Genomic Correlates of Outcome in Tumor-Infiltrating Lymphocyte Therapy for Metastatic Melanoma

Abstract

Purpose:

Adoptive cell therapy (ACT) of tumor-infiltrating lymphocytes (TIL) historically yields a 40-50% response rate in metastatic melanoma. However, the determinants of outcome are largely unknown.

Experimental Design:

We investigated tumor-based genomic correlates of overall survival (OS), progression-free survival (PFS), and response to therapy by interrogating tumor samples initially collected to generate TIL infusion products.

Results:

Whole exome sequencing (WES) data from 64 samples indicated a positive correlation between neoantigen load and OS, but not PFS or response to therapy. RNA sequencing analysis of 34 samples showed that expression of PDE1C, RTKN2, and NGFR were enriched in responders who had improved PFS and OS. In contrast, the expression of ELFN1 was enriched in patients with unfavorable response, poor PFS and OS, whereas enhanced methylation of ELFN1 was observed in patients with favorable outcomes. Expression of ELFN1, NGFR and PDE1C was mainly found in cancer-associated fibroblasts and endothelial cells in tumor tissues across different cancer types in publicly available single cell RNA sequencing datasets, suggesting a role for elements of the tumor microenvironment in defining the outcome of TIL therapy.

Conclusions:

Our findings suggest that transcriptional features of melanomas correlate with outcomes after TIL therapy and may provide candidates to guide patient selection.

INTRODUCTION

The treatment landscape for metastatic melanoma has been revolutionized by immunotherapies. Combination immune checkpoint blockade (ICB) approaches achieve objective response rates of 40-60% (1,2). Historically, cell-based immunotherapy using tumor-infiltrating lymphocytes (TIL) yields similar response rates as the antibody-based immunotherapies but has had more difficulties reaching the mainstream because of the challenges of manufacturing approaches and the toxicities related to IL-2 administration, limiting this treatment to a few centers worldwide. Nonetheless, adoptive cell therapy (ACT) of TIL, has led to durable responses, even in patients refractory to checkpoint inhibitors but to a lesser extent (3–7). However, recent studies have shown promising results with a commercial TIL product in patients with metastatic melanoma and in patients with cervical cancer, supporting the potential for more widespread application of TIL therapy in the future. Therefore, there is a need to find predictive biomarkers that allows to stratify patients who would clinically benefit from TIL therapy.

Previously, our group reported TIL attributes that correlated with clinical responses in metastatic melanoma patients treated with TIL therapy (6,8). Higher number of TIL infused, and higher proportion of CD8⁺ TIL and CD8⁺ BTLA⁺ (B and T lymphocyte attenuator) TIL in the infusion product correlated with improved clinical response (8). While these features raise hypotheses about strategies to improve the efficacy of TIL, they are not helpful for selecting patients for this treatment. We also identified important clinical features, including low serum LDH levels and fewer prior systemic therapies, at the time of treatment as predictors of improved response to TIL therapy however these are general markers of improved therapeutic response and are not unique to TIL therapy (6,9).

There is a strong rationale to evaluate tumor features for associations with TIL outcomes. In addition to improving patient selection, the identification of molecules or pathways associated with resistance could suggest rational combinatorial approaches to improve outcomes. In a small cohort of patients treated with TIL therapy, tumor mutation load was associated with favorable outcome (9).

In this study, we aim to identify genomic and transcriptomic factors that correlate to outcomes in melanoma patients receiving TIL therapy. We performed whole exome sequencing (WES) and RNA sequencing (RNA-Seq) of tumor tissues used to generate the TIL product for 64 metastatic melanoma patients before they were treated with TIL therapy as part of clinical trials at MD Anderson Cancer Center (MDACC, n=55) or at the H. Lee Moffitt Cancer Center and Research Institute (MCC, n=9). Our integrated analysis of the molecular features of the tumor tissues with patient response, progression-free survival (PFS), and overall survival (OS) has identified new candidate predictors of clinical outcomes for TIL therapy.

METHODS

Clinical cohort development and study design

All tumor tissues utilized for this study were obtained from stage III or IV metastatic melanoma patients enrolled on the TIL ACT clinical trial MDACC (Institutional Review Board approved protocol #2004-0069, NCT00338377) and MCC (Institutional Review Board protocol, #18309, NCT01005745, NCT01701674, and NCT01659151). All studies were approved by the United States Food and Drug Administration and the Institutional Review Board of each independent center. The studies were written and conducted in accordance with the principles from the Declaration of Helsinki and all the patients signed written informed consent prior to treatment initiation. Tumor samples were obtained from 73 patients treated at MDACC and 10 patients treated at MCC. Patients from MCC were recruited in three different trials: TIL ACT (n = 5), TIL + ipilimumab (n = 2), and TIL + BRAF inhibitor (n = 3). At MDACC, patients were enrolled in 3 cohorts of this trial: TIL ACT (n = 65), TIL + BRAF inhibitor (n = 3), and TIL with or without dendritic cell (DC) infusion (n = 5).

For all trials, patients were lymphodepleted 7 days prior to intravenous infusion of autologous TIL. TIL were expanded from 1-3 mm³ tumor fragments in RPMI based media with the addition of 6,000 IU/mL of IL-2 every 2-3 days for 3-5 weeks as previously described(8,10). Additionally, patients received high-dose IL-2 (720,000 IU/kg) every 8 hours beginning one day after infusion for a maximum of 15 doses at MDACC or over 5 days at MCC. Patients treated at MD Anderson received a second course of high dose IL-2 three weeks later. Further information regarding trial design are described in Forget, Haymaker et al. 2018, Pilon-Thomas, et al. 2012, Mullinax, et al. 2018, and Atay et al., 2019 respectively (6,7,10,11).

