherapy and Anti-PD-1/Anti-CTLA-4 Combined Therapy	Gide et al32	European Nucleotide Archive: PRJEB23709
Integrative molecular and clinical modeling of clinical outcomes to PD1 blockade in patients with metastatic melanoma	Liu et al33	Database of Genotypes and Phenotypes: phs000452.v3.p1
The Cancer Genome Atlas	Liu et al28	National Cancer Institute Genomic Data Commons Data Portal
Oligonucleotides
Software and algorithms
Bcl2fastq (v2.20.0)	Illumina	https://emea.support.illumina.com/downloads/bcl2fastq-conversion-software-v2-20.html
CellRanger (v6.0.1)	10X Genomics	https://www.10xgenomics.com/support/software/cell-ranger/latest
CiteSeqCount (v1.4.3)	Zenodo	https://doi.org/10.5281/zenodo.2590196
R (v4.0.0) Arbor Day	The R Foundation for Statistical Computing	https://www.r-project.org/
Seurat (v4.0.5)	Hao et al59	https://satijalab.org/seurat/
Single-cell variational inference (scVI)	Gayoso et al60; Lopez et al61	https://scvi-tools.org/
Scanpy	Wolf et al62	https://scanpy.readthedocs.io/en/stable/
Anndata	Wolf et al62	https://anndata.readthedocs.io/en/latest/
msigdbr	Igor Dolgalev66	https://CRAN.R-project.org/package=msigdbr
Velocyto (v0.17)	La Manno et al20	https://velocyto.org/
scVelo (0.2.4)	Bergen et al67	https://scvelo.readthedocs.io/en/stable/
ranger (v0.12.1)	Wright et al68	https://cran.r-project.org/web/packages/ranger/index.html
Caret (v6.0-86)	Kuhn et al69	https://cran.r-project.org/web/packages/caret/index.html
miloR (v1.1.1)	Dann et al22	https://www.bioconductor.org/packages/release/bioc/html/miloR.html
pySCENIC (v1.1.2-2)	Aibar et al24; Van de Sande et al23	https://pyscenic.readthedocs.io/en/latest/
ggraph (v2.0.5)	Thomas Lin Pedersen70	https://cran.r-project.org/web/packages/ggraph/index.html
Tidygraph (v1.2.0)	Thomas Lin Pedersen71	https://tidygraph.data-imaginist.com/
TCGAbiolinks (v2.14.1)	Colaprico et al72	https://bioconductor.org/packages/release/bioc/html/TCGAbiolinks.html
DESeq2 (v1.26.0)	Love et al73	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
GSVA (v1.36.3)	Hanzelmann et al74	https://www.bioconductor.org/packages/release/bioc/html/GSVA.html
lme4 (v1.1-27.1)	Hothorn et al75	https://cran.r-project.org/web/packages/lme4/index.html
multcomp (v1.4-17)	Hothorn et al75	https://cran.r-project.org/web/packages/multcomp/index.html
afex (v1.0-1)	Singmann et al76	https://cran.r-project.org/web/packages/afex/index.html
survival (v3.2-11)	Therneau et al77; Therneau et al78	https://cran.r-project.org/web/packages/survival/index.html
survminer (v0.4.9)	Kassambara et al79	https://cran.r-project.org/web/packages/survminer/index.html
survMisc (v0.5.5)	Dardis et al80	https://cran.r-project.org/web/packages/survMisc/index.html
Other

Highlights.

• Rela+nivo enhances response to T cell receptor & IFN-γ signaling in CD8+ T cells

• Cytotoxic/exhaustion gene modules are co-expressed following rela+nivo

• PRDM1, BATF, ETV7 & TOXdrive rela+nivo cytotoxicity/exhaustion profile

• Cytotoxicity/exhaustion profile is associated with clinical benefit