Qualitative Analysis of Tumor-Infiltrating Lymphocytes across Human Tumor Types Reveals a Higher Proportion of Bystander CD8+ T Cells in Non-Melanoma Cancers Compared to Melanoma

Simple Summary Human tumors are often infiltrated by T cells; however, it remains unclear what proportion of T cells infiltrating tumors are bystander and non-tumor specific. We have investigated qualitative characteristics of these tumor-infiltrating lymphocytes (TILs) based on their gene-expression in the tumor-microenvironment or on their response to autologous tumor cells in vitro. Despite a considerable inter-sample variability, we found the overall proportion of bystander (non-tumor reactive) TILs to be remarkably high. Importantly, we observed a higher proportion of bystander TILs in non-melanoma tumors, compared to melanoma. This study suggests that immunotherapeutic strategies, especially when applied to non-melanoma tumors, should be tailored to reinvigorate the small proportion of tumor-reactive T cells infiltrating the tumor-microenvironment. Abstract Background: Human intratumoral T cell infiltrates can be defined by quantitative or qualitative features, such as their ability to recognize autologous tumor antigens. In this study, we reproduced the tumor-T cell interactions of individual patients to determine and compared the qualitative characteristics of intratumoral T cell infiltrates across multiple tumor types. Methods: We employed 187 pairs of unselected tumor-infiltrating lymphocytes (TILs) and autologous tumor cells from patients with melanoma, renal-, ovarian-cancer or sarcoma, and single-cell RNA sequencing data from a pooled cohort of 93 patients with melanoma or epithelial cancers. Measures of TIL quality including the proportion of tumor-reactive CD8+ and CD4+ TILs, and TIL response polyfunctionality were determined. Results: Tumor-specific CD8+ and CD4+ TIL responses were detected in over half of the patients in vitro, and greater CD8+ TIL responses were observed in melanoma, regardless of previous anti-PD-1 treatment, compared to renal cancer, ovarian cancer and sarcoma. The proportion of tumor-reactive CD4+ TILs was on average lower and the differences less pronounced across tumor types. Overall, the proportion of tumor-reactive TILs in vitro was remarkably low, implying a high fraction of TILs to be bystanders, and highly variable within the same tumor type. In situ analyses, based on eight single-cell RNA-sequencing datasets encompassing melanoma and five epithelial cancers types, corroborated the results obtained in vitro. Strikingly, no strong correlation between the proportion of CD8+ and CD4+ tumor-reactive TILs was detected, suggesting the accumulation of these responses in the tumor microenvironment to follow non-overlapping biological pathways. Additionally, no strong correlation between TIL responses and tumor mutational burden (TMB) in melanoma was observed, indicating that TMB was not a major driving force of response. No substantial differences in polyfunctionality across tumor types were observed. Conclusions: These analyses shed light on the functional features defining the quality of TIL infiltrates in cancer. A significant proportion of TILs across tumor types, especially non-melanoma, are bystander T cells. These results highlight the need to develop strategies focused on the tumor-reactive TIL subpopulation.

S4 of S16   . The pie charts illustrate the relative distribution of the seven combinations of three T cell functions generated by tumor-reactive CD8 + (A) and CD4 + (B) T cells in two different cohorts: MM PD-1 naïve and MM PD-1 res. The red, green and blue slices represent cells expressing either three, two or one of the three T cell functions analyzed, respectively. CD8 + and CD4 + TILs in the two MM cohorts had a comparable polyfunctionality (Permutation test, p = 0.9 and p = 0.3, respectively). (A,B) The figure shows a graphical presentation of SPICE data analyses. In these panels, the recognition of Y TILs was tested against TCLs or TCLs+IFNγ, and only the highest value reported. T cells were gated on cells expressing at least one of the three T cell functions analyzed (TNF, IFNγ and CD107a). Median values are shown. S7 of S16   Tables   Table S1. Overview of all samples used in the study (in vitro data). *Actual value used when testing for potential differences in reactivity among pairs tested separately with Y TILs and REP TILs. S12 of S16 *Actual value used when testing for potential differences in reactivity among pairs tested separately with Y TILs and REP TILs. Table S4. Single-cell RNA-sequencing datasets accession numbers (in situ data).

