Role of immune response, inflammation and tumor immune response–related cytokines/chemokines in melanoma progression

Abstract

To investigate the role of tumor cytokines/chemokines in melanoma immune response, we estimated proportions of immune cell subsets in melanoma tumors from TCGA, followed by evaluation of association between cytokine/chemokine expression and these subsets. We then investigated association of immune cell subsets, chemokines, and cytokines with patient survival. Finally, we evaluated immune cell TIL score for correlation with melanoma patient outcome in a separate cohort. There was good agreement between RNA-seq estimation of T-cell subset with pathologist-determined TIL score. Expression levels of cytokines IL-12A, IFNG, IL-10, and chemokines CXCL9 and CXCL10 were positively correlated with PDCD1, CTLA-4, and CD8+ T cell subset, but negatively correlated with tumor purity (Bonferroni-corrected P<0.05). In multivariable analysis, higher expression levels of cytokines IFNG and TGFB1, but not chemokines, were associated with improved OS. Higher expression level of CD8+ T cell subset was also associated with improved OS (HR = 0.06, 95% CI = 0.01-0.35, P = 0.002). Finally, multivariable analysis showed that patients with brisk TIL score had improved melanoma-specific survival compared to those with non-brisk score (HR = 0.51, 95% CI = 0.27-0.98, P = 0.0423). These results suggest that expression of specific tumor cytokines represent important biomarkers of melanoma immune response.

INTRODUCTION

Tumors consist of a variety of cells, including cancer, immune, stromal, and endothelial cells. Tumor-infiltrating lymphocytes (TILs) play an important role in regulating tumor immunity and treatment response in melanoma and other cancers (Li et al., 2016, Pages et al., 2005, Sato et al., 2005, Schumacher et al., 2001). RNA sequencing (RNA-seq) deconvolution tools can identify cellular compositions and functions of immune infiltrates in tumors (Hunt et al., 2018, Li et al., 2016, Newman et al., 2015, Racle et al., 2017).

The immune system can protect the host against tumor development and progression. Cytokines and chemokines are critical mediators of local and systemic immune responses and interactions among cells. For example, IL-2 can stimulate cytotoxic T lymphocytes and is a longstanding treatment option for patients with advanced melanoma (Vuoristo et al., 2001). Tumor interferon (IFN)-γ (IFNG) expression has been associated with a better clinical outcome across multiple cancer types (Fridman et al., 2012). A recent investigation has identified correlation of PD-L1 expression on tumor cells with both lymphocytic infiltration and intratumor IFNG expression (Taube et al., 2012). An immunization strategy using IL-12p70–producing dendritic cells has demonstrated a potential therapeutic role for this cytokine in advanced melanoma (Carreno et al., 2013). Furthermore, transforming growth factor-β (TGFB) and IL-10 were found to be involved in inhibiting activity of T regulatory cells, and decreased serum levels of IL-10 were found to be associated with tumor growth and poor melanoma patient survival (Huang et al., 1999).

Chemokines help immune cells home to the tumor microenvironment. For example, one study found that the chemokines CX3CL1, CXCL9, and CXCL10 were linked to intense T cell infiltration (Harlin et al., 2009). Another study reported that elevated levels of these chemokines were associated with improved patient outcomes (Mlecnik et al., 2010). Additionally, IFNs could promote the secretion of these chemokines in melanoma cells by increasing endothelial expression of E-selectin (Clark et al., 2008, Dengel et al., 2010). However, IFNs can also activate immune regulatory mechanisms, including indoleamine 2,3-dioxygenase, PD-1, and PD-L1 in peripheral tissues, which may limit T-cell activation (Erdag et al., 2012). Therefore, identifying phenotypes of immune homing receptors/ligands that influence immune cell recruitment is important, because it would elucidate how the immune microenvironment might influence melanoma patient outcome and treatment response and help identify targets for more selective personalized therapy. In the current study, we applied the-Tumor Immune Estimation Resource (TIMER) approach to investigate cellular compositions based on melanoma RNA-seq data in The Cancer Genome Atlas (TCGA) (Li et al., 2016); we subsequently evaluated cytokine/chemokine expression associations with immune cell subsets, and with patient outcome measures. We further explored associations between tumor immune cell infiltration and melanoma patient outcome in a separate cohort of melanoma patients who were thoroughly staged, and consistently evaluated, treated, and followed. Finally, since genome-wide investigation can assist in identifying genetic determinants of immune response (Siemers et al., 2017), we evaluated the potential role of TIL-related genetic variants in melanoma patient outcomes in this same cohort of melanoma GWAS patients.