At time of tumor harvest for TIL generation, a portion of the tumor was preserved as a Formalin-Fixed, Paraffin-embedded (FFPE) block for histopathological diagnosis. For the MDACC and MCC treated patients, blood samples collected at time of surgery for TIL harvest or, prior to lymphodepletion before TIL infusion (baseline) were processed as peripheral blood mononuclear cells (PBMCs). PBMCs were preserved by the MDACC Imunomonitoring Core Lab and MDACC Melcore Lab and by the Pilon-Thomas laboratory at MCC. Response to TIL therapy was measured using Immune-Related Response Criteria (irRC) at MDACC and RECIST v1.1 at MCC. Responders are defined as patients with complete response (CR) or partial response (PR), while patients with stable disease (SD) or progressive disease (PD) are considered non-responders. Metadata for patients enrolled in the trials are presented in Tables 1 and 2 and Supplementary Tables S1 and S2 and associated patient specific data is reported as Supplementary Tables S3–S6.

Table 1.

Patient Metadata for samples used for WES

Number of patients		Total	Complete Responder	Partial Responder	Stable Disease	Progressive Disease
Age
	<30	7	0	4	2	1
	31-40	13	1	2	4	6
	41-50	18	4	5	6	3
	51.60	18	2	3	8	5
	>61	8	1	2	4	1
Gender
	Male	37	5	6	14	12
	Female	27	3	10	10	4
Stage
	IIIC	4	2	1	1	0
	M1a	1	0	0	0	1
	M1b	10		2	5	2
	M1c	49	5	13	18	13
Type of Melanoma
	Cutaneous	45	7	9	15	14
	Acral	5	0	1	4	0
	Mucosal	3	0	2	1	0
	Unknown Primary	3	0	1	2	0
	Undetermined	8	1	3	2	2
Prior Systemic Therapies
	0	6	1	2	0	3
	1	19	3	3	9	4
	2	14	2	5	7	0
	3	11	2	3	4	2
	4+	14	0	3	4	7
Prior Progression on:
	Anti-PD-1	8	0	2	4	2
	Ipilimumab	23	3	5	10	5
	High Dose IL-2	21	4	4	6	7
BRAF inhibitors (not sorafenib)		6	0	1	3	2
	Biochemotherapy	25	1	7	10	7
Brain Metastases at time of treatment
	Yes	18	0	6	7	5
	No	46	8	10	17	11
LDH at TIL Infusion
	Elevated	18	0	4	10	4
	Not Elevated	46	8	12	14	12

Table 2.

Patient Metadata for samples used for RNA Sequencing

	Number of patients	Total	Complete Responder	Partial Responder	Stable Disease	Progressive Disease
Age
	<30	5	0	2	2	1
	31-40	4	0	0	0	2
	41-50	8	1	2	2	0
	51-60	12	1	1	1	5
	>61	5	0	2	2	0
Gender
	Male	25	2	3	13	7
	Female	9	0	4	4	1
Stage
	IIIC	1	0	0	1	0
	M1a	1	0	0	0	1
	M1b	4	0	2	2	0
	M1c	28	2	5	14	7
Type of Melanoma
	Cutaneous	28	2	5	13	8
	Acral	1	0	0	1	0
	Mucosal	0	0	0	0	0
	Unknown Primary	1	0	0	1	0
	Undetermined	4	0	2	2	0
Prior Systemic Therapies
	0	1	0	1	0	0
	1	12	0	2	7	3
	2	9	1	1	6	1
	3	5	1	1	2	1
	4+	7	0	2	2	3
Prior Progression on:
	Anti-PD-1	6	0	1	3	2
	Ipilimumab	15	1	2	8	4
	High Dose IL-2	11	1	3	4	3
BRAF inhibitors (not sorafenib)		6	0	1	3	2
	Biochemotherapy	14	1	4	5	4
Brain Metastases at time of treatment
	Yes	11	0	3	5	3
	No	23	2	4	12	5
LDH at TIL infusion
	Elevated	11	0	0	9	2
	Not Elevated	23	2	7	8	6

DNA and RNA extractions

FFPE specimens were sequentially cut in 10 μm sections. The initial 2 shavings were discarded, and the following 7-12 shavings were placed in an RNase free microcentrifuge tube. Samples were immediately placed at −80 °C until extraction was performed. DNA and RNA were extracted utilizing the Qiagen AllPrep DNA/RNA FFPE Kit (#80234) with the Qiagen deparaffinization solution utilized (#19093). After extraction, DNA and RNA underwent melanin removal utilizing the Zymo Research OneStep PCR Inhibitor Removal Kit (#D6030). PBMCs acquired prior to treatment were used as germline controls and extracted for DNA using the Qiagen AllPrep DNA/RNA Mini Kit (#80204). Samples were quantified for the 260/280 ratio, RIN, and DV200 with a Nanodrop 1000.

Whole Exome Sequencing (WES)

Whole exome sequencing was performed at the Broad Institute on DNA from FFPE tissue of 83 samples. Matched PBMC DNA was sequenced as germline controls and fingerprinting confirmed patient sample concordance. The library construction and hybrid capture, preparation of libraries for cluster amplification and sequencing, and cluster amplification and sequencing were performed as outlined in Forget and Haymaker et al (6). Sequencing reads were aligned to hg19 using BWA (12). Five samples were removed due to less than 5x mean exon coverage. Cross-contamination of samples was estimated using ContEst (13). Four samples were removed due to possible contamination.

WES analysis of genetic alterations

Tumor Purity, SNVs, and Indels

ABSOLUTE was used to call tumor purity and the cancer cell fraction (CCF) of mutations (14). Ten samples had poor ABSOLUTE solution and were manually determined to have low purity (less than 20%). These samples were excluded, leaving 64 samples used for analysis. Single-nucleotide variants (SNVs) calling was performed with MuTect and FFPE artifact filtering was performed as outlined previously (6). Short insertions and deletions (indels) were called using Strelka (15). The SNVs and indels were filtered to remove artifacts, based on allelic fractions observed in panels of normal samples.

