An activation to memory differentiation trajectory of tumor infiltrating lymphocytes informs metastatic melanoma outcomes

Abstract

There is a need for better classification and understanding of tumor infiltrating lymphocytes (TILs). Here we applied advanced functional genomics to interrogate 9000 human tumors, and multiple single-cell sequencing sets using benchmarked T cell states, comprehensive T cell differentiation trajectories, human and mouse vaccine responses, and other human TILs. Compared to other T cell states, enrichment of T memory/resident memory programs was observed across solid tumors. Trajectory analysis of single-cell melanoma CD8⁺ TILs also identified a high fraction of memory/resident memory-scoring TILs in anti-PD-1 responders, which expanded post therapy. In contrast, TILs scoring highly for early T cell activation, but not exhaustion, associated with non-response. Late/persistent, but not early activation signatures, prognosticate melanoma survival, and co-express with dendritic cell and IFN-γ response programs. These data identify an activation-like state associated to poor response and suggest successful memory conversion, above resuscitation of exhaustion, is an under-appreciated aspect of successful anti-tumoral immunity.

ETOC BLURB

Jaiswal et al. highlight the need for improved tumor infiltrating lymphocyte (TIL) classification by showing current transcriptome assignments may misclassify early activated/effector TILs as exhausted. The study surveys 9000 solid tumors, multiple single-cell RNA sequencing sets, mouse and human models, and scoring methods to reclassify TILs and associate melanoma survival to T cell memory/resident memory.

Graphical Abstract

INTRODUCTION:

T cells are critical to tumor immune surveillance, editing, and eradication. CD8⁺ tumor-infiltrating lymphocytes (TILs) are key targets of immune checkpoint blockade (ICB) and used as adoptive cell therapy. CD8⁺ TILs are associated with improved survival in multiple cancer types (Naito et al., 1998; Schumacher et al., 2001; Gooden et al., 2011). A more accurate understanding of how human T cells differentiate in the tumor microenvironment (TME) could better inform strategies to sensitize non-responsive patients to ICB, by intervening upon TILs formation and differentiation. Despite efforts to classify TILs into T cell archetypes, it remains unclear which state assignments faithfully capture T cell heterogeneity and behavior in the TME. This is particularly challenging when a handful of markers that are used to classify TILs are shared by multiple T cell states. A broad (comprehensive) and deep (high-resolution) approach would better distinguish TILs status and associate TILs status to patient outcomes.

That tumor reactive TILs display reduced IFN-γ production in response to cognate antigen stimulation and express increased PD-1 (Zippelius et al., 2004; Ahmadzadeh et al., 2009) prompted their classification as “exhausted” T cells. These TILs enrich for an exhaustion gene signature observed in models of chronic lymphocytic choriomeningitis virus (LCMV) antigen-driven T cell dysfunction (Wherry et al., 2007), largely due to shared expression of immune checkpoints (Baitsch et al., 2011). More recent efforts to discern individual TILs status use transcriptome profiling by single cell sequencing (scRNA-seq). High dimensional scRNA-seq data provides a basis for unbiased cell clustering, however the assignment of individual clusters to the exhaustion state is frequently made based on the expression of immune checkpoints (Zheng et al., 2017; Zhang et al., 2018; Li et al., 2019 ). However, immune-checkpoint molecules are also expressed during early T cell activation (Chikuma et al., 2009; Ahn et al., 2018; Avery et al., 2018). PD-1, for example, is expressed in a variety of other T cell states including self-renewing populations in tumors (Siddiqui et al., 2019), stem cells (Liu et al., 2018), and long-lived resident memory T cells (T_RM) (Hombrink et al., 2016; Wang et al., 2019). PD-1 upregulation in the viral setting is not required for the exhausted phenotype. Rather, PD-1 signaling opposes the terminal differentiation of exhausted T cells (Odorizzi et al., 2015). Hence, immune checkpoints lack specificity to the T exhaustion state, and PD-1 may serve to prevent the terminal differentiation of multiple self-renewing populations.

In addition, T cells co-expressing immune checkpoints span a wide functional spectrum from activated to dysfunctional. Some checkpoint-expressing cells retain high clonality, tumor reactivity and proliferative capacity (Zhang et al., 2018; Li et al., 2019). Immune checkpoint expression does not distinguish early T cell exhaustion, which is IL-15 rescuable, from late exhausted cells, whose changes in chromatin accessibility suggest terminal differentiation which cannot be rescued by IL-15 (Schietinger et al., 2016; Philip et al., 2017). Given that the expression of immune checkpoints is not unique to exhausted T cells, and checkpoint-expressing TILs are not uniformly dysfunctional, the characterization of TILs based on immune checkpoints is likely insufficient to capture TILs status or differentiation.

Recent work identified stem-like TCF1⁺ TILs in human cancers and murine cancer models. The current view is that these stem-like cells give rise to terminally differentiated (TD) exhausted/dysfunctional TILs (Jansen et al., 2019; Miller et al., 2019; Siddiqui et al., 2019). This model is based on a comparison of TILs to stem-like exhausted progenitors generated during chronic LCMV whereby the exhaustion progenitor population (CXCR5⁺ TCF1⁺) is capable of giving rise to TIM3⁺ TD exhausted cells (Blackburn et al., 2008; Im et al., 2016). Additional subsets between stem-like and TD exhausted cells were also described, each harboring differences in proliferative potential and effector-like behavior (Hudson et al., 2019; Beltra et al., 2020). Although TCF1⁺ stem-like TILs are thought to give rise to exhausted cells, they are linked to positive outcomes in ICB-treated melanoma patients (Sade-Feldman et al., 2018; Miller et al., 2019). Confounding this, TCF1 is also expressed on memory cells, where it promotes differentiation and longevity (Zhou et al., 2010), which might be favorable in the TME. Given that exhaustion progenitors and memory T cells share TCF1 expression, it is unclear if TIL differentiation occurs along one axis or another.

Different terms are used to describe hyporesponsive T cell states in the context of tumors, including T cell anergy, dysfunction and exhaustion (Blank et al., 2019). Tumor-driven T cell hyporesponsiveness can be induced by sub-optimal priming, which may be due to insufficient co-stimulation (e.g. signal 2 loss, anergy), chronic antigen/TCR stimulation (e.g. exhaustion), immune checkpoint signaling (e.g. PD-L1/PD-L2) and/or other inhibitory signals or cytokine cues within the TME (Schietinger and Greenberg, 2014; Philip and Schietinger, 2021). It is possible that multiple mechanisms operate within the TME, with some T cells receiving chronic antigen cues and differentiating to an early rescuable hyporesponsive state, or a late non-rescuable exhausted/dysfunctional state, while other T cells undergo sub-optimal engagement, which can potentially be intervened upon.

Another complexity in understanding TILs states is that some human TILs from epithelial tumors and melanoma share features of tissue retention seen in T_RM cells, which are long-lived, non-recirculating sentinels that survive in the absence of antigen and mediate tissue barrier defense (Gebhardt et al., 2009; Wakim et al., 2010; Wijeyesinghe et al., 2021). In support of TILs residency status, dominant TIL TCR clonotypes vary between different metastatic melanoma skin sites from the same patient, suggesting a lack of recirculation (Boddupalli et al., 2016). Additionally, TILs express cellsurface markers shared by T_RM, such as CD69, CD49a, and CD103 (Salerno et al., 2014; Webb et al., 2014; Djenidi et al., 2015; Ganesan et al., 2017; Duhen et al., 2018; Edwards et al., 2018; Savas et al., 2018; Egelston et al., 2019). However, CD69 is also present on activated T cells, acting as a “stand your ground” signal promoting T differentiation in the LN (Baeyens et al., 2021). CD103 expression is induced by TGFβ signaling, and also observed on T effectors entering tissue 4 days post priming (Ericsson et al., 2004; Hardenberg et al., 2018). Given high TGFβ levels in the tumor microenvironment (Loffek, 2018), CD103 expression on TILs may not be a precise marker for T_RM status. Importantly, the relative enrichment and relatedness of resident memory TIL phenotypes against other TIL classification states including dysfunction, activation, and exhaustion was not tested, making it difficult to prioritize scoring of one state over another. This has made it challenging to distinguish residency programs from others, such as exhaustion, where some overlaps between these programs are now being noted.

Overall, the diversity and complexity of various approaches have made it challenging to provide an unbiased classification and deconvolve TIL programming. Here, we used an integrated approach to systematically classify both bulk tumors and individual human TILs using stringently defined, large T cell differentiation state trajectories, and large signatures, derived across many model systems. Our findings point to a revised model of TILs differentiation, wherein poor outcomes and a lack of ICB responsiveness may represent failed TILs progression from an early activation-like state, and/or failed generation of persistent memory. These data highlight the need for improved TILs classification methods and suggest combination therapies with ICB to promote successful memory formation.

RESULTS:

Protective T_RM and human melanoma TILs co-express T cell inhibitory checkpoints, activation, and residency markers.

Our understanding of T cell differentiation, and benchmarked states of differentiation stems from basic immunology where the behavior of T cells is studied during acute and chronic infections. However it is unclear how faithfully these states capture TILs differentiation in the TME. We sought to initially profile CD8⁺ T cells generated after infections, and CD3⁺CD8⁺ melanoma TILs for expression of immune checkpoints, activation, and residency markers which are commonly used to classify T cells in both settings.