Assessment of TIL Reactivity Against TCLs or FTDs in Vitro
Autologous TILs were thawed and rested overnight in RPMI 1640 (Cat. No 72400-021, Gibco, Thermo Fisher Scientific, Waltham, MA, USA) containing 10% Human serum (H4522, Sigma-Aldrich/Merck KGaA, Darmstadt, Germany) and 1% PenStrep (Cat No 15140122, Gibco, Thermo Fisher Scientific). For the analysis of TIL reactivity against FTDs, single-cell suspensions were thawed, washed and used immediately after a trypan blue viability count. Tumor-specific immune activation was assessed with a 5 to 8-hour co-culture assay of TILs and autologous TCLs or FTDs (effector/target ratio of 3:1) in the presence of anti-CD107a antibody and Brefeldin A (dilution of 1:1000, GolgiPlug™, Cat No 555029, BD Biosciences, San Jose, CA, USA) or a combination of Brefeldin A and Monensin (dilution of 1:1000, GolgiStop™, Cat No 554724, BD). Intracellular staining of TILs and acquisition with a flow cytometer were carried out using standard methods. Boolean gating was performed to obtain the proportion of cells positive for at least one marker among TNF, IFNγ and CD107a. A specific antitumor response was defined as the detection of a response greater than twice the background (i.e., TILs alone), with a difference of > 0.5% from the background and a minimum of 50 positive flow cytometry events after subtraction of the control. A value of 0.5% was used as the limit of sensitivity. Unspecific T cell activation was ruled out by co-culturing selected TILs with a panel of allogeneic melanoma TCLs in our institution's cell line bank, with no upregulation of functional markers detected when co-culturing with at least one TCL.

Antibodies
The following antibodies were used for the flow cytometry assays

Characterization of TIL Polyfunctionality
For polyfunctional characterization of tumor-reactive TILs, flow data were primarily analyzed in FlowJo V10 (BD). Lymphocytes were selected based on a plot of FSC-A vs SSC-A. Doublets were removed by gating FSC-A vs FSC-H. Subsequently, cells negative for Live/Dead Fixable Dead Cell Stain Near-IR (NIR) were gated as live cells and CD3 + TILs were gated in a plot of CD3 vs CD56. CD4 + and CD8 + populations were identified in a CD4 vs CD8 plot. Finally, TNF + , IFNγ + and CD107a + subpopulations, within the CD4 and CD8 compartments, were gated in a plot of TNF or IFNγ or CD107a versus CD4 or CD8, respectively. Subsequently, Boolean combination gates were made for the three functional markers (TNF, IFNγ, and CD107a), generating seven gates of tumor-reactive TILs, each showing the percentage of CD8 + or CD4 + TILs expressing a unique combination of the three markers (TNF + /IFNγ + /CD107a + , TNF + /IFNγ + /CD107a -, TNF + /IFNγ -/CD107a + , TNF -/IFNγ + /CD107a + , TNF + /IFNγ -/CD107a -, TNF -/IFNγ + /CD107a -, TNF -/IFNγ -/CD107a + ) and one gate of not reactive TILs (TNF -/IFNγ -/CD107a -). Only the data regarding the seven gates of tumor-reactive TILs were used for the following analyses and exported into Pestle 2.0, where they were formatted according to manufacturer's instructions, and the background was subtracted. In SPICE, thresholds were set at 0.1. Comparison of pie charts was performed using a partial permutation test followed by comparison of bar charts using a Wilcoxon rank sum test in case of significant difference between the respective pies, as described previously [9]. All values calculated in SPICE were expressed as median unless otherwise specified. Three to 12 samples were selected for these analyses from each tumor type according to the following criteria: 1) Only samples with a proportion of tumor-reactive T cells > 0.5% were selected; 2) For each reported patient, only the sample with highest reactivity obtained from Y TILs tested against TCLs or TCLs + IFNγ was utilized; 3) Within each tumor type, samples were selected along all the range of reactivity, with an equal representation of samples with high (>66 percentile), intermediate (33-66 percentile) or low (<33 percentile) reactivity.