Mutations and Neoantigen prediction

Annotating mutations, calculating the nonsynonymous mutations, and calling mutated genes were performed as outlined in Forget and Haymaker et al (6). The HLA class I types of each patient were called with OptiType (16). Selected mutations were confirmed using remaining DNA after WES for Sanger sequencing. The region of interest was amplified using custom PCR primers. Sanger sequencing was performed on a 3730xl DNA Analyzer (Thermo Fisher/Applied Biosystems) using BigDye™ Terminator v3 chemistry (Thermo Fisher/Applied Biosystems). Mutation analysis was performed using SeqScape® Software v2.5 (Thermo Fisher/Applied Biosystems). Neoantigens were predicted using NetMHCpan 4.0 and default parameters (17). This method integrates both binding affinity and mass spectrometry eluted ligand data. The predicted neoantigen load was determined by an eluted ligand likelihood (ELL) in the top 2%.

RNA profiling and analysis

Quality Control

550 ng of mRNA regardless of dV200 was used for sample preparation and strand-specific cDNA synthesis at the Broad Institute (Illumina TruSeq RNA Access Kit) for 53 pre-treatment samples, of which 34 met the quality-control threshold of having at least 12,500 genes detected. A transcriptome capture approach that targets 21,415 genes, representing 98.3% RefSeq exome, was used to enrich for the mRNA. RNA-Seq of 76 bp reads was performed on Illumina machines in two batches. Fastq files were aligned using STAR, which also provided raw read counts for each gene (18).

Gene expression profiling using droplet digital PCR (ddPCR) assay

One μg of RNA was used from 16 FFPE tissue samples (same RNA preparation as for RNA-seq, when available) and from 2 positive and 2 negative control cell line samples presenting either high (U20S ATCC #HTB-96, MCF7 ATCC #HTB-22) or low expression of ELFN1 (Jurkat ATCC TIB-153, HEK293 ATCC #CRL-1573), was used for cDNA synthesis using the High-capacity cDNA Reverse Transcription kit (without RNase inhibitor; cat. #4368814, Thermo Fisher Scientific). The selection of control samples was performed based on The Protein Atlas gene expression, public available datasets (http://www.proteinatlas.org, Nov 04 2020) . Next, the obtained cDNA was used to set up the ddPCR reaction following the manufacturer’s instructions of the QX200 Droplet Digital PCR System using supermix for probes (no dUTP) (cat. #1863023; Bio-Rad). Each sample was run in duplicates and was run with non-template control. Briefly, the ddPCR assay was performed in 20 μL reactions containing 2x ddPCR supermix (no dUTP, Bio-Rad), two set of IDT 20x PrimeTime qPCR Assay with primers and Taqman probes (cat. #Hs.PT.58.14371910, cat. #Hs.PT.58v.45621572; IDT) at a 3.6:1 ratio (final concentrations of 900 nM and 250 nM, respectively), targeting ELFN1 gene (labeled with FAM) and HPRT1 (labeled with HEX) - the house keeping gene, and cDNA (50 ng of control cell line samples and 250 ng of FFPE tissue samples) or PCR grade water. The PCR reaction mixture with the addition of droplet generation oil for probes (Bio-Rad) was portioned into droplets using the QX200 Droplet Generator (Bio-Rad) according to the manufacturer’s protocol. The droplets were carefully collected and next PCR amplified in C1000 Thermal Cycler (Bio-Rad) using following optimized conditions: 1 cycle at 95°C for 10 min, 40 cycles of 94°C for 1 min and 57.1°C for 1 min, 1 cycle of 98°C for 10 min, and a 12°C hold. Fluorescence signal was read in the FAM and HEX channels in QX200 droplet reader (Bio-Rad). Primer and probe sequences for target and reference genes are provided in Supplementary Table S7. The ddPCR data analysis was performed using QuantaSoft™ Analysis Pro Software (version 1.0.596; Bio-Rad), with positive and negative droplet populations detected using two-dimensional graphs. The gene expression was calculated from generated Poisson concentrations (copies/μL) and the data is presented as a ratio of ELFN1 gene to housekeeping gene and graphically displayed using GraphPad Prism Software version 8.0.0. Examples of this analysis can be found in Supplementary Fig. S1.

Methylation array

DNA (500 ng) from 30 FFPE tissue blocks at MDACC was utilized for methylation analysis at the John Wayne Cancer Institute & USC Epigenome Center as previously described (19). The Infinium HumanMethylation450 BeadChip was used to profile the methylation level of single-CpG–sites for 30 samples. After preprocessing the data, the calculated betta-values of M sites for each sample were converted to M-values to avoid heteroscedasticity for highly methylated or unmethylated CpG sites (20,21). The M-values of 50 M sites for 20 responders (CR/PR) and 10 non-responders (PD) were then used for unsupervised hierarchical clustering of the samples in terms of their methylation profiles ordered by Msites location within the ELFN1 gene. Before the clustering, the M-values were centered and normalized by rows and columns. Hierarchical clustering was implemented by Cluster 3.0 software using the centroid linkage algorithm and the Pearson correlation as the similarity measure. Visualization was done by Java TreeView 3.0. The mean levels of methylation for responders and non-responders were calculated using normalized and centered M-values and visualized as boxplots using ‘beeswarm’ R library.

To establish the difference in methylation status between responders (PR/CR) and non-responders (PD), differentially methylated regions (DMRs) between PD and CR/PR tumors in each gene/promoter presented on the array were identified using mGSEA, an R package that implements a Gene Set Enrichment Analysis method to identify DMRs. Genes with significantly differentially methylated regions (p adj < 0.05) were further filtered using the cancer gene census list downloaded from COSMIC (15Apr2021). Overlap of the filtered genes with gene ontology gene sets was implemented using the Molecular Signatures Database (MSigDB) (v7.4).

Statistical Analysis

Statistical analyses were performed in R 3.5. For correlations of non-silent mutation load and neoantigen load with OS and PFS, the log10 (mutation or neoantigen load) was used in a Cox proportional hazards model, to preserve the quantitative values of mutation and neoantigen load. Maftools was used for the sample-by-gene mutation plot (22). For the analysis of mutations in specific genes, each gene was classified as having at least one non-silent mutation or not in each patient, and the binary results were used in Mann-Whitney tests for responder vs non-responder comparisons or in Cox proportional hazards models for correlations with OS or PFS.