To profile CD8⁺ T cells after infections, we queried publicly available datasets of CD8⁺ T naive (T_naive), effector-memory (T_EM)_, central memory cells (T_CM), and T_RM isolated from murine skin, lung, gut, or brain after herpes simplex virus (HSV), influenza, LCMV (Mackay et al., 2013), and vesicular stomatitis virus (VSV) infection (Wakim et al., 2012). Strikingly, we observed T_RM express high levels of immune checkpoint transcripts (Pdcd1, Ctla4, Lag3, Havcr2, Tigit) when compared to T_naive, T_EM, and T_CM (Fig 1A). Conserved high expression of immune checkpoint transcripts was seen across tissues and infections, with the highest Pdcd1 levels seen on skin and lung T_RM. Ctla4 and Tigit transcripts were elevated across skin, lung and gut T_RM, while Havcr2 transcripts were high in skin T_RM, relative to naïve T cells or T_CM/T_EM (Fig 1A). In addition, T_RM across these sites were also observed to co-express high levels of immune activation transcripts (Gzma, Gzmb, Gzmk, Ifng, and Prf1) (Fig 1B). Additionally, we noted that co-expression of both immune checkpoints and activation transcripts is controlled by the transcription factor Runx3, a master regulator of tissue residency and cytotoxicity (Wang et al., 2018) using available data on Runx3 modulation of CD8⁺ T cells (Milner et al., 2017). Loss of Runx3 led to loss of activation and immune checkpoint transcripts, while enforced expression of Runx3 drove high expression of both (Fig 1B).

Figure 1: Human and mouse T_RM, and human metastatic melanoma TILs co-express immune activation, inhibitory checkpoint molecules, and tissue-residency markers.

A. Relative expression of immune checkpoint transcripts in murine T_RM across organs and infections: influenza (lung), lymphocytic choriomeningitis virus (LCMV) (gut), herpes simplex virus (HSV) (skin) and vesicular stomatitis virus (VSV) (brain), or in dendritic epidermal T cells (DETC). Transcript levels are expressed as a normalized ratio to naive T cells from each respective data set and includes GSE47045(Mackay et al., 2013) (n=3 samples per condition) or GSE3915 (Wakim et al., 2012) (brain, n=3–5 replicates per condition). Significance values for the expression level of transcripts relative to naïve T cells were calculated using an unpaired Student’s t test.

B. Left: Heat map of immune checkpoint and activation transcripts in murine T_RM across tissue sites and infections generated using GSE47045 (Mackay et al., 2013). Right: Heat map of immune checkpoint and activation molecules in T cells infected with Runx3 cDNA-expressing retrovirus generated using GSE106107 (Milner et al., 2017). RNA-seq analyses of Runx3-deficient (fl/fl) or overexpressing cells (+++) (n=2 biological replicates of approximately 20 million paired-reads per sample).

C. Steady-state human skin T_RM from pooled patient epidermal sheets. Expression of PD-1 on epidermal CD45⁺CD3⁺ CD8⁺ T cells. CD8⁺ T cells were sub-gated further into CD69⁺CD103⁺ or CD69⁺CD103⁻ T_RM, or recirculating CD103⁻CD69⁻ T cells as defined previously (Watanabe et al., 2015).

D. Analysis of 1500 human T cell melanoma TILs isolated from 13 patients including lymphoid and non-lymphoid metastases (GSE72056 (Tirosh et al., 2016)). Top: T cells rank ordered by T_RM score. Bottom: Heat map of the top vs. bottom 82 T_RM genes in rank ordered T cells.

E. Correlation between the expression of immune checkpoint and activation markers and TILs T_RM scores, based on genes upregulated or downregulated in T_RM.

F. Representative PD-1 flow cytometry of CD45⁺HLA-DR⁻ CD8⁺ T cells stained for CD103 and CD69 isolated from a LN melanoma metastasis.

G. Percentage CD69⁺ CD103⁺ expressing CD8⁺ T cells isolated from the same individuals non-involved LN (NIN) or tumor involved LN (TIN) with metastatic melanoma (n=5 paired specimens). Significance values were calculated using a paired Student’s t test.

H. Percent PD-1⁺ or percent granzyme B⁺ of CD3⁺ CD8⁺ T cells in each of the 4 quadrants based on CD69 and CD103 expression testing 19 patient specimens with metastatic melanoma as compared to PBMCs from 3 healthy volunteer donors (ND). Samples were taken from PBMC of healthy volunteers (n=3), NIN (n=5) and TIN (n=6), skin metastases (n=6) or small bowel metastases (n=7). Data shown as mean ± SD. Significance values were calculating using an unpaired Student’s t test. *, p≤0.05; **, p≤0.01; ***, p≤0.001; ****, p≤0.0001.

See also Figure S1 and Tables S1–S2.

T_RM are non-recirculating and express tissue retention proteins such as CD69 and CD103 (Gebhardt et al., 2009; Kumar et al., 2017; Behr et al., 2018). CD69 associates with and inhibits sphingosine 1-phosphate receptor-1 (S1P1) function, which inhibits T cell egress from tissue, and suggests tissue residency (Bankovich et al., 2010; Mackay et al., 2015); CD103 (α_Eβ₇) is important for epithelial T_RM formation (Casey et al., 2012) and epidermal retention (Mackay et al., 2013). Therefore, to confirm these observations on the protein level and in humans, we analyzed CD3⁺ CD8⁺ T cells from human surgical skin samples, testing co-expression of PD-1, CD69 and CD103. Cell-surface PD-1 protein is highly expressed on epidermal CD69⁺CD103⁺ and CD69⁺CD103⁻ human skin T_RM (Fig 1C, S1A). Collectively these data indicate T_RM co-express tissue residency markers, immune checkpoints traditionally associated with T cell exhaustion, and immune activation transcripts.

Next we sought to profile human TILs. Some human TILs have been described as T_RM-like with properties of residency including expression of CD69⁺ and CD103⁺ (Boddupalli et al., 2016; Corgnac et al., 2020). We wanted to test if like T_RM, TILs with tissue-residency properties co-express immune checkpoints and activation transcripts. We generated a T_RM core/consensus signature spanning multiple infections and sites using public data sets (Mackay et al., 2013) (Table S1–S2). We scored the single-cell RNA-sequencing transcriptomes of 1500 individual CD3⁺ TILS previously isolated from 13 metastatic melanoma patients using (Tirosh et al., 2016) against this core T_RM signature of 121 transcripts, of which 76 were detectable by single cell. Roughly two-thirds of human melanoma TILs expressed a high T_RM score (Fig 1D). TILs bearing a high T_RM signature had high expression of both immune checkpoints (CTLA4, PDCD1, TIGIT, HAVCR2, LAG3) and immune activation markers (IFNG, PRF1, GZMA, GZMB and GZMK), while for TILs with a low T_RM score, the opposite was true (Fig 1E). These data suggest that TILs with residency programs co-express immune checkpoint and activation transcripts.

To validate this observation on the protein level, we conducted multi-parametric flow cytometry profiling of human melanoma TILs. In one representative patient, of CD3⁺ TILs, 90% of CD8⁺ T cells express CD69, suggesting residency (Fig 1F). Roughly one-third of CD8⁺ CD69⁺ T cells also co-expressed CD103 (Fig 1F), which binds E-cadherin and enables target tumor and epithelial cell lysis (Le Floc’h et al., 2007). All CD8⁺CD69⁺ T cells were marked by high cell-surface PD-1 protein (Fig 1F). We profiled an additional 19 metastatic melanoma patients, analyzing immune activation, checkpoint, and residency marker expression on CD3⁺CD8⁺ TILs. Metastatic sites surveyed include melanoma tumor involved lymph nodes (TIN, n=6), non-involved lymph nodes (NIN, n=5), skin metastases (n=6), and small bowel metastases (n=7). Of these, TIN and NIN were isolated as paired surgical specimens from the same 5 patients. A significant accumulation of CD8⁺ T cells co-expressing T_RM markers CD69⁺CD103⁺ was noted in TIN when compared to T cells isolated from NIN of the same donors, suggesting tumor-specific induction of residency programs (Fig 1G, S1B). CD8⁺ TILs co-expressing CD69 and CD103 from metastases were largely KLRG1^low and CD45RO⁺ suggesting prior antigen experience (Fig. S1B–G). Across various sites, the fraction of CD8⁺ TILs expressing PD-1 and granzyme B was highest within the CD69⁺CD103⁺ population, in contrast to low PD-1 on circulating T cells (Fig 1H). PD-1 cell-surface levels were also highest on CD69⁺ CD103⁺ TILs (Fig. S1C). Taken together, these data indicate that human TILs which enrich for either T_RM transcripts or T_RM cell-surface markers suggesting tissue residency, also co-express immune checkpoint and immune activation markers.

TILs, primary cancers, and metastatic melanoma display a higher relative enrichment of persistent memory and resident memory programs.

Given that both TILs and T_RM co-express markers common to resident memory T cells, activated T cells, and exhausted T cells, we sought to deconvolve these states from each other. To do this we conducted a global comparative analysis using methods to discern the relative enrichment of different T cell differentiation states. Using multiple public data sets, we generated T cell program signatures that are either comprehensive (non-exclusive), in which we did not remove transcripts that overlap multiple states, or unique to a particular T cell state, by exclusion of overlapping transcripts (exclusive). We scored individual melanoma CD3⁺ TILs against the top 400 genes present in activated (T_act), exhausted (T_exh), memory (T_mem), and T_RM signatures (Fig 2A, Table S1–S2). Strikingly, an individual TIL could score for enrichment of more than one T cell state signature, preventing assignment using any single signature (Fig 2B). Rather, these data demonstrate comparative classification methods are necessary.

Figure 2: Human melanoma and other cancers show a relative enrichment of T memory and resident memory programming.

A. Derivation of viral T cell signatures. Top: schema depicting derivation of non-exclusive signatures. Bottom: Venn diagram depicting removal of shared transcripts from signatures to generate exclusive signatures. For T_act and T_mem, early and late signatures were unionized before removing overlapping transcripts.