RESULTS

Immune cell composition of melanoma tumors

We analyzed RNA-seq data using TIMER deconvolution to estimate immune cell subsets in patients whose data are available in TCGA (Li et al., 2016). TIMER has an input matrix of built-in reference gene expression signatures for each major immune cell subset, which are collectively used to estimate the relative proportions of each cell type of interest. We estimated immune cell subsets. We then assessed the relationship between immune cell subsets inferred using the TIMER approach and pathologist-established TIL scores recorded in TCGA. We consulted with clinical pathologists to stratify the variable TIL scores (0-6) into two categories, using a cutoff value of 2 and treating each score as a binary outcome (i.e., ≥2 or ≥2), as this resulted in the best alignment with standardized MD Anderson TIL coding. We found that the proportion of the tumor represented by immune cells estimated using TIMER predicted the pathologist-determined TIL score in 80% of cases (P < 0.001, data not shown). These results demonstrate that mRNA gene expression-based immune cell subsets inferred with deconvolution methods are generally consistent with TIL scores determined by pathologists.

Chemokines and cytokines associated with immune cell infiltrates in melanoma tumors

We used TIMER deconvolution to estimate immune cell subsets and tumor purity based on RNA-seq data for the melanoma cohort in TCGA (Snyder et al., 2014). We then assessed associations of cytokine and chemokine gene expression with tumor purity, CD8+ T cell abundance, and overall survival (OS; Table 1, Table S1). We focused primarily on the cytotoxic T cell subset because these cells can directly kill tumor cells and because the presence of cytotoxic T cells at the tumor site predicts favorable outcomes across multiple cancers (Li et al., 2016, Pages et al., 2005, Sato et al., 2005, Schumacher et al., 2001). We found that both PDCD1 and CTLA4 gene expression levels were associated with the CD8+ T cell subset (PDCD1: r² = 0.53, P = 7.97×10⁻³³; CTLA4: r² = 0.38, P = 7.40×10⁻¹⁷; Table 1, Table S1) and with the CD4+ T cell subset (PDCD1: r² = 0.35, P = 4.11×10⁻¹⁴; CTLA4: r² = 0.11, P = 2.56×10⁻²; Table S1) in all melanoma. Those results remained significant in metastatic melanoma but became less significant in primary melanoma due to small sample size (Table S1). With the exception of CX3CL1, gene expression of cytokines and chemokines listed in Table 1 and Table S1 was positively correlated with PDCD1 expression, CTLA-4 expression, CD4+ T cell, CD8+ T cell, Neutrophil and Dendritic cell in the tumor specimens, and negatively correlated with tumor purity (P<0.05 after Bonferroni correction) in all melanoma and metastatic melanoma patients. These data suggest that these cytokines and chemokines are important biomarkers of tumor immune response.

Table 1.

Relationship between cytokine or chemokine gene expression levels and tumor immune response or overall survival in 462 patients with melanoma whose sequencing data were available in The Cancer Genome Atlas.