For the transcriptomic analysis, samples with fewer than 12,500 detected genes were excluded for quality control, leaving 34 samples for subsequent analyses. The raw read counts for each gene were fed into DESeq2 for differential expression analysis (23). Only protein-coding genes were analyzed. Genes with zero read counts in 75% or more of the samples were excluded to prevent outliers from influencing results. This removed 2290, or 12.4%, of the protein-coding genes in the 34 pre-treatment samples. We used log2 (OS +1),log2 (PFS +1), and response as the variables of interest in DESeq2, and batch effects were controlled by introducing a batch covariate into the model (23). As shown in the logarithmic formula above, OS and PFS are used as numerical values to find their correlation with gene expression and no cut-offs in survival are used.

For gene set enrichment analysis (GSEA), we took DESeq2-normalized counts of protein-coding genes filtered as described in the above section and applied ComBat to control for batch effects (24). GSEA was run using the graphical user interface with phenotype permutations and p = 1 weighting. Signal2Noise was used as the gene ranking metric for dichotomous comparisons of responders vs non-responders. Pearson correlation was used as the gene ranking metric to assess correlations between gene expression and continuous survival data, which were transformed as log (OS + 1) and log (PFS + 1).

For analysis of genes associated with TIL features of interest (number of TIL infused, number of CD8⁺ TIL infused, fraction of CD8⁺ TIL), we selectively examine the four genes (ELFN1, NGFR, RTKN2, and PDE1C) whose expression correlated with all three outcome metrics. For each of the TIL features, we ran DESeq2, using the RNA sequencing batch as a covariate, and selectively examined the four genes. P-values were not adjusted for multiple comparisons.

For analysis of ddPCR and its correlation with RNA-seq, we used the Pearson’s correlation and Mann-Whitney t-test in GraphPad Prism version 8.0.0. The statistical differences in the mean methylation levels were evaluated by Wilcoxon tests in R.

Data availability

TCGA data referenced within this work available at the Cancer Genomics hub website (http://cghub.edu). The whole exome sequencing, RNA sequencing, and methylation data sets referenced in this work are found as supplementary data. Further supporting data is available upon request to the corresponding author.

RESULTS

Study design

To investigate the association between the genetic and transcriptomic makeup of the tumor and clinical outcome in TIL therapy, we evaluated the baseline melanoma tumor tissue samples from patients treated at MDACC and MCC (Table 1). Following surgical resection, tumor samples were first utilized to grow TIL for clinical use. The leftover tissues were used to generate genome-wide tumor mutational burden and DNA methylation profiles in combination with transcriptomic profiles (Fig. 1). Characterization of the TIL infusion product and detailed clinical benefit following TIL therapy were previously reported (6,8,10,11). Here we describe the results of our in-depth molecular characterization of the tumor samples collected from melanoma patients before receiving TIL therapy and identify potential biomarkers of survival outcomes and response to TIL therapy.

Figure 1. Resected tumor workflow on the metastatic melanoma ACT of TIL study.

Schematic depicting the distribution of the metastatic melanoma tumor tissue after surgical resection at MDACC (n=64) and MCC (n=9). Numbers in circles indicate initial sample prioritization and distribution for 1) generation of the clinical TIL product for patient treatment and 2) formalin-fixed, paraffin embedded (FFPE) samples. The 73 patients with generated clinical TIL product, thus had 73 corresponding FFPE samples. DNA was extracted from all 73 FFPE samples, 64 samples passed quality control (QC) for mutation profiling via WES and 30 samples were also used for DNA methylation profiling (7 samples passed DNA methylation QC but did not pass WES QC). Extracted RNA from 34 patients passed QC for transcriptomic profiling for RNA-Seq, with remaining RNA or subsequently extracted RNA utilized for ddPCR from 16 samples.

Mutational profile of TIL treated patients is similar to melanoma TCGA with the exception of an enrichment for mutated KCNQ2 and SFTA3

Tumor mutation burden (TMB) and mutations in specific genes have been shown to be associated with outcomes in patients receiving ICB, in several tumor types (25–28). Thus, we compared the results of WES of pre-treatment tumor samples with clinical outcomes in 64 metastatic melanoma patients treated with TIL therapy. We first compared our results to The Cancer Genome Atlas (TCGA) study of cutaneous melanoma (29). We found that the mutations detected in both cohorts were similar overall. The median number of mutations in our cohort, including single-nucleotide variants (SNVs) and short insertions and deletions (indels), was 455.5 (range 21–1941, Fig. 2A), which is comparable to what it was previously reported in the in TCGA in 333 melanoma samples, indicating we had sufficient sensitivity to detect the vast majority of mutations in these samples, which gave us confidence that the bioinformatic approach employed to correct the bias from the formalin fixation of the tissue was adequate (29). Most of the SNVs detected were C>T transitions, and the overall mutational signature was highly consistent with UV-induced mutations found in melanoma as defined by COSMIC Signature 7 (Fig. 2B) (30). Non-negative factorization categorized the SNVs in our cohort into four separate SNV mutation signatures, with the most predominant signature (S4) closely resembling COSMIC Signature 7 (Supplementary Fig. S2) (30). The close semblance of the overall mutation signature to UV-induced signature suggests a very low false-positive rate in our mutation calls.

Figure 2. Mutation analysis of pre-treatment tumors from TIL treated patients identifies enriched recurrently mutated genes in this cohort.

A, Mutation load in the TCGA cutaneous melanoma cohort and our combined TIL cohort from MDACC and MCC, consisting of the pre-treatment tumor of the 64 patients treated with TIL. Red line identifies median number of mutations. B, The upper panel shows the mutation signature of SNVs in the TIL patient cohort (n=64), while the lower panel depicts the signature of UV-induced mutations as described by COSMIC Signature 7. The two signatures have a cosine similarity of 0.99. C The MutSig2 plot of the recurrent mutations (left side) found within the sample set (q< 0.1 is significantly enriched) and the frequency (in percentage) of the mutations (right side). The type of mutation, response to therapy, and cancer center of origin for each sample are denoted below the main plot.