B. Scoring of 2608 single cell melanoma T cells by the expression of exclusive viral T cell signatures. Cells were rank ordered by T_RM scores. Scores are defined by the average expression (log2-transformed and centered) of upregulated (e.g. T_act-high) signatures minus the average expression of downregulated (e.g. T_act-low) signatures.

C-D. Enrichment of exclusive viral T cell signatures in (C) primary (n=80) or metastatic (n=368) melanoma, or (D) multiple primary cancers. Data in (C) shown as boxplots, with top/middle/bottom lines indicating the 3^rd quartile, median, and 1^st quartile, respectively. Whiskers indicate minimum and maximum values, and dots represent outliers. Significance values are shown for the comparison between primary and metastatic melanoma using the Wilcoxon rank sum test. Data in (D) shown as a bubble plot, with both color and size of points indicating mean enrichment score.

See also Figure S2 and Tables S1–S2.

Our single cell analysis suggested a conserved pattern of signature enrichment when using either exclusive or non-exclusive signatures (Fig 2A, S2A), with both memory and resident memory consistently scoring above T cell activation and exhaustion. However, because single cell analysis captures a fraction of the transcriptome, we reasoned that it might not be of sufficient data depth to truly rank enrichment of one signature over. Therefore, we conducted comparative enrichment analysis of T cell state signatures against bulk primary (n=80) and metastatic melanoma (n=368) specimens. Again, we noted the same pattern of signature rank scoring. Circulating and resident memory signatures scored above activation and exhaustion (Fig 2C, S2B). Interestingly, both primary and metastatic melanoma maintained this ranked classification, suggesting the early vs. late tumor microenvironment did not drastically change differentiation status. Rather, a statistically significant higher enrichment score was present in metastatic above primary melanoma suggesting tumor metastases may further enforce such programs (Fig 2C). In addition, we tested if this ranked enrichment pattern is present in other primary cancers by pan- genome analysis of bulk sequencing data across the TCGA primary cancer database. Patterns of ranked enrichment were similar across primary cancers (n=9084 samples, Fig 2D, S2C–D). Collectively, these data suggest individual TILs, primary cancers, and metastatic melanoma display a higher relative enrichment of persistent memory and resident memory programs.

ICB non-responding melanoma tumors have a high fraction of activated TILs while responding tumors expand a T resident/long term memory cluster suggesting persistence.

Given that individual TILs express markers associated with activation, exhaustion, and tissue residency and score for more than one T cell signature, we considered whether TILs form a unique developmental trajectory, and whether this may be poorly captured by any single cell state. TILs may differentiate in a manner either intermediate to, or distinct from benchmarked states. Therefore, we analyzed single cell melanoma TILs using monocle software, which orders single cells based on transcriptome differences to construct an unbiased differentiation trajectory (Trapnell et al., 2014). We examined CD3⁺CD8⁺ TILs that had been isolated from both ICB responding and non-responding lesions (32 patients total, 24 αPD-1 recipients and 8 combo αCTLA-4/αPD-1 recipients) (Sade-Feldman et al., 2018). These were divided into 5 trajectory clusters, with 3 dominant clusters (1,3,5) encompassing most cells. Two minor clusters contained relatively few cells (2,4) and were not analyzed further (Fig 3A, S3A–D). Non-responders had a higher fraction of cluster 1 cells than responders, whereas cluster 3 was represented at a higher frequency in samples from αPD-1 responders (Fig 3B, S3E–F) driven primarily by a higher frequency of cluster 3 cells post-response (Fig 3C–D, S3G). Also, within non-responders, a statistically significant increase in cluster 1 over cluster 3 (p< 0.01) and cluster 5 (p < 0.05) was observed, whereas responders had a statistically significant increase of cluster 3 over cluster 1 (p < 0.01).

Figure 3: Three major single cell trajectory states define responder and non-responder status.

A. Pseudotime trajectory for CD8A⁺CD3E⁺ co-expressing T-cell clusters identified by pre-treatment vs. post-treatment and response status (responder vs. non-responder), respectively. Cells are colored by assigned pseudotime cluster.

B. Relative percentage of cells (CD8A⁺CD3E⁺) from each pseudotime cluster in non- responder vs. responder samples. Dots represent individual samples. Differences between groups were modelled using mixed-effects modeling (MEM) (see STAR Methods), bars graphs and error bars show estimated marginal mean ± SEM. *, p≤0.05; ** p≤0.01; *** p≤0.001; **** p≤0.0001.

C. Relative percentage of cells from each pseudotime cluster in all samples separated by baseline and post treatment in responders and non-responders.

D. Relative percent of CD8A⁺CD3E⁺ cells derived from each cluster per patient, and averaged across all patient samples, separated by pre/post treatment status and by non-responsive/responsive sites. Differences in cluster frequencies between non-responders and responders was modelled using MEM (see STAR Methods). Data shown as estimated marginal mean ± SEM. *, p≤0.05; ** p≤0.01; *** p≤0.001; **** p≤0.0001.

E. Left: heatmap depicting the expression of pseudotime cluster Differentially Expressed Genes (DEGs) in CD8A⁺CD3E⁺ co-expressing TILs (FDR <0.05). Right: top pseudotime cluster DEGs (FDR < 0.01, sorted by |logFC|) are annotated.

See also Figure S3, Table S1 and Table S3.

To test if trajectory cluster assignment was driven by biological factors such as cell cycle, which have been documented to influence single cell gene expression (Barron and Li, 2016), we regressed the cell cycle effect from input cells before recomputing the pseudotime trajectory. Doing so did not measurably change the trajectory shape and cluster assignments (Fig. S3H–I). We also sought to test whether pre-selecting TILs by markers used to distinguish bystander T cells and exhausted T cells modifies the association of a cluster to response status. We selected CD3⁺CD8⁺ TILs based on ENTPD1/CD39, a marker used to help distinguish tumor antigen specific T cells from CD39^low bystanders, (Simoni et al., 2018), and TOX, a transcription factor that marks and promotes T cell exhaustion (Alfei et al., 2019; Khan et al., 2019; Mann and Kaech, 2019; Scott et al., 2019). However, even after dividing cells based on ENTPD1 and TOX expression, ICB non-responder sites had higher fractional representation of cluster 1, while responders had a higher fraction of cluster 3 cells (Fig. S3J–K). Collectively this suggests pre-classification of cells based on commonly used individual classifiers such as those used to mark exhausted cells does not modify the association of key TILs clusters with response status. Hence, these data identify cluster 1 as non-responder associated, and cluster 3 as responder enriched post therapy. We also computed differentially expressed genes (DEGs) specific to each of the major clusters (Fig 3E, Table S1, S3). Cluster 1 was notable for the down-regulation of stem and memory genes IL7R and TCF7, the transcription factors FOSB, FOSL2, and JUN, and nuclear receptor NR4A3. Cluster 3 showed an opposing trend with increased IL7R, FOSL2, and NR4A3. Cluster 5 was distinguished from the other 2 clusters by a smaller handful of transcripts.

To better understand how cluster-specific transcripts relate to the biology of T cells, we conducted pathway analysis of cluster DEGs. Cluster 1 DEGs positively enriched for G1/S and oxidative phosphorylation pathways, and G2/M checkpoints, and negatively enriched for AP-1 downstream transcripts and pathways related to protein translation (Fig 4A, Table S4). In contrast, cluster 3 DEGs positively enriched for AP-1 related transcripts, G2/M transition, and negatively enriched for oxidative phosphorylation pathways (Fig 4A, Table S4). Cells in cluster 5 had relatively few DEGs and showed negative enrichment of allograft rejection and cytokine signaling pathways (Fig 4A, Table S4).

Figure 4: Non- responder cluster most closely associates to early activation signatures while responder associated cluster 3 and 5 are most closely associated to resident and central memory respectively.

A. Pathway overrepresentation analysis of DEGs defining each pseudotime cluster. Positive or negative enrichment values indicate that pathway enrichment is driven by DEGs respectively up- or downregulated in cells in that cluster.

B. Changes in pseudotime cluster program expression (average z-score) during T cell differentiation in response to (L to R): LCMV Armstrong infection (GSE10239 (Sarkar et al., 2008); GSE41867 (Doering et al., 2012)), LCMV cl. 13 infection (GSE41867), autochthonous tumor induction (GSE89307 (Philip et al., 2017)), vaccinia virus infection (GSE79805 (Pan et al., 2017)), and human yellow fever vaccination (naïve → Activation: GSE26347; naïve → activation → memory: GSE100745 (Akondy et al., 2017)). Data points and error bars indicate mean and 95% CI of the relative enrichment of cluster DEGs. Black line and grey ribbon indicate the same for an equal number of random transcripts. Significance values are shown for the comparison between signature enrichment at a given timepoint and the timepoint immediately preceding it, which was carried out using mixed-effects modeling (see statistics section). For vaccinia virus infection, a linear regression model was fitted (shaded) and the change in cluster program activity from d0 to d90 is annotated along with significance of the rate of change. For GSE100745, significance values for enrichment at a given timepoint vs the timepoint preceding it were calculated using a Student’s t test. *, p≤0.05; ** p≤0.01; *** p≤0.001; **** p≤0.0001.

See also Figure S4 and Tables S1, S4, and S5.