Processing of TCGA Data
The Cancer Genome Atlas (TCGA) sample information were downloaded via the TCGAbiolinks R package [10][11][12] . Cancer types data were retrieved via the TCGAbiolinks R package to generate the sub-categories of Skin cutaneous melanoma (SKCM), Kidney renal clear cell carcinoma (KIRC), Ovarian serous cystadenocarcinoma (OV) and Sarcoma (SARC). TCGA study abbreviations used in this study are reported at https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables/tcga-studyabbreviations. The samples information table was filtered to remove Formalin-Fixed Paraffin-Embedded (FFPE) samples and replicates. Samples labeled as Metastatic (for SKCM) or as Primary Solid Tumor (for KIRC and SARC) or any tumor (for OV) and as RNA-Seq were selected. The TCGA HTSeq-Counts RNA-Seq data were downloaded using the TCGA biolinks R package and normalized using the DESeq2 R package, with the variance stabilizing transformation vst() function. A gene list of 17 genes (B2M, TAP1, TAPBPL, CALR, PSMB9, PSMB10, ERAP1, PDIA3, NLRC5, RFX5, PSME1, PSME2, PSME3, CIITA, HSP90AB1, HSP90AA1, and HSP90B1) that reflected antigen processing and presentation machinery (APM) was obtained from Thompson et al. [13].The vst-normalized count data were used for the calculation of an APM score based on the mean normalized expression of the above mentioned 17 genes for each selected sample. TCR richness data for the selected samples were obtained from Thorsson et al. [14] (Table S1) and calculated from RNA-seq datasets as previously described [14]. Plots and statistical tests for the TCGA data were produced using the ggstatsplot R package [15].

T cell Transcriptomics Single-Cell Data from Public Domain
One dataset [16] was downloaded but not used for further analysis due to a significantly lower average number of detected genes per cell. Of the twelve colorectal cancer patients from Zhang et al [5], four were excluded due to the microsatellite unstable (MSI) status of their disease. The exceedingly high TMB of these samples could bias the analysis due to the high disproportionality of tumor-reactive T cells (this condition is normally found in only 15% of all patients with colorectal cancer [17]. Therefore, only microsatellite stable (MSS) colorectal cancer samples were considered for the analyses in this study. Additionally, only 4 tumor samples from adult patients (1 papillary renal cell carcinoma and 3 clear cell renal cell carcinomas) were included from a kidney cancer dataset [18] . The melanoma datasets [6][7][8] included samples from both anti-PD-1 therapy naïve and previouslytreated patients. Treatment with anti-PD-1 is known to induce early infiltration of intratumoral T cell, especially in connection with an objective response [19], however we did not include any tumor biopsy from regressing lesions under treatment with anti-PD-1 in the in vitro study. Hence, samples from patients whose lesions regressed after anti-PD-1 therapy were excluded from the in situ analyses to maintain consistency with our in vitro data. The remaining samples (n = 67) were classified as MM PD-1 naïve (n = 39) or MM PD-1 res (n = 28). Both MM PD-1 naïve and MM PD-1 res biopsies were available for eight patients, but only the biopsies collected at the earlier time point (MM PD-1 naïve) were considered for these analyses (in total 59 samples from an equal number of individual patients). For those patients where multiple MM PD-1 res biopsies were available, the one collected at the latest time point was chosen. For CD4 + TILs, only samples from MM PD-1 naïve patients (n = 39) were considered (the proportion of tumor-reactive TILs was significantly higher in MM PD-1-res samples, therefore the two groups could not be merged), whereas both MM PD-1 naïve and MM PD-1 res samples (n = 59) were considered for CD8 + TILs (no difference in the proportion of tumor-reactive TILs between MM PD-1 naïve and MM PD-1 res samples) ( Figure S5A and Figure S10). For patient Mel129 from Jerby-Arnon et al [7], data from two biopsies collected at the same time point were available, and the mean values of the two were used. Patient Mel78 from Tirosh et al [6] was excluded S15 of S16 from the analyses due to the extremely low number of T cells isolated from the tumor tissue, whereas, no CD4 + T cells could be identified from patient P1207 from Zhang et al. after data processing [5]. As different gene IDs were used by the various authors who published these datasets, we used the HUGO Gene Nomenclature Committee (https://www.genenames.org/download/custom/, accessed on 21/11/2019) to convert gene IDs to a NCBI GeneID in each dataset. Only T cells isolated from tumor tissues were utilized. Additionally, only those genes expressed in at least three cells and only those cells expressing at least 500 genes were considered for these analyses. Normalization of the read counts was performed with the SCTransform function in the Seurat R package. Identification of CD4 + and CD8 + T cells was based on CD4, CD8A and CD8B expression and on the following criteria: CD4 T cells = CD4+ AND CD8A-AND CD8B-CD8 T cells = CD4-AND (CD8A+ OR CD8B+) Cells were considered "positive" or "negative" for the expression of a specific gene according to a manually set threshold based on a bimodal distribution of the gene expression across the different datasets.