In addition to UV-induced mutations, driver mutations such as BRAFV600E and NRASQ61R have been well characterized in melanoma regarding their impact on tumor development, frequency, and therapeutic benefit. In our patient cohort, we identified six recurrent mutated genes (Fig. 2C), including BRAF, NRAS, CDKN2A, TP53, KCNQ2, and SFTA3. Comparing our cohort to previously published datasets, we find that BRAF, TP53, and CDKN2A were mutated at similar rates, within 10% of what was previously reported [52%, 27%, and 11%, respectively (Fig. 2C)] (29,31). However, NRAS was less frequently mutated in our cohort (14% vs. 26–28%), while HRAS and KRAS were mutated at similar rates [3% and 2%, respectively (Fig. 2C)] (29,31). The complete list of mutations per patient can be found in Supplementary Table S8. We did not identify genes whose mutation was correlated with response, OS, or PFS.

We identified recurrent mutations in KCNQ2 and SFTA3 (MutSigCV q = 0.03 and 0.04, respectively) which were more frequent than reported in TCGA (12% vs. 1.9% for KCNQ2, 6% vs. 0.6% for SFTA3, Fig. 2C)(29). KCNQ2 had missense mutations in 6 patients and splice site mutations in 2 patients, SFTA3 had missense mutations in 4 patients, with a hotspot (in 3 of the 4 patients) with C to T transitions at chr14:36946289, leading to Glu50Lys. Sanger sequencing confirmed the presence of these mutations (Supplementary Fig. S3; Supplementary Table S9). Notably, TIL therapy can only be offered if TIL are adequately expanded from the tumor samples, which was the case for 60-70% of the patients accrued on these trials. Thus, treated patients have undergone a selection process that may influence the mutational profiles observed and may explain the differences compared to TCGA (32).

Mutational burden fails to correlate with response and progression-free survival while predicted neoantigen load correlates with overall survival

Previous studies reported a positive correlation between high tumor mutation burden and immunotherapeutic approaches. Also, previous reports suggested an association of high tumor mutation burden with PFS and OS in TIL treated patients (25). In our cohort, we observed a positive trend between longer OS and greater non-silent mutation load (p = 0.07 for OS, Fig. 3A), however no correlation was observed with PFS or response (p = 0.14, p = 0.77, Supplementary Fig. S4A and B). Of note, sample purity did not correlate with any outcomes, indicating there was no bias in the proportion of tumor cells within the samples between the groups (Supplementary Fig. S5).

Figure 3. Non-silent mutation burden and predicted HLA class I neoantigen load associate with overall survival.

The three graphs represent pre-treatment tumor samples from 64 TIL treated patients. A, Graph showing non-silent mutation load and OS of patients color-coded by their response to TIL therapy. The p-value was generated using the Cox proportional hazards model. B, Predicted neoantigen load versus OS color-coded by patient response status. The dotted lines in both plots indicates ongoing survival. Statistical differences were determined by Cox proportional hazards model.

Tumor mutation burden correlates with the generation of neoantigens, which can induce a powerful T-cell mediated immune response leading to improved tumor clearance in patients . The neoantigen load can be calculated by predicting the ability of the specific HLA molecules expressed by the patient’s tumor cells to bind peptides derived from mutated antigens. As expected, we observed a strong correlation between predicted HLA class I neoantigen load and non-silent mutation burden (Pearson’s r = 0.99) (Supplementary Fig. S6A). We detected an association between OS and high tumor neoantigen load (p = 0.042, Fig. 3B). However, there was no association of neoantigen load with PFS (p = 0.12) and clinical response (p = 0.78; Supplementary Fig. S6B and C). Of note, several patients with partial or complete response to therapy have ongoing progression-free survival, suggesting the possibility that the p value could change after a longer follow-up (Supplementary Fig. S4A and S6B).

Nonetheless, Figure 3B and S6B show that the small population of patients with fewer than 200 predicted neoantigens uniformly have poor outcome (SD or PD to TIL therapy with short OS and PFS survival). This data suggests that a very low tumor burden leads to poor outcome however the reverse was not true. Overall, many patients with high mutation load or neoantigen load did not have favorable survival or response to therapy, prompting the need for further investigation into other biomarkers of response.

Expression of PDE1C, RTKN2, NGFR, and ELFN1 is associated with PFS, OS and response to TIL therapy

To gain further insight into potential biomarkers and novel targets to improve TIL therapy, we turned to transcriptomic profiling and performed RNA-Seq on the tumor samples (n = 34, Fig. 1, Table 2). We identified nine genes enriched in responders and one gene enriched in non-responders (Fig. 4A). Then, we determined genes associated with PFS. Fifty-two genes were enriched in patients with improved PFS while 31 genes were enriched in patients with poor PFS (Fig. 4B). When assessing genes related to OS, 60 genes were enriched in patients with longer OS and 47 genes were enriched in in patients with shorter OS (Fig. 4C) (FDR q = 0.1 for all). In these analyses, PFS and OS are on a continuum and no cut-offs are used between long and short survival (see Methods). The complete list of significantly enriched genes is annotated in Supplementary Table S10. Of note, we did not find any gene set enriched using gene set enrichment analysis (GSEA) within our cohort.

Figure 4. Identification of gene expression profile associated with outcome to TIL therapy.