In T cells, metabolism and cell cycle are influenced by the quality of T cell stimulation (Macian et al., 2001; Ohtsuka et al., 2016; Tan et al., 2017), which can influence T cell differentiation. To understand when during T cell development these melanoma TILs cluster programs are enforced, we queried each against early to late T cell differentiation trajectories. We used temporal trajectories occurring in response to several distinct initiating cues: productively cleared gut LCMV Armstrong infection, resulting in protective T cell memory (Sarkar et al., 2008; Doering et al., 2012), chronic LCMV clone 13 (LCMV cl. 13) infection driving T cell exhaustion/dysfunction (Doering et al., 2012), a time course of autochthonous liver cancer in which SV40 large T antigen (TAG) is an oncogenic driver leading to distinct phases of early and late dysfunction in the setting of chronic antigen (Philip et al., 2017), and protective vaccinia virus skin scarification (VACVss) which generates circulating T_CM, T_EM, and skin T_RM (Pan et al., 2017). We also tested two human datasets of naïve, effector, and memory CD8⁺ T cell programs from individuals who received the human yellow fever (YF) vaccine (Akondy et al., 2017). We applied sample-wise gene set enrichment, which allows us to map dynamic modulation of each cluster program. The trajectory cluster 1 (non-responder associated) program was significantly upregulated at early timepoints following all mouse and human initiating challenges, and was downregulated in all late time points (Fig 4B). These data show that non-responder enriched clusters are associated with T cell activation states irrespective of tissue site, species, and acute or chronic antigen exposure.

In contrast, cluster 3 transcripts (responder-associated) did not score highly during early or late acute (cleared) or chronic (exhaustion associated) LCMV, or during any early or late stage of chronic SV40-TAG liver tumorigenesis driven T cell dysfunction. Rather cluster 3 DEGs were only steadily upregulated as skin T_RM are generated by VACVss and when long-lived memory CD8⁺ T cells were generated 4–12 years after YF vaccine (Fig 4B). These data relate cluster 3 programming to long-lasting human memory T cell programming, and tissue resident memory states. Finally, in direct contrast to non-responder associated cluster 1 transcripts, cluster 5 DEGs were significantly downregulated acutely at early time points during all initiating stimuli (Fig 4B). We also scored individual TILs for enrichment of viral T cell signatures (Seurat’s AddModuleScore function, (Tirosh et al., 2016; Butler et al., 2018)). Again, cluster 1 highly expressed T_act genes d4.5 and d6 post infection while transcripts up in resident memory cells had higher activity within cells in clusters 3 and 5 (Fig. S4A–D). This matched our findings scoring cluster DEGs in T cell kinetic trajectories and occurred irrespective of whether TILs were isolated from responder or non-responder samples and using state unique (exclusive) or comprehensive (non-exclusive) signatures (Fig. S4A–D). We then tested non-responder vs. responder associated DEGs independent of cluster assignment (Table S1, S5). Global non-responder DEGs were upregulated at early timepoints, similar to cluster 1 transcripts (Fig. S4E). Taken together, these data suggest that human TILs can be modeled along a representative spectrum of early activation-like to persistent memory-like clusters. Increased frequency of early activated-like T cells was associated with a lack of responsivity, whereas increased presence of persistent memory/resident memory-like T cells was associated to positive αPD-1 responses in patients.

Viral and vaccine driven T cell states more strongly overlap the transcriptome of human melanoma TILs clusters than canonical pre-clinical models of dysfunction or exhaustion

Next, we sought to compare non-responder and responder associated melanoma TILs to T cell programs from pre-clinical models currently being used to capture TILs differentiation, and to human CD8⁺ T cell responses generated after protective immunization. In order to generate a comparison across multiple disparate models, we applied over-representation analysis which tests the relative overlap of melanoma TILs clusters 1,3, and 5 DEGs above those occurring by random chance. We scored overlaps with T cell signatures derived from 19 unique pre-clinical and clinical models. The overlap was computed between trajectory cluster DEGs and signatures from: a) virally generated activation, memory, resident memory and exhausted T cells, b) murine B16 melanoma TILs dysfunction/activation gene modules (Singer et al., 2016), c) TILs derived from an autochthonous murine liver cancer model both with and without Tox (Philip et al., 2017; Scott et al., 2019), a transcription factor driving T cell dysfunction in tumor (Scott et al., 2019) and exhaustion to virus (Khan et al., 2019; Charmoy et al., 2021), d) exhausted subsets and their progenitors generated using a murine model of chronic LCMV cl. 13 infection (Im et al., 2016; Hudson et al., 2019; Beltra et al., 2020), e) 4 different models of T cell anergy, another state of tolerized T cells (Safford et al., 2005; Zha et al., 2006; Provine et al., 2016; Brignall et al., 2017), and f) human vaccine responses to yellow fever immunization (Akondy et al., 2017) (Fig 5A–B, Table S1,S6,S7). Considering both the significance and the magnitude (odds ratio) of the overlap, cluster 1 transcripts again most strongly overlapped genes present during murine and human T cell activation (24/190 genes overlapping murine T activation up, 39/190, and 65/190 genes overlapping human YF activation up, Fig 5A–B, S5A). Cluster 1 transcripts more weakly overlapped memory, and Tox-dependent T cell signatures (Fig 5A). If we then scored individual single cells for each of these signatures, we observed that in addition to the enrichment of activation signatures in cluster 1 TILs, enrichment of intermediate and terminally differentiated exhausted programs exists (CD101⁻TIM3⁺, (Hudson et al., 2019) and CXCR5⁻TIM3⁺, (Im et al., 2016), Fig. S5B). However, when we removed activation transcripts from the exhaustion signatures, exhaustion programs no longer associated with cluster 1 (Fig. S5C–E), indicating the association of exhaustion signatures to cluster 1 was driven by shared expression of activated T cell transcripts. A lesser association of cluster 1 to murine memory was observed (Fig 5A, S5F), but other murine memory transcripts were downregulated in cluster 1, and human memory did not overlap cluster 1 (Fig 5A–B). Collectively, these data reinforce our prior observation that cluster 1 is most closely associated to early human and mouse T cell activation, using independent scoring methods.

Figure 5: Human melanoma TILs associated to ICB non-response/response show higher overlap with early viral activation and resident memory signatures than multiple exhaustion, dysfunction, anergy models.

A-B. Left: schemas depicting generation of (A) murine T cell signatures, or (B) human vaccine response signatures. Right: Over-representation analysis of pseudotime cluster DEGs with corresponding signatures. Size indicates significance of enrichment (−log10(p value)). Color indicates magnitude of signature overlap (log2(Odds Ratio)). Grey squares indicate non-significant enrichment (FDR < 0.05).

See also Figure S5, Tables S1, S6 and S7

In addition, over-representation analysis revealed cluster 3 transcripts overlapped murine T_RM, and long-term human memory (4–12yrs) (19/263 genes overlap murine T_RM up, 27/263 genes overlapping human memory up, Fig 5A–B, S5G), as previously observed using sample-wise gene set enrichment of T cell differentiation trajectories (Fig 4). To a lesser extent and with lower significance and magnitude, cluster 3 transcripts overlapped some transcripts present 14 days after yellow fever immunization. Also consistent with our prior findings, murine and human activation genes were downregulated in cluster 5, (12/86 down genes overlapping murine activation, 22/86 down genes overlapping human activation, Fig 5A–B). Taken together, these results expand and validate the prior findings that αPD-1 non-responder associated melanoma TILs clusters overlap T cell activation, but not exhaustion, exhaustion progenitors, or anergy. Responder-associated clusters overlap some transcripts found in murine T_RM and human long-term memory.

Other human TILs characterized as exhausted or dysfunctional associate to early activation T cell states

Next we tested these observations in 3 additional patient cohorts, which include TILs isolated from melanoma and non-melanoma solid tumors. TILs sets were previously generated from: a) human kidney, bladder, and prostate cancer patients and described as stem- vs terminally differentiated (TD)-like (Jansen et al., 2019), b) hepatocellular carcinoma (HCC) patients, that expressed high immune checkpoints (Zheng et al., 2017), and c) melanoma distinguished by using gene modules co-expressed with LAG3 or FGFBP2 to mark dysfunctional and cytotoxic populations, respectively (Li et al., 2019) (Table S1,S8). We used DEGs from 3 sets of TILs which had been previously annotated as exhausted or dysfunctional (Jansen_TD-like, Zheng_LAYN, and Li_Dysfunction). We observed these DEGs overlapped each other and non-responder associated cluster 1 in melanoma, suggesting conservation of programming across some cancer types (Fig 6A, S6A–C). DEGs from these previously annotated exhausted or dysfunctional TILs strikingly overlapped murine and human T activation state signatures. In kidney cancer TILs (Jansen_TD-like) 60/307 transcripts identified overlap murine T activation exclusive signatures, and 71/307 and 114/307 transcripts overlap human T effector programs after yellow fever vaccination, analyzing 2 separate yellow fever data sets (Fig 6A). When scored using sample-wise gene set enrichment to analyze dynamic changes in program induction, these other human TIL signatures were again closely associated with murine and human activation (Fig 6B). In contrast, signatures from stem-like TILs showed an opposing trend, and were transiently downregulated during early activation, but later upregulated during memory timepoints following cleared LCMV infection and YF human vaccination (Fig 6B, Jansen_Stem-like). These results suggest that TILs from other cancers share conserved programs with each other and with melanoma associated clusters, and their programs also relate closely to those seen during viral activation or memory.

Figure 6: Human TILs classified as dysfunctional/exhausted from independent datasets associate to pseudotime cluster 1, and early activation over terminal exhaustion.

A. Overrepresentation analysis of human TILs signatures with left: pseudotime cluster DEGs, or right: exclusive viral T cell signatures and human vaccine response signatures. Size indicates significance of enrichment (−log10(p value)). Color indicates magnitude of signature overlap (log2(Odds Ratio)). Grey squares indicate non-significant enrichment (FDR < 0.05).