Gene	PDCD11		CTLA-41		Tumor purity1		CD8+ T cell subset2		Overall survival3
	r2	P	r2	P	r2	P	r2	P	HR4	95% CI4	P
PDCD1	-	-	0.58	1.97E-43	−0.62	5.28E-50	0.53	7.97E-33	0.88	0.83-0.92	7.04E-7
CTLA4	0.58	1.97E-43	-	-	−0.48	4.40E-28	0.38	7.40E-17	0.88	0.83-0.93	2.93E-5
IL12A	0.30	2.46E-11	0.35	1.37E-14	−0.28	9.47E-10	0.30	3.08E-10	0.80	0.73-0.88	3.59E-6
IFNG	0.88	4.34E-155	0.57	1.09E-41	−0.57	5.46E-41	0.65	8.80E-54	0.85	0.80-0.90	4.08E-8
TGFB1	0.40	<1.00E-200	0.38	3.22E-17	−0.44	1.46E-22	0.07	1.30E-1	0.90	0.82-1.00	4.63E-2
IL10	0.61	1.05E-49	0.57	4.06E-41	−0.55	2.47E-32	0.50	2.02E-29	0.85	0.78-0.92	8.83E-5
CX3CL1	0.10	2.43E-2	0.10	5.07E-2	−0.06	2.02E-1	−0.01	8.57E-1	0.93	0.83-1.04	2.09E-1
CXCL9	0.85	<1.00E-200	0.54	1.15E-37	−0.57	5.67E-40	0.64	1.94E-52	0.89	0.85-0.93	1.59E-7
CXCL10	0.76	<1.00E-200	0.47	1.13E-26	−0.49	2.40E-28	0.60	3.23E-43	0.87	0.83-0.91	1.94E-8

¹Spearman correlation test.

²Purity-corrected partial Spearman correlation.

³Univariate analysis.

⁴HR: hazards ratio; CI: confidence interval.

Immune response mediators that predicted melanoma patient survival

We next assessed the association between cytokines, chemokines, and immune cells and their subsets, and clinical outcome, with adjustment for clinical phenotype.

Cytokines and chemokines

We found that immune response–related cytokines and chemokines were correlated with OS in the melanoma TCGA. Increased expression of these cytokines and chemokines, except IL-10 and CX3CL1, significantly contributed to favorable OS (P<0.05 after Bonferroni correction; Table 1). After adjustment for sex, age at diagnosis, disease stage, tumor purity, and CD8+ T cells, two cytokines—IFNG and TGFB1—remained associated with OS, but none of the chemokines significantly contributed to OS in all melanoma and metastatic melanoma patients (Table 2, Table S2). Thus, our data suggest that IFNG and TGFB1 may be important biomarkers of tumor immune response and melanoma patient clinical outcome.

Table 2.

Expression of cytokines and chemokines and other clinical variables associated with overall survival in 462 patients with melanoma whose sequencing data were available in The Cancer Genome Atlas.¹

	Univariate: CD8+ T cell subset			Multivariable: CD8+ T cell subset			Multivariable: cytokines			Multivariable: chemokines
Variable	HR	95% CI	P	HR	95% CI	P	HR	95% CI	P	HR	95% CI	P
CD8+ T cell subset	0.06	0.02-0.24	1.11E-5	0.06	0.01-0.35	0.002	0.57	0.06-5.22	0.620	0.34	0.05-2.32	0.274
Purity	-	-	-	1.17	0.50-2.74	0.712	0.72	0.30-1.74	0.468	0.95	0.40-2.25	0.914
Age	-	-	-	1.02	1.01-1.03	0.001	1.02	1.01-1.03	0.001	1.02	1.01-1.03	0.001
Male sex	-	-	-	0.94	0.69-1.29	0.698	0.95	0.69-1.30	0.726	0.96	0.70-1.32	0.798
White race	-	-	-	0.30	0.14-0.66	0.003	0.25	0.11-0.57	0.001	0.31	0.14-0.68	0.004
Stage 2	-	-	-	1.27	0.81-1.97	0.298	1.23	0.79-1.92	0.362	1.15	0.74-1.80	0.537
Stage 3	-	-	-	1.70	1.12-2.56	0.012	1.71	1.12-2.59	0.012	1.70	1.13-2.58	0.012
Stage 4	-	-	-	3.94	1.97-7.88	<0.001	4.30	2.14-8.67	<0.001	4.52	2.24-9.15	<0.001
IL-12A	-	-	-	-	-	-	0.94	0.84-1.05	0.280	-	-	-
IFNG	-	-	-	-	-	-	0.86	0.77-0.96	0.008	-	-	-
TGFB1	-	-	-	-	-	-	0.84	0.72-0.99	0.033	-	-	-
IL-10	-	-	-	-	-	-	0.98	0.86-1.12	0.737	-	-	-
CX3CL1	-	-	-	-	-	-	-	-	-	0.94	0.84-1.06	0.340
CXCL9	-	-	-	-	-	-	-	-	-	0.91	0.79-1.05	0.176
CXCL10										0.97	0.84-1.11	0.623

¹HR: hazards ratio; CI: confidence interval. Boldface type indicates statistical significance.