Volcano plots of genes enriched by response A, PFS B, and OS C. Genes associated with response to therapy, longer OS, and longer PFS are listed on the right of each plot, with the inverse indicated on the left. The most significantly enriched genes are labeled on the plots. Plots were generated using log (PFS or OS +1) as a continuous variable in DESeq2, where PFS and OS are in months. D, Euler diagrams represent the overlapping genes enriched in response, OS, or PFS. The upper diagram represents genes associated with improved OS, improved PFS, and response, with NGFR, PDE1C, and RTKN2 enriched in these criteria. The lower diagram identifies ELFN1 enriched in patients with poor OS, poor PFS, and lack of response to therapy. Graphs E and F depict 5 samples from responder patients and 11 samples from non-responder patients. E, Scatterplot depicts correlation between ELFN1 expression by RNA-Seq (transcripts per million; tpm) and the mean ddPCR ratio of concentration of target gene to reference gene (ELFN1/HPRT1) per sample, plotted with the line of best fit. Statistical differences were calculated by Spearman correlation and the r² value is represented on the graph. Samples are color coded by response to TIL therapy. F, Graph depicts the mean ddPCR ratio of concentration of target gene to reference gene (ELFN1/HPRT1) per sample stratified by response to TIL therapy. Statistical differences were determined by Mann-Whitney test. Error bars represents mean ± standard deviation.

To identify the most prominent genes whose expression was associated with good or poor outcome, we looked for genes at the intersection of all three outcome metrics. We found PDE1C, RTKN2, and NGFR to be enriched in patients’ samples associated with therapeutic response, improved PFS, and improved OS (Fig. 4D). In turn, we found ELFN1 to be consistently enriched in patients who did not respond to therapy and had poor OS and PFS (Fig. 4D).

Given our interest in finding genes associated with resistance to TIL therapy, we focused on the characterization of ELFN1. As a complementary approach we assessed ELFN1 expression utilizing the droplet digital PCR (ddPCR) assay for those patients who had RNA available after RNA-Seq (Supplementary Table S6). Using the housekeeping gene HPRT1 as a reference, ELFN1 expression was highly concordant with that quantified by RNA-Seq (Pearson’s correlation r² = 0.94; p <0.001, Fig. 4E). Consistent with the RNA-Seq analysis, we confirmed an increase in ELFN1 expression in non-responders compared to responders to TIL-therapy (p = 0.0055; Fig. 4F). Therefore, we believe that ELFN1 may play a role in resistance to TIL therapy.

Previous factors that have been associated with response to TIL therapy include attributes of the infusion product such as the number of TIL infused, the proportion of CD8 TIL and/or total number of CD8 TIL infused (8,10,33,34). ELFN1 expression was negatively correlated with the number of CD8 TIL infused (p= 0.045, Supplementary Fig. S7A), consistent with its association with poor outcome. In contrast, NGFR was positively correlated with the total number of TIL infused, number of CD8 TIL infused, and the proportion of CD8 TIL infused (p= 0.008, 0.0001, 0.03, respectively, Supplementary Fig. S7B–D), also consistent with its correlation with good outcome. While ELFN1 showed a range of expression across patients and a general pattern of lower expression in patients infused with higher numbers of CD8⁺ TIL, the pattern of expression of NGFR, PDE1C and RTKN2 was different and shows selected high expression restricted to a few good responder patients for each gene. Thus, the analysis of expression of these 3 markers individually with infusion TIL features was influenced by outliers and did not allow to discern firm generally applicable associations, despite significant p values for certain NGFR associations. Indeed, RTKN2 negatively correlated with the proportion of CD8 TIL infused, which is at odds with its correlation with good outcome (p=0.006, Supplementary Fig. S7E) and PDE1C did not show a significant correlation with any of these TIL features.

In silico exploration of the expression pattern of ELFN1, NGFR and PDE1C in tumors using publicly available scRNAseq datasets from the pan-cancer blueprint of stromal cell heterogeneity (http://blueprint.lambrechtslab.org) showed the expression of the 3 genes to be confined to clusters mapping to cancer-associated fibroblasts or endothelial cells in the 4 cancer types included in the analysis (breast cancer, lung cancer, colorectal cancer and ovarian cancer, Supplementary Fig. S8A) (35). Clustering of the 30,292 fibroblasts found in all samples showed that ELFN1, NGFR and PDE1C are expressed in distinct subpopulations of fibroblasts (Supplementary Fig. 8B). Finally, the clustering of all endothelial cells also demonstrated that ELFN1 and PDE1C are expressed in distinct endothelial cell populations.

ELFN1 is differentially methylated between responders and non-responders

The differences in ELFN1 mRNA expression between responders and non-responders could be due to underlying differences in genetics, differential activation of pathways, differences in epigenetic state, or a combination of these factors. To evaluate the role of epigenetics in determining these expression patterns, we interrogated DNA methylation across the entire ELFN1 gene. We determined that indeed, ELFN1 is hypermethylated in responding patients (with either complete or partial response, n=20) and hypomethylated in non-responders (for this purpose, a category that includes only patients with progressive disease, n=10) (p=0.04, Fig. 5A; Supplementary Table S8). When focusing on methylation at the 3’ untranslated region (UTR), 5’ UTR, body, and transcription start site (TSS), we observed two distinct clusters of patients, corresponding with the response to therapy (p = 0.0068, Fig. 5B). The first cluster (CL1) features mostly responder patients, with methylation of ELFN1 primarily in the 5’ UTR (promoter region) and consistent with lack of mRNA expression. In contrast, the second cluster (CL2) features mostly patients with progressive disease, with methylation of ELFN1 at the 5’ region, which may suggest a mechanism to regulate ELFN1 expression. The response to therapy and PFS data added to the right of the figure allow to appreciate that responders found in cluster 2 do not have ongoing responses or any response lasting beyond one year. In sum, the methylation data supports the differential mRNA expression of ELFN1 between responders and non-responders, which suggests that ELFN1 may represent a gene associated with resistance to TIL therapy. The differentially methylated oncogenes between responders (CR and PR) and non-responders (PD) are listed in Supplementary Table S11.

Figure 5. ELFN1 DNA methylation status in pre-treatment tissue samples correlates with response.

A, Global mean methylation of ELFN1 in 30 tumors harvested for TIL propagation, comparing responders (blue, n = 20, CR+PR) to non-responders (red, n = 10, PD) patients. Statistical differences were determined using a Wilcoxon test. Error bars represents mean ± standard deviation. B, Unsupervised hierarchical clustering of the 30 tumors shown in (a) by methylation site within ELFN1. Statistical differences were determined using Fisher’s exact test. Location of the methylation site (M sites) is presented within the gene (5’ to 3’, top of the heat map) with the CpG islands in maroon. Each patient’s clinical response and PFS is annotated on the right. The location of ELFN1 on chromosome 7 is also shown above.