B. Relative enrichment of human TILs signatures in timecourses of T cell differentiation. Signatures from TILs labeled as dysfunctional/exhausted are bracketed. Data points and error bars indicate mean and 95% CI of the relative enrichment of TILs signatures. Black line and grey ribbon indicate the same for an equal number of random transcripts. Significance values are shown for the comparison between signature enrichment at a given timepoint and the timepoint immediately preceding it, which was carried out using mixed-effects modeling (see statistics section). For vaccinia virus infection, the slope of DEG enrichment over the 90 day timecourse (shaded) was calculated using lm, with slopes and significance values annotated. For GSE100745, significance values for enrichment at a given timepoint vs the timepoint preceding it were calculated using a Student’s t test. *, p≤0.05; ** p≤0.01; *** p≤0.001; **** p≤0.0001.

See also Figure S6, Tables S1 and S8.

Persistent T cell signatures stratify melanoma survival and correlate with each other suggesting co-expression in the same tumors, while early activation signatures do not predict survival or correlate.

Next we tested if we could leverage these observations to predict melanoma survival. We scored metastatic melanoma tumors for high, medium, and low expression of each T cell state signature, and tested if this distinguished overall survival outcomes. Our TCGA test cohort was comprised of 355 patients with metastatic melanoma with 15 year survival outcome data, and who did not receive immunotherapy. T cell signatures of early activation did not stratify patient survival. Instead, strikingly, all late differentiation signatures including memory, resident memory, and exhaustion correlated with survival (Fig 7A, n = 355 patients, 15 year test cohort). This observation could not be explained by the presence of immunologically “hot” tumors, or those with a simple T cell infiltrate, as T cell transcripts associated to activation did not stratify survival (Fig 7A, S7A). T_RM signatures also stratified primary and metastatic melanoma survival (n= 80 primary melanoma specimens, T_RM non-exclusive p<.05, T_RM exclusive p=0.0522, n = 349 metastatic melanoma specimens, T_RM core p<.05, Fig. S7B–d). Next, we tested an independent dataset of 121 metastatic melanoma patients treated with αPD-1 (Liu et al., 2019) to validate these findings, and test relevance to current immuno-oncology regimens. Again, T cell signatures representing late differentiation states stratified survival, whereas early activation signatures did not (Fig 7B, n = 121, validation cohort). This trend was maintained despite the shorter length of follow-up of the validation cohort (50 vs. 400 months) (Fig 7B). Additionally, T cell state signatures better predicted melanoma survival than commonly used single T cell markers, as these markers no longer stratify survival in the validation cohort of ICB recipients (Fig. S7E). Collectively these data indicate an activation to persistent memory spectrum can prognosticate melanoma patient survival in cohorts receiving, and not receiving, ICB.

Figure 7: Late persistent memory states and exhaustion positively stratify overall melanoma survival, and the response to ICB, and correlate with each other.

A-B. Patient survival stratification testing non-exclusive T cell signatures in (A) metastatic melanoma specimens (n=355), or (B) metastatic melanoma patients treated with αPD-1 (n = 121). High, medium, and low expression of each signature was used to stratify patient survival.

C. Correlation scores between exclusive T cell signatures, a DC maturation signature and an IFN-γ response signature, comparing samples across metastatic melanomas irrespective of metastatic site (n=367, Pearson), primary cutaneous melanomas (n=103, Pearson), all primary cancers (n = 9127, Pearson), and in metastatic melanoma patients treated with αPD-1 (n = 121, Spearman).

D-E. Patient survival stratification testing d. human yellow fever effectors and TD-like TILs from kidney/bladder/prostate cancer, or e. yellow fever memory and stem-like TILs signatures from kidney/bladder/prostate cancer in left: metastatic melanoma specimens (n=355), or right: metastatic melanoma patients treated with αPD-1 (n = 121). Survival statistics represent Mantel-Cox analysis. *, p≤0.05; **, p≤0.01; ***, p≤0.001.

See also Figure S7.

Because memory, resident memory and exhaustion have been reported to derive from the same KLRG1⁻ memory progenitor (Angelosanto et al., 2012; Mackay et al., 2013), we tested if all persistent signatures might stratify survival because they are co-expressed within the same tumors. Due to a high overlap of transcripts, non-exclusive signatures are highly correlated (Fig. S7F). However, testing exclusive signatures, with no overlapping transcripts, we found that T_mem, T_exh and T_RM signature expression correlate with each other, but not with T cell activation (Fig. 7C). The most highly correlated states were T_mem and T_RM across melanoma metastases irrespective of site or cancer (n=367 TCGA metastatic melanoma, Pearson’s correlation coefficient r=0.88; primary cutaneous melanoma n=103 TCGA SKCM, r=0.83; all primary cancers r=0.83, and metastatic melanoma treated with αPD-1 n=121 Spearman correlation coefficient ρ=0.71, Fig 7C). This would be anticipated if TILs undergo memory-like differentiation broadly. Strikingly we also observed a correlation between expression of resident memory and late exhaustion signatures (Fig 7C, metastatic melanoma r=0.73, TCGA SKCM r=0.61, TCGA Pan Cancer r=0.72, Metastatic melanoma treated with αPD-1 ρ=0.54). All persistent T cell states were correlated closely with both a 227-gene skin Dendritic Cell (DC) signature (Fig 7C, Ex: T_RM metastatic melanoma r = 0.93, TCGA SKCM r = 0.85, TCGA Pan Cancer r = 0.88, Metastatic melanoma treated with αPD-1 ρ=0.81), and a 200-gene IFN-γ response signature that stratifies melanoma survival (Nirschl et al., 2017) (Fig 7C, Ex: T_RM metastatic melanoma r = 0.82, TCGA SKCM r = 0.79, TCGA Pan Cancer r = 0.74, Metastatic melanoma treated with αPD-1 ρ=0.63). These data suggest late immune programs favoring memory persistence (memory, resident memory and exhaustion) co-occur and are co-expressed with DC and IFN-γ response programming in melanoma. Further validating these findings, both Jansen_Stem-like and YF memory transcripts stratify melanoma survival in both test and validation cohorts, while signatures from TD-like TILs and effector responses do not (Fig 7D–E, S7G). Collectively these data demonstrate that a spectrum of activation to persistent memory/resident memory can help to stratify melanoma survival with persistence and memory/resident memory associating to DCs and IFN-γ responsivity, and positively predicting patient overall survival and ICB response.

DISCUSSION:

Here, we apply a systems biology approach to multiple single cell and bulk RNA-seq datasets in order to gain a better understanding of CD8⁺ TILs differentiation and relatedness to patient survival and ICB response. Several key themes emerged. First, we found that current approaches that assign TILs status based on expression of a small handful of transcripts such as immune checkpoints are too reductive to discern TILs status. Instead, comparative scoring methods are required to better classify TILs differentiation. This knowledge could then be used to frame variables shaping TILs differentiation such as the TME, and also to consider therapeutic interventions to direct TILs differentiation along one trajectory or another.

Second, we find that melanoma TILs can be modeled along an early activation-like to memory-like spectrum using multiple systems-based approaches spanning 7 detailed kinetic trajectories of T cell differentiation and 33 signatures generated in T cells after viral infections, exhaustion/exhaustion intermediates, anergy, tumor-dysfunction models, human vaccine responses, and other human cancers. The over-representation of cluster 1 in ICB non-responders which scores for T cell activation, but shows striking enrichment of G1/S blockade and loss of translation may suggest failed memory conversion. Because this early-activated like state did not overlap T cell exhaustion or exhaustion progenitors, selection of exhausted T cells may need to be performed on tumor-antigen specific cells and not performed as a global classifier, as is currently used based on redundant markers. Despite this, clinically relevant TILs classification can still be performed in the absence of such selection. Strikingly terminally differentiated TILs from kidney, bladder, and prostate cancer previously considered exhausted, have a striking overlap of 114 of 307 genes to human yellow fever effector programming, suggesting relevance of these observations beyond melanoma. Conversely, a melanoma responder-associated cluster enriched during skin T_RM formation, and transcripts overlapped found in long-lived memory T cells, present 4–12 years after human immunization. Given these findings, cluster 3 may not represent memory or resident memory directly, but rather may represented a shared progenitor population to long-lived memory cells.

Third, the tumor microenvironment may better support differentiation towards either an early activation or memory-like state at large. On the bulk level and across multiple data sets, persistent signatures are co-expressed in the same tumors, while activation signatures do not co-express with memory. It is possible for both ICB responsive and non-responsive lesions to be found within the same patient (Sade-Feldman et al., 2018; Topp et al., 2021) suggesting the TME may therefore dictate differentiation. Supporting this, when analyzing individual melanoma TILs, non-responding tumors have higher proportion of activation scoring TILs, while responding tumors have more memory-like TILs. Interestingly, DC maturation programs and IFN-γ response signatures co-express in the same tumors scoring for persistent signatures, and both have been linked to positive clinical outcomes (Nirschl et al., 2017). This analysis suggests that these populations may co-exist in a favorable tumor microenvironment (TME), raising the need to understand how persistent anti-tumoral memory is generated during LN priming, or preferentially sustained/reinforced within the TME. Our prior work showed TMB stratifies melanoma survival (Nirschl et al., 2017), and TMB was later found to predict of ICB response (Litchfield et al., 2021); however TMB scoring does not relate to DC and IFN-γ scoring in tumors, suggesting TMB is an independent prognostic factor and therefore not directly correlated with these memory T programs (Nirschl et al., 2017).