Immune cells

In TCGA patients, we found that an increased CD8+ T cell subset was associated with improved OS in univariate analysis (HR = 0.06, 95% CI = 0.02-0.24, P = 1.11×10⁻⁵; Table 2, Table S2). After adjustment for sex, age at diagnosis, disease stage, and tumor purity, the CD8+ T cell subset remained a significant predictor of OS (HR = 0.06, 95% CI = 0.01-0.35, P = 0.002; Table 2). We separately analyzed the relationship between the CD8+ T cell subset and OS in primary (N=102) and metastatic (N=348) melanoma tumor cohorts within TCGA, and found that the CD8+ T cell subset was associated with OS in the metastatic melanoma tumor cohort (HR=0.06, 95% CI = 0.01-0.42, P = 0.004), but not the primary melanoma tumor cohort (HR=0.83, 95% CI = 0.002-423, P = 0.953) in multivariable analysis (Table S2). We did not detect any significant association between CD4+ T cell and OS in all melanoma, primary melanoma and metastatic melanoma patients (Table S3).

In this study we did not have access to melanoma tumor specimens and therefore did not directly measure tumor immune cell subsets or expression levels of chemokines/cytokines. However, because TCGA data demonstrated that CD8+ subsets were correlated with TIL score, we next used TIL score as a surrogate marker of CD8+ subsets and assessed whether pathologist-determined TIL score was correlated with melanoma-specific survival (MSS) (Figure 1) in patients from MDACC who were part of a previously published genome-wide association study (GWAS) (Amos et al., 2011). Of the 1804 melanoma patients in the dataset, 1396 had undergone evaluation for lymphocyte infiltration of their primary tumor. 13.0% of these 1396 patients had a brisk immune score, represented by a band-like infiltrate along the entire deep edge of the invasive component and interacting with the tumor cells. We found that a brisk immune score was associated with prolonged MSS, and MSS durations were significantly different across three immune score groups: absent, non-brisk, and brisk (log-rank P = 0.0003; Figure 1). After adjustment for sex, age at diagnosis, and stage at diagnosis, multivariable analysis showed that patients with a brisk immune score were less likely than those with a non-brisk score to die from melanoma (HR = 0.51, 95% CI = 0.27-0.98, P = 0.0423, data not shown). These results confirm that tumor lymphocyte infiltration can contribute independently to melanoma patient outcome. Since tumor specimens were not available for the majority of patients in the GWAS cohort, we did not attempt to evaluate T-cells subsets or expression levels of tumor chemokines and cytokines in this cohort.

Figure 1.

Kaplan-Meier melanoma-specific survival curves for patients from The University of Texas MD Anderson Cancer Center with various levels of tumor infiltration by lymphocytes. Absent=0, only scattered lymphocytes; non-brisk=1 and 2, anything that is neither minimal nor brisk; brisk=3, band-like infiltrate along the entire deep edge of the invasive component and interacting with the tumor cells.

Evaluation of TIL-associated SNPs with Patient Outcome Measures

To investigate potential genetic mechanisms linking TIL with melanoma patient outcomes, we selected SNPs with a nominal P-value<10⁻⁵ with regard to predicting TIL (brisk vs non-brisk/minimal) in our study population. Forty-one SNPs from different gene regions reached a significance level of <10⁻⁵ (Table S4). Among these SNPs, several had P values <0.05 for association with disease-free survival (DFS); none was associated with MSS or OS. In particular, the C allele in the rs10032022 SNP (nearby LOC105377253 gene) was associated with non-brisk TIL infiltration score and was nominally associated with increased disease recurrence (HR=1.17, P=0.0163, Table S1). Two other SNPs (rs4603413 and rs1760828, within the CXXC6P1 gene) also reached nominal significance level of <0.05 for predicting DFS; paradoxically, their reference A1 alleles were associated with brisk TIL infiltration but poorer DFS. These preliminary data suggest that genetic variation contributing to TIL infiltration might also contribute to risk of melanoma recurrence; future validation of these findings will be required (Table S4).