DISCUSSION

There is currently no validated predictive biomarker of response to TIL therapy. While there are reports suggesting that neoantigen-specific TIL may mediate at least in part the responses to TIL therapy (25,36), here we show that mutational burden or neoantigen burden do not correlate with response to TIL therapy or progression-free survival. An association was found between neoantigen burden and overall survival, pointing to a benefit of high mutational load, possibly for response to subsequent lines of therapy. Our work demonstrates that the expression of four genes in the pre-treatment tumor microenvironment was tied to outcome to TIL therapy, where PDE1C, RTKN2 and NGFR are associated with improved outcome and ELFN1 marks tumors resistant to TIL therapy. To our knowledge, this is the largest cohort of TIL ACT treated patients in which interrogation of pre-treatment metastatic melanoma tumors by WES, transcriptomic, and epigenetic profiling was performed.

Mutation-reactive TIL are commonly found in TIL expanded from melanoma and other solid tumors, and their presence has been found in select expanded TIL products whose infusion yielded complete responses in melanoma patients (25,36–40). Furthermore, the infusion of an enriched population of mutation-specific TIL has shown clinical benefit in other solid tumor types (36,38,40). A similar lack of association between high tumor mutation burden (TMB) and response to TIL therapy was reported before, although this prior analysis of 24 melanoma patients’ pretreatment tumor tissue did find an association between TMB as well as neoantigen burden with PFS and OS following TIL therapy (9). Our results are slightly different as we did not find an association with PFS. The small number of patients included in the two studies may account for this variability. In addition, since melanoma is a tumor type with one of the highest TMB, our study may not have included enough patients on the lower end of the mutational burden spectrum to find an association. In fact, our study does show that the small number of patients with fewer than 200 predicted neoantigens have poor outcome. Interestingly, high TMB has been correlated with response to checkpoint blockade immunotherapy in melanoma but has also been correlated with higher levels of CD8⁺ T-cell infiltration (41). It is possible that lack of T-cell infiltration in tumors with low TMB led to inability to grow TIL and exclusion from treatment on our study. Taken together, our analysis of somatic mutations and response to TIL therapy suggests that above a certain threshold, the number of mutations or the number of predicted neoantigens are not important drivers of the response to TIL therapy in melanoma. This data is not in contradiction with reported evidence that mutation-specific TIL may contribute to tumor control, but argues that other aberrantly expressed tumor determinants may also significantly contribute to tumor rejection and that one must not underestimate the importance of tumor-associated antigens derived from post-transcriptionally modified or overexpressed proteins, as well as cancer testis antigens, for example.

The finding of significant rates of KCNQ2 and SFTA3 mutations in pre-treatment samples was unexpected. Compared to prior TCGA melanoma analyses, we used an updated algorithm, (MutSig2CV) to detect recurrent mutations. MutSig2CV considers the mutation abundance above background, clustering in hotspots, and the evolutionary conservation of the mutation site, to assess significance (42). However, this alone does not explain our findings, as we detected higher rates of mutations in these genes in our cohort relative to TCGA. Notably, only patients for whom TIL could successfully be expanded (success rate around 65%) were included in this cohort (32). Therefore, the sample set may be skewed towards less immunosuppressed and more infiltrated tumors. Though it is not yet known what role the two genes play in melanoma biology, there is evidence supporting that both could impact the tumor immune microenvironment. KCNQ2 encodes a subunit of a voltage-gated potassium channel that regulates neuronal excitability (43). Elevated potassium concentration in melanoma tumors have been found to be inhibitory to infiltrating T cells thus mutations affecting the amount of potassium in the tumor microenvironment could potentially affect T-cell function (44). SFTA3 encodes a surfactant protein expressed in lung, thyroid, and cornea with putative immunoregulatory function (45,46). Other surfactant proteins, SP-A and SP-D, have been found to act as pattern recognition molecules promoting the phagocytosis of pathogens (47). A similar role is emerging for SFTA3. Indeed, exposure of a human lung alveolar type II cell line A549 to bacterial lipopolysaccharide (LPS) was found to upregulate expression of SFTA3, while the addition of recombinant SFTA3 to alveolar macrophages augmented their phagocytic activity (46,48). In our study, SFTA3 harbored a hotspot mutation in three samples. Conceivably SFTA3 could play a role in the regulation of myeloid cell activation in melanoma. Despite being more frequently found in our cohort, neither KCNQ2 nor SFTA3 mutation correlated with outcome to TIL therapy, and both proteins are expressed at very low levels in melanoma. Further studies will be required to evaluate the association between these mutations and TIL expansion and any causal roles they may play.