In viral infection models, successful stimulation of T cells by cross-presenting DC during priming in the draining LN is crucial for memory and resident memory formation (Henrickson et al., 2013; Kim et al., 2014; Iborra et al., 2016). Memory and resident memory T cells derive from the same KLRG1⁻ progenitor population, whose fate commitment is tied to inflammatory signals received during T cell priming (Joshi et al., 2007; Mackay et al., 2013). However, after LN priming, antigen re-encounter in tissue promotes persistent memory, as seen during T_RM formation (Khan et al., 2016; Muschaweckh et al., 2016). Re-stimulation of TILs in the TME may be a crucial driver of persistent memory. In support of this, TCF1⁺ stem-like TILs from kidney, prostate and bladder cancer co-localized with APCs (Jansen et al., 2019), indicating that APCs in tumor may help to sustain them. DC are necessary for ICB efficacy in mouse models, and an increased frequency of DC in human tumors is associated with ICB response (Salmon et al., 2016; Sanchez-Paulete et al., 2016; Barry et al., 2018). A role for DCs in improving sub-optimal LN priming, and intra-tumoral recall could both be important.

In addition, this study highlights our need to better understand on a functional level how persistent TILs with some properties overlapping memory/resident memory actively participate in anti-tumoral recall and/or predict positive overall outcomes and response to ICB. Memory populations possess potent cytotoxic capabilities (Barber et al., 2003; Steinbach et al., 2016). The T_RM marker CD103 promotes TIL cytotoxicity towards tumor cells (Le Floc’h et al., 2007). However, in addition to their killing ability, T_RM “sound the alarm” locally in tissues upon recall, and recruit other immune cells such as circulating memory T cells, DC, and natural killer cells to their location (Schenkel et al., 2013; Schenkel et al., 2014). This is interesting to consider given that in murine models ICB response is associated with T cell entry through tumor-associated high endothelial venules (TA-HEVs), and an increase in TA-HEV endothelial cell frequency and maturation increases the fraction of tumor-infiltrating stem-like CD8⁺ T cells (Asrir et al., 2022). Interestingly, cluster 3 remains higher in responder samples even when we filter on ENTPD1⁻ expressing TILs that are predicted to be bystanders (Simoni et al., 2018; Scheper et al., 2019). TILs persistence therefore may be a desirable metric even without selecting tumor antigen-specificity, and could be explained if tumor antigen specific memory cells recruit other bystander memory cells into tumors, as has been seen during T_RM recall. Another possibility is if memory in the TME serves as a benchmark to broader organismal immunity, which could protect against metastatic seeding or growth. In support of this concept, expanded T cell clonotypes in tumor also expand in peripheral tissue and blood (Wu et al., 2020). In pre-clinical models systemic immunity is necessary for anti-tumoral immune responses (Spitzer et al., 2017). Further work is needed to understand whether both tumor-specific and bystander resident memory-like TILs directly or indirectly respond to immune checkpoint blockade and by what mechanisms. Further work may also focus on understanding if, and how immune checkpoints limit global memory beyond the TME. It will be critical to understand the role of immune checkpoints like PD-1 in the generation, maintenance and recall of memory cells, including resident memory T cells.

These findings also suggest it will be important to establish variables driving an early activation-like state, without progression to memory. When considering factors that limit TIL differentiation towards memory we found that biological pathways defining non-response associated early activated TIL clusters include a lack of peptide translation, G2/M checkpoints, and AP-1 down-regulation. Peptide translation is actively regulated throughout CD8 T cell differentiation (Araki et al., 2017). Translation is compromised after sub-optimal T cell stimulation (Tan et al., 2017), and cell cycle arrest in S phase has also been observed in hypo-proliferative effector T cells stimulated with high concentrations of antigen (Ohtsuka et al., 2016). Sub-optimal stimulation of T cells results in NFAT signaling without AP-1 cooperation and may lead to anergy and exhaustion programming (Macián et al., 2002; Martinez et al., 2015). Although we observe AP-1 programming down in cluster 1 program pathways, we did not observe overlap with in vivo exhaustion or canonical anergy models including progenitor exhaustion states, suggesting this early activation TILs state that is associated to non-response status requires deeper characterization. The biological pathways driving cluster 1 programming are suggestive of sub-optimal T cell stimulation, yet it is unknown if insufficient stimulation occurs during anti-tumor T cell priming locally in the draining LN, how this can be distinguished in individuals who have both ICB responsive and non-responsive sites, and whether inadequate recall in the TME drives these changes. It is also possible that immunosuppressive features of the TME, such as metabolism, ER stress, and glucose availability, may contribute to an early activation arrest, terminal effector differentiation, or halted memory differentiation.

Our findings suggest the need to better classify and annotate TILs along their T cell differentiation trajectory. The key distinction of early activation halting from canonical exhaustion suggests serial therapeutics which first generate and then promoting durable T cell memory are likely to complement αPD-1 blockade in melanoma. Future work will test variables enabling simultaneous memory conversion and retention in tissue and tumor alike, and model therapeutic drug strategies to promote durable memory phenotypes in these settings. This may be achieved by enhanced immune priming, optimizing immune checkpoint inhibitor and drug combinations, and promoting micro-environmental cues that favor T cell memory.

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, Niroshana Anandasabapathy (nia9069@med.cornell.edu)

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

• This paper analyzes existing, publicly available data. These accession numbers for the datasets are listed in the key resources table. Datasets consulted included murine T_RM microarrays (GEO: GSE47045, GSE3915), Runx3 T cell RNA-seq (GEO: GSE106107), melanoma TILs single cell RNA-seq (GEO: GSE72056), ICB treated melanoma TILs single cell RNA-seq (GEO: GSE120575), TCGA bulk RNA-seq (URL: https://www.cancer.gov/tcga), αPD-1 +/− αCTLA-4 metastatic melanoma bulk RNA-seq (DOI: https://doi.org/10.1038/s41591-019-0654-5), murine LCMV Arm + cl 13 T cell microarrays (GEO: GSE41867), murine LCMV Arm T cell microarray (GEO: GSE10239), murine SV40-TAG TIL bulk RNA-seq (GEO: GSE89307), murine VACV–SIINFEKL T cell microarray (GEO: GSE79805), Human Yellow Fever vaccine T cell microarray (GEO: GSE26347), Human Yellow Fever vaccine T cell bulk RNA-seq (GSE100745), B16 dysfunctional CD8+ TILs RNA-seq (DOI: https://doi.org/10.1016/j.cell.2016.08.052), SV40-TAG bulk RNA-seq (GEO: GSE126973), LCMV cl 13 RNA-seq (GEO: GSE84105, DOI: https://doi.org/10.1016/j.immuni.2019.11.002, https://doi.org/10.1016/j.immuni.2020.04.014), anergy microarray and RNA-seq (DOI: https://doi.org/10.1038/ni1193, https://doi.org/10.4049/jimmunol.1602033, GEO: GSE5960, GSE73001), human kidney, bladder, prostate cancer TILs RNA-seq (GEO: GSE140430), hepatocellular carcinoma TILs scRNA-seq (DOI: https://doi.org/10.1016/j.cell.2017.05.035), and melanoma TILs scRNA-seq (DOI: https://doi.org/10.1016/j.cell.2018.11.043).

• All original code has been deposited at Zenodo and is publicly available as of the date of publication (DOI: https://doi.org/10.5281/zenodo.6419916). DOI is listed in the key resources table.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
anti-human CD279 (PD-1) Antibody (clone EH12.2H7)	Biolegend	Cat#329904
Mouse IgG1, κ Isotype Ctrl Antibody (clone MOPC-21)	Biolegend	Cat#400108
anti-human CD3 Antibody (clone UCHT1)	Biolegend	Cat#300428
CD103 (Integrin alpha E) Monoclonal Antibody (clone B-Ly7)	Invitrogen	Cat#17–1038–42
anti-human CD8 Antibody (clone SK1)	Biolegend	Cat#344724
anti-human CD49a Antibody (clone TS2/7)	Biolegend	Cat#328318
anti-human CD69 Antibody (clone FN50)	Biolegend	Cat#310906
anti-human CD4 Antibody (clone RPA-T4)	Biolegend	Cat#300512
anti-human CD45 Antibody (clone HI30)	Biolegend	Cat#304029
anti-human CD3 (clone SK7)	Biolegend	Cat#344811
anti-human CD3 (clone OKT3)	Biolegend	Cat#317317
Mouse Anti-Human HLA-DR (clone L243)	BD Biosciences	Cat#335796
anti-human TIM-3 (clone F38–2E2)	Biolegend	Cat#345001
anti-human CD45RO (clone UCHL1)	Biolegend	Cat#304202
anti-human GrzB (clone GB11), intracellular	Biolegend	Cat#515406
anti-human CD127 (clone A019D5)	Biolegend	Cat#351346
Human TruStain FcX (Fc Receptor Blocking Solution)	Biolegend	Cat#422301
Biological Samples
Healthy patient skin samples	Weill Cornell Medicine, Department of Dermatology	https://dermatology.weill.cornell.edu/
	Brigham and Women’s Hospital, Department of Dermatology	https://www.brighamandwomens.org/dermatology
Melanoma TILs samples	University of Virginia School of Medicine, Division of Surgical Oncology	https://surgery.virginia.edu/divisions-of-surgery/surgical-oncology-division/
Chemicals, peptides, and recombinant proteins
Collagenase 4	Worthington Biochemical	Cat#LS004188
Dispase II	Roche	Cat#04942078001
DNase I	Roche	Cat#10104159001
Critical commercial assays
Fixation/Permeablization Kit	BD Biosciences	Cat#554714
Fixable Aqua Dead Cell Stain Kit	ThermoFisher Scientific	Cat#L34966
Deposited data
Murine TRM microarrays	(Mackay et al., 2013)	GEO: GSE47045
	(Wakim et al., 2012)	GEO: GSE3915
Runx3 T cell RNA-seq	(Milner et al., 2017)	GEO: GSE106107
Melanoma TILs single cell RNA-seq	(Tirosh et al., 2016)	GEO: GSE72056
ICB treated Melanoma TILs single cell RNA-seq	(Sade-Feldman et al., 2018)	GEO: GSE120575
TCGA bulk RNA-seq	TCGA Research Network	https://www.cancer.gov/tcga
Anti-PD-1 +/- anti-CTLA-4 metastatic melanoma bulk RNA-seq	(Liu et al., 2019)	Supplementary Data 2; https://doi.org/10.1038/s41591-019-0654-5
Murine LCMV Arm + cl 13 T cell microarrays	(Doering et al., 2012)	GEO: GSE41867
Murine LCMV Arm T cell microarray	(Sarkar et al., 2008)	GEO: GSE10239
Murine SV40-TAG TIL bulk RNA-seq	(Philip et al., 2017)	GEO: GSE89307
Murine VACV–SIINFEKL T cell microarray	(Pan et al., 2017)	GEO: GSE79805
Human Yellow Fever vaccine T cell microarray	(Akondy et al., 2017)	GEO: GSE26347
Human Yellow Fever vaccine T cell bulk RNA-seq	(Akondy et al., 2017)	GEO: GSE100745
B16 Dysfunctional CD8+ TILs RNA-seq (murine signatures)	(Singer et al., 2016)	https://doi.org/10.1016/j.cell.2016.08.052
SV40-TAG bulk RNA-seq (murine signatures)	(Scott et al., 2019)	GEO: GSE126973
LCMV cl 13 RNA-seq (murine signatures)	(Im et al., 2016)	GEO: GSE84105
	(Hudson et al., 2019)	https://doi.org/10.1016/j.immuni.2019.11.002
	(Beltra et al., 2020)	https://doi.org/10.1016/j.immuni.2020.04.014
Anergy microarray and RNA-seq (murine signatures)	(Safford et al., 2005)	https://doi.org/10.1038/ni1193
	(Zha et al., 2006)	GEO: GSE5960
	(Provine et al., 2016)	GEO: GSE73001
	(Brignall et al., 2017)	https://doi.org/10.4049/jimmunol.1602033
Human Kidney, Bladder, Prostate Cancer TILs RNA-seq (human TILs signatures)	(Jansen et al., 2019)	GEO: GSE140430
Hepatocellular carcinoma TILs scRNA-seq (human TILs signatures)	(Zheng et al., 2017)	https://doi.org/10.1016/j.cell.2017.05.035
Melanoma TILs scRNA-seq (human TILs signatures)	(Li et al., 2019)	https://doi.org/10.1016/j.cell.2018.11.043
Code for TILs scoring and survival analysis	This paper	https://doi.org/10.5281/zenodo.6419916
Software and algorithms
Picard	Broad Institute	http://broadinstitute.github.io/picard/
Bowtie2	(Langmead et al., 2009)	http://bowtie-bio.sourceforge.net/index.shtml
RSEM v1.2.3	(Li and Dewey, 2011)	https://github.com/deweylab/RSEM
Seurat v3	(Stuart et al., 2019)	https://satijalab.org/seurat/index.html
Monocle 2	(Trapnell et al., 2014; Qiu et al., 2017a; Qiu et al., 2017b)	http://cole-trapnell-lab.github.io/monocle-release/docs/
Bioconductor-GSVA	(Hanzelmann et al., 2013)	https://www.bioconductor.org/packages/release/bioc/html/GSVA.html
nmle	CRAN	https://cran.r-project.org/web/packages/nlme/index.html
clusterProfiler	(Yu et al., 2012)	https://guangchuangyu.github.io/software/clusterProfiler/
ConsensusPathDB	(Cavill et al., 2011)	http://cpdb.molgen.mpg.de/
Survival R package (3.1–8)		https://cran.r-project.org/src/contrib/survival_3.1-8.tar.gz

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Patient samples

Samples from healthy patient skin and melanoma TILs were provided as deidentified with all identifiers removed, and are therefore not classified as Human Subjects research as per the Department of Health and Human Services Title 45, Part 46 Section 4. Normal human skin was obtained from discarded surgical skin specimens. PBMCs were obtained from healthy volunteers. Melanoma tumor specimens, tumor-involved lymph node, and non-involved lymph node from patients with metastatic melanoma were obtained with written informed consent and Institutional Review Board approval (University of Virginia IRB # 10598) to use the samples for research.

METHOD DETAILS

Genomics

T_RM microarray analysis was performed using GSE47045 (Mackay et al., 2013) and GSE39152 (Wakim et al., 2012). Datasets of murine and human T cell differentiation trajectories include: LCMV Armstrong infection (GSE10239 (Sarkar et al., 2008), GSE41867 (Doering et al., 2012)), LCMV cl. 13 infection (GSE41867), autochthonous tumor induction (GSE89307 (Philip et al., 2017)), vaccinia virus infection (GSE79805 (Pan et al., 2017)), and human yellow fever vaccination (GSE26347, GSE100745 (Akondy et al., 2017)). Signatures used to score melanoma TILs were derived from other human TILs datasets of melanoma and non-melanoma cancers, including a. DEGs from TILs described as stem-like vs terminally differentiated and isolated from human prostate, kidney, and bladder cancers (Jansen et al., 2019), b. a gene module expressed by immune checkpoint high TILs isolated from hepatocellular carcinoma (HCC) patients (Zheng et al., 2017), and c. gene modules co-expressed with LAG3 or FGFBP2 used to identify dysfunctional vs cytotoxic melanoma TILs, respectively (Li et al., 2019). Additionally, signatures were curated from murine T cell datasets including a. viral T cell signatures, b. Uncoupled Dysfunction/Activation Gene Modules in CD8⁺ TILS from B16F10 bearing mice (Singer et al., 2016), c. signatures from Early/Plastic (d5–7) vs Late/Fixed (d14–60) dysfunctional (Philip et al., 2017) or WT vs TOX-knockout (Scott et al., 2019) CD8⁺ TILs in inducible, autochthonous murine liver cancer models, d. Signatures derived from various exhausted T cell subsets identified in murine models of LCMV cl. 13 infection: stem-like CXCR5⁺TIM3⁻ vs TD-like CXCR5⁻TIM3⁺ (Im et al., 2016), as well as intermediate populations that have recently been characterized (Hudson et al., 2019; Beltra et al., 2020), and e. signatures from tolerized T cells generated by CD3 or TCR stimulation without costimulation (Safford et al., 2005; Zha et al., 2006), priming without CD4 help (Provine et al., 2016), or ionomycin treatment without PMA (Brignall et al., 2017). See Table S1 for more information. Signatures with GSE numbers listed were derived by calculating DEGs using the package limma based on voom normalized raw counts in R unless otherwise specified. Signatures without GSE numbers listed were derived from DEG lists attached to the corresponding paper.

Cell isolation from human skin and tumor

Normal human skin was obtained from discarded surgical skin specimens. Subcutaneous fat was discarded, and 1 × 1cm² piece of skin was incubated with 1.0mg/ml dispase II (Roche) overnight at 4’C to separate epidermis from dermis. Both epidermis and dermis were cut into small pieces and incubated with 1.0mg/ml collagenase IV (Kim et al., 2015a; Kim et al., 2015b) and 0.1mg/ml DNase I (Roche) in RPMI 1640 medium (Gibco) with 10% FBS (HyClone) for 2hr and followed by enzymatic inhibition with 10mM EDTA (Gibco). Cell suspensions were passed through 70μm cell strainer (BD) and subsequently used for FACS staining. Freshly excised tumor metastases were gently minced into small pieces, and incubated in 2.5mg/ml collagenase D and 1mg/ml DNaseI solution for 30 min at 37°C. Tumor was then passed through a 70μm cell strainer followed by an 80/40% Percoll gradient centrifugation step to obtain total tumor immune infiltrates.

Melanoma TILs preparation and Single cell sequencing

Methods related to single cell sequencing are described elsewhere (Tirosh et al., 2016). Resected tumors were transported in DMEM (ThermoFisher Scientific) on ice immediately after surgical procurement. Tumors were rinsed with PBS (Life Technologies). A small fragment was stored in RNA-protect (Qiagen) for bulk RNA and DNA isolation. Using a scalpel, the remainder of the tumor was minced into tiny cubes <1 mm³ and transferred into a 50 ml conical tube (BD Falcon) containing 10 ml pre-warmed M199-media (ThermoFisher Scientific), 2 mg/ml collagenase P (Roche) and 10U/μl DNase I (Roche). Tumor pieces were digested for 10 minutes at 37°C, then vortexed for 10 seconds and pipetted up and down for 1 minute using pipettes of descending sizes (25 ml, 10 ml and 5 ml). If necessary, this process was repeated twice more until a single-cell suspension was obtained. The cell suspension was then filtered using a 70μm cell strainer (ThermoFisher Scientific) and residual cell clumps were discarded. The suspension was supplemented with 30 ml PBS (Life Technologies) with 2% fetal calf serum (FCS) (Gemini Bioproducts) and immediately placed on ice. After centrifugation at 580 × g, at 4°C for 6 minutes, the supernatant was discarded and the cell pellet was re-suspended in PBS with 2% FCS and placed on ice prior to staining for FACS. Single-cell suspensions were stained with CD45-FITC (VWR) and Calcein-AM (Life Technologies) per manufacturer recommendations. CD45⁺ single cells were sorted into 96-well plates chilled to 4°C, pre-prepared with 10μl TCL buffer (Qiagen) supplemented with 1% beta-mercaptoethanol (lysis buffer) in each well. Single-cell lysates were sealed, vortexed, spun down at 3700 rpm at 4°C for 2 minutes, immediately placed on dry ice and transferred for storage at −80°C. RNA and DNA were isolated using the Qiagen minikit following the manufacturers recommendation.