Contrary to the findings by Lauss et al., we did not observe enrichment of the HLA class I antigen presentation machinery genes in patient tumors with good clinical outcome to TIL therapy, nor was there an indication of HLA class I deficiency in patients with poor outcome (9). We also did not see a particular immune signature (lymphoid or myeloid) that could be associated with either good or bad clinical outcome (9). One caveat was that the difficulty in extracting high-quality RNA from FFPE tumor tissue limited the sample size of our RNA analysis to 34. Prior studies support that RNA sequencing using RNA extracted from melanoma FFPE samples is feasible and adequately reflects the underlying biology of the disease (49). Nonetheless, we recognize that RNA recovered from FFPE tissue contains more degradation than when extracted from fresh samples. Even though the samples for which we present analysis passed our quality control metrics and we feel confident in the differentially expressed genes that we found, RNA degradation may have prevented the identification of a broader transcriptomic signature. Additional larger cohorts will be required to determine the relevance of these metrics to outcome. Nevertheless, we were able to identify three genes (PDE1C, NGFR and RTKN2) associated with improved outcome to TIL therapy and one gene (ELFN1), associated with poor outcome. More recently high NGFR expression has been found to mark melanoma tumor cells resisting T-cell killing and to drive immunotherapy resistance (50). NGFR is expressed by so called “melanoma initiating tumor cells” that are de-dedifferentiating, becoming slow cycling and downmodulating the expression of differentiation antigens such as Mart-1 (51,52). In our dataset NGFR was associated with good outcome. It is possible that the number of NGFR high melanoma tumor cells in our samples was low across the cohort, or not different between responders and non-responders. Besides the known expression of NGFR by the melanoma tumor cells, we turned to single cell RNA sequencing to identify cells from the tumor microenvironment expressing our genes of interest. Our use of publicly available scRNAseq datasets suggests that ELFN1, NGFR and PDE1C are expressed in fibroblasts and/or endothelial cell populations in several tumor types. The fact that 3 of the 4 molecules we identified are differentially expressed by fibroblasts and endothelial cells is remarkable. It is possible that the genes we identified define specific cell populations with contrasting pro and anti-tumor roles. Fibroblasts are heterogenous cell populations in the tumor microenvironment. Single cell RNA sequencing studies have linked some fibroblasts subsets with immunotherapy resistance in cancer (53), but others have been associated with anti-tumor function, and in general the full heterogeneity of fibroblasts populations in cancer is not defined and probably underestimated (54). Further studies will be needed to elucidate the pattern of expression of NGFR, ELFN1 and PDE1C in the melanoma tumor TME and to understand if they are functionally important in controlling the disease or if they are markers of a cell subset playing an important role in pro or anti-tumor control of melanoma.

It is possible that the genes we have identified are not identifying a cell subset but are directly playing an important role in facilitating or hindering response to TIL therapy. Three of our four genes of interest were previously reported to have an impact on the immune system which could perhaps help explain our findings. For example, PDE1C degrades cyclic AMP, a secondary messenger that inhibits effector T cells (55,56). Thus, PDE1C expression in the tumor microenvironment may prevent suppression of anti-tumor effectors. Surprisingly, NGFR was reported to bind the costimulatory molecule CD80 expressed by antigen-presenting cells, in a manner that is competitive with CTLA-4 binding to CD80, thus the NGFR-CD80 interaction may protect T cells against suppression through CTLA-4 (57,58). Finally, RTKN2 was reported to be highly expressed in CD4⁺ human T cells where it inhibited apoptosis through NF-KappaB signaling or expression of downstream BCL-2 gene (59). In the context of TIL therapy, RTKN2 could potentially protect T cells from apoptosis, favoring persistence and anti-tumor activity post-transfer.

Only ELFN1 was associated with poor PFS and OS, and lack of clinical response. First identified in interneurons in the hippocampus, ELFN1 is essential to regulate the properties of the synaptic connection (60). ELFN1 is an allosteric modulator of glutamate receptor type III, mainly metabotropic glutamate receptor 7 (mGluR7) and mGluR6 (61,62). The metabotropic glutamate receptors type I and II have been shown to be involved with immune system modulation, as both of them are expressed in T cells (63). To date, glutamate receptor type III expression has not been reported in T cells (63). Since it has been demonstrated that ELFN1 cannot interact with types I and II, this makes it highly improbable that ELFN1 would exert any type of regulation on T cells (62). To our knowledge, ELFN1’s expression has never been reported in melanoma cancer cells, but it has been described in other cancers such as breast and ovarian cancers (64,65). Further study in melanoma tumor cells would thus be needed to establish its expression pattern and further characterize its role. Interestingly, we noted an enhanced methylation of ELFN1 in patients’ baseline tumor tissue that responded to therapy when looking at the global methylation profile of this gene. We observed two distinct clusters of methylation, particularly with hypo-methylation in non-responding patients in the body and 3’ CpG island. This suggests that ELFN1 expression may not be dictated canonically by methylation of the promoter region of the gene, but rather methylation within the body or 3’ CpG island may lead to gene repression. Though this data supports a role for ELFN1 in resistance to TIL therapy, methylation of ELFN1 in the body and 3’ CpG island of a few non-responder patients found in the first cluster intermingled with long term responders argues that ELFN1-dependent and ELFN1-independent resistance mechanisms are at play in these patients.

In conclusion, we found that neoantigen load, transcriptional, and epigenetic states correlate with treatment outcomes by assessing a cohort of metastatic melanoma patients receiving TIL therapy. However, a clear limitation of the study is the small sample size, despite our inclusion of patients accrued at two major centers (the MD Anderson and the H. Lee Moffitt Cancer Centers). Therefore, it will be essential to validate these results in larger cohorts of patients treated with TIL therapy, and equally as important to continue the profiling and discovery efforts on future patients to understand and target mechanisms of resistance. These efforts will be especially important in the context of patients that may be heavily pre-treated with other immunotherapy agents such as checkpoint blockade, in order to maximize the utility of TIL ACT in high unmet need patient populations.

TRANSLATIONAL RELEVANCE.

The transfer of autologous tumor-infiltrating lymphocytes (TIL) to treat metastatic melanoma has produced durable responses however mechanisms responsible for resistance to therapy have not yet been elucidated. We interrogated the characteristics of the tumor tissue used to derive TIL in a cohort of patients treated at two institutions to understand the attributes of the tumor microenvironment prior to therapy correlating with response or resistance. No association between the neoantigen load and response to therapy was found, though it was associated with overall survival. Gene expression studies revealed that tumors from patients who experienced a good outcome to therapy expressed higher levels of NGFR, PDE1C and RTKN2 while patients who did not respond to therapy had elevated expression levels of the ELFN1 gene, combined with a distinct ELFN1 DNA methylation pattern. Moreover, investigation of the pattern of expression of ELFN1, NGFR and PDE1C in publicly available single cell RNA sequencing datasets demonstrated preferential expression of the 3 genes in cancer-associated fibroblasts and epithelial cells in multiple tumor types, suggesting a potential role for these microenvironmental elements in defining the response to TIL therapy. These findings point to transcriptional and epigenetic differences associated with response to TIL therapy warranting further investigation.