Whole transcriptome amplification, library preparation, and RNA-seq analysis

Adapted from Tirosh et. al. (Tirosh et al., 2016) Whole Transcriptome Amplification (WTA) was performed with a modified SMART-Seq2 protocol, as described previously (Picelli et al., 2013; Trombetta et al., 2014), with Maxima Reverse Transcriptase (Life Technologies) used in place of Superscript II. WTA products were cleaned with Agencourt XP DNA beads and 70% ethanol (Beckman Coulter) and Illumina sequencing libraries were prepared using Nextera XT (Illumina), as previously described (Trombetta et al., 2014). The 96 samples of a multiwell plate were pooled together, and cleaned with two 0.8× DNA SPRIs (Beckman Coulter). Library quality was assessed with a high sensitivity DNA chip (Agilent) and quantified with a high sensitivity dsDNA Quant Kit (Life Technologies). Samples were sequenced on an Illumina NextSeq 500 instrument using 30bp paired-end reads. Exome sequences were captured using Illumina technology and Exome sequence data processing and analyses were performed using the Picard and Firehose pipelines at the Broad Institute. The Picard pipeline (http://picard.sourceforge.net) was used to produce a BAM file with aligned reads. This includes alignment to the hg19 human reference sequence using the Burrows-Wheeler transform algorithm (Li and Durbin, 2009) and estimation of base quality score and recalibration with the Genome Analysis Toolkit (GATK) (http://www.broadinstitute.org/gatk/) (McKenna et al., 2010). All sample pairs passed the Firehose pipeline including a QC pipeline to test for any tumor/normal and inter-individual contamination as previously described (Berger et al., 2011; Cibulskis et al., 2013).

Following sequencing on the NextSeq, BAM files were converted to merged, demultiplexed FASTQs. Paired-end reads were then mapped to the UCSC hg19 human transcriptome using Bowtie with parameters “-q --phred33-quals -n 1 -e 99999999 -l 25 -I 1 -X 2000 -a -m 15 -S -p 6”, which allows alignment of sequences with single base changes such as due to point mutations. Expression levels of genes were quantified as E_i,j=log₂(TPM_i,j/10+1), where TPM_i,j refers to transcript-per-million (TPM) for gene i in sample j, as calculated by RSEM v1.2.3 in paired-end mode. TPM values were divided by 10 since we estimate the complexity of our single cell libraries to be on the order of 100,000 transcripts and would like to avoid counting each transcript ~10 times, as would be the case with TPM, which may inflate the difference between the expression level of a gene in cells in which the gene is detected and those in which it is not detected. When evaluating the average expression of a population of cells by pooling data across cells (e.g., all cells from a given tumor or cell type) the division by 10 was not required and the average expression was defined E_p(I)=log₂(TPM(I)+1), where I is a set of cells. For each cell, we quantified the number of genes for which at least one read was mapped, and the average expression level of a curated list of housekeeping genes. We then excluded all cells with either fewer than 1,700 detected genes or an average housekeeping expression (E, as defined above) below 3. For the remaining cells, we calculated the pooled expression of each gene as (Ep), and excluded genes with an aggregate expression below 4, which defined a different set of genes in different analyses depending on the subset of cells included. For the remaining cells and genes, we defined relative expression by centering the expression levels, Er_i,j=E_i,j−average[E_i,1...n]. For the single cell scoring of non-exclusive signatures a union of T cell activation sets A and B (day 6 and day 4.5), and a union of memory T cell states A and B were used.

T cell classification and T_RM scoring

T cells were identified based on high expression of CD2 and CD3 (average of CD2, CD3D, CD3E and CD3G, E>4), and were further separated into CD4⁺, Tregs and CD8⁺ T cells based on the expression of CD4, CD25 and FOXP3, and CD8 (average of CD8A and CD8B), respectively. We defined a two signature T_RM high vs. low score (defined as the relative expression of the T_RM-up signature minus that of the T_RM down signature, log2) based on the average expression T_RM specific genes as previously described across mouse lung, skin and gut T_RM (Mackay et al., 2013) and applied across a compendium of CD3⁺ TILs single cell transcripts from 19 human donors. Included human homologues and T_RM high and low genes are annotated in Table S2, and used to generate a heat map sorted by the relative expression of the T_RM score. Specifically, expression data was quantified using RSEM in transcripts per million (TPM), then log-transformed as log2(TPM+1), and then centered for each gene across all cells. The centered data was averaged across sets of genes to define signature scores. We then subtracted a control score from the signature score, which is defined using the same process on randomly selected gene-sets.

Pan Cancer TCGA analysis and Validation set (αPD-1) analysis

TCGA RNAseq level 3 processed data and the corresponding clinical annotations were obtained from the UCSC cancer genome browser (https://genome-cancer.ucsc.edu/proj/site/hgHeatmap/). Gene level expression estimates are reported using RSEM normalized counts and were restricted to only include those sequenced on the Illumina HiSeq platform from the same center, UNC. For each sample the expression level of a signature was measured using ssGSEA in the `gsvà R package. For primary vs metastatic melanoma enrichment, values were Benjamini Hochberg corrected, and a Wilcoxon rank sum was used, to compare the distributions in the box plots (metastatic vs. primary exclusive signatures and metastatic vs. primary non-exclusive signatures). For SKCM (skin cancer melanoma), patients were stratified into high, medium, and low groups using quantiles, with the high group corresponding to the upper quartile, the low group the lower quartile, and medium the remaining two quartiles. All statistical tests for the above data were run using R version 3.1.1. When examining correlation of T cell state signature enrichment in the validation set, the Spearman correlation coefficient was used due to the presence of outliers.

Single cell responder, non-responder analysis to immune checkpoint inhibitors

Processed and normalized TPM counts and patient specific per cell/ sample annotation were downloaded from GEO NCBI for Single cell dataset GSE120575 (www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE120575). Seurat R package was used to further process and analyze the data. Counts matrix was natural log transformed and scaled to create the Seurat data object (Butler et al., 2018). Seurat data object was imported as input for cellular trajectory analysis using the tool Monocle (Trapnell et al., 2014; Qiu et al., 2017a; Qiu et al., 2017b; Zeng et al., 2017). Cell types positive for both CD8a and CD3e molecular markers were pre-filtered for downstream analysis. Pseudotime and relevant states were computed using Monocle’s gene expression tracing to model dynamic biological trajectory trees. When recomputing pseudotime trajectories with cell cycle effect regressed out, we first regressed out cell cycle scores during data scaling of the input Seurat object before importing it for cellular trajectory analysis using Monocle.

DEGs for each major pseudotime cluster relative to the other major clusters (e.g. cluster 1 vs cluster 3 and cluster 5) were calculated using the FindMarkers function in the Seurat package. Gene-name based pathway analysis on cluster DEGs was conducted using online webtool ConcensusPathDB (Cavill et al., 2011). Gene Signature scoring was performed for each cell using AddModuleScore function in the Seurat package. The enrichment scores for select signatures across each cell were used as input for the heat maps. The method of scoring pseudotime cluster signature enrichment during viral timecourses (see Genomics Section for dataset information) was adapted from the method described above (see T cell classification and T_RM scoring section). Viral microarray datasets were transformed such that expression matrices were in the format log2(normalized intensity + 1). Viral RNA-Seq datasets were transformed such that expression matrices were in the format log2(FPKM + 1) and then subsequently filtered on genes with at least one read mapped across all samples. Signature enrichment at a given timepoint was compared to that of the timepoint immediately preceding using mixed-effects modeling, with a random intercept considered for each RNA-seq or microarray sample scored, using the lme function from the nmle package in R. Overrepresentation analysis of pseudotime cluster DEGs in pre-clinical model signatures (see Genomics Section for dataset information) was carried out using the clusterProfiler package in R (Yu et al., 2012).

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical details for experiments can be found in the corresponding figure legend. For indicated experiments, when complex correlation structures were detected within experiments (e.g. TILs were collected from the same patient pre/post therapy), a linear mixed-effects model (MEM) approach was used, accounting for the correlation structure between samples. In general, a random intercept was considered for each experimental unit (e.g. patient, biological replicate), and fixed factors including pseudotime cluster, group, or time, and its interactions. Models were fitted using the lme function from the nmle package in R. Models including heterogeneity of variance and other random effects were also fitted and the model with the minimum Akaike information Criteria (AIC) was chosen as the optimal model. Marginal means were estimated from the optimal model, using the emmeans packages and hypotheses of interest were tested with appropriate contrasts.

Supplementary Material

2

Table S2: List of viral T cell signatures used for scoring. Related to Figures 1–2

3

Table S3: List of major pseudotime cluster DEGs. Related to Figure 3

4

Table S4: Results of pathway analysis of pseudotime cluster DEGs. Related to Figure 4

5

Table S6: List of murine T cell signatures used for scoring. Related to Figure 5

6

Table S7: List of human viral T cell signatures used for scoring. Related to Figure 5

7

Table S8: List of human TILs signatures used for scoring. Related to Figure 6

HIGHLIGHTS.

• Improved global TILs classification methods are required to deconvolve cell states.

• αPD-1 non-responder TILs and dysfunctional TILs score for T activation, not exhaustion

• αPD-1 response and patient survival associates with late T cell memory/T_RM scoring.

• Persistent programs co-express with DC maturation and IFN-γ response programs.

Abbreviations:

TILs

Tumor infiltrating lymphocytes

T_RM

Tissue resident memory T cells

DC

dendritic cells

dLN

draining lymph node

VACV

vaccinia

HSV

herpes simplex virus

LCMV

Lymphocytic choriomeningitis virus

PD-1

programed cell death receptor 1

PD-L1 and PD-L2

programmed death-ligand 1 and 2

αPD-1

anti-PD-1

migDC

migratory DC

OVA

ovalbumin

T_ACT

T activation

T_EM

T effector memory

T_CM

T central memory

T_exh

T exhaustion

scRNA-seq

Single cell sequencing

IFN-γ

interferon gamma