DISCUSSION

In the current study, we estimated immune cell subsets in the melanoma cohort of TCGA using the TIMER approach based on RNA-seq data, and found that immune cell subsets agreed well with the pathologist-determined TIL score in 80% of the cases. We found that several cytokines and chemokines were associated with immune response, as denoted by the CD8+ T cell subset. We further observed an association of OS with selected cytokines and chemokines as well as intra-tumor immune response. Finally, we verified the correlation between TILs and MSS in a cohort of patients from our institution, and identified preliminary associations between germline genetic polymorphisms and TIL infiltration. These results demonstrate that systematic evaluation of chemokines and cytokines influencing lymphocyte infiltration may help identify high-risk melanoma patients and potentially suggest new or more selective immunotherapeutic approaches.

Reference gene profiles from different sources (circulating or tumor tissues), different types of high-throughput tumor data (transcriptional or epigenetics), and different deconvolution approaches can generate non-unanimous immune cell subset results. Because no sorted or purified immune cell subsets were reported from TCGA, we could not use TCGA to confirm the estimated results. However, previous studies used sorted immune cells as ground-truth data to validate these deconvolution approaches and guided our selection of estimated immune cell subsets from TCGA data. T-cell subsets estimated using TIMER accurately predicted pathologist-determined TIL score in 80% of cases, suggesting that deconvolution can yield accurate estimations of T-cell subsets and provide reliable tumor purity estimates. Subsequent analyses demonstrated that inferred T-cell subsets estimates could help account for heterogeneity in disease progression: an increased proportion of CD8+ T cells infiltrating a tumor was independently associated with improved patient outcome.

Identifying cytokines and chemokines associated with immune response can help elucidate mechanisms regulating those immune responses in the tumor microenvironment and provide T cell activation/inhibition data more specific than those conveyed by pathologist-determined TIL score. Using TIMER deconvolution and pathway analysis to estimate the abundance of immune cell types, Li et al identified significant chemokine-receptor networks for immune infiltration in a variety of cancers. In that study, CD8+ T cell levels were associated with an enriched subset of chemokine-receptor pairs, including CCL3,4,5–CCR1,5 and XCL1,2–XCR1, and macrophage abundance was related to the CXCL12–CXCR4 pair in thyroid, head and neck, stomach, and colon cancers (Li et al., 2016). We evaluated selected inflammation and tumor immune response-related cytokines and chemokines and found that several cytokines were correlated with CD8+ T cell abundance, suggesting that these molecules could be potentially targeted to enhance CD8+ T cell responses, or to identify patients more likely to benefit from immunotherapy.

The current study has limitations. While TCGA does not have prior systemic therapy (other than IFNA) and the MDACC GWAS consists of a cohort of patients recruited prior to the widespread availability of modern checkpoint inhibitor immunotherapies, some melanoma patients included in these studies might have received systemic treatments, including immune-based treatments, that could have affected their survival. Limitation of clinical annotation, lack of detailed pathology information, and lack of prior treatment data in the TCGA cohort and the multiplicity of (generally ineffective) systemic therapies administered to the MDACC GWAS cohort had prevented us from evaluating potential influences of systemic therapy, including immunotherapy, in the patients included in the current study. Melanoma tumors in TCGA included a mix of primary tumors and metastases, and there were very few primaries and metastases from the same patient. The association between tumor T-cell subsets and melanoma outcome was only significant in the metastatic tumors, which could be due to smaller sample size in the primary tumor cohort (102 primary tumors vs 348 metastases), or due to the smaller number of events among patients who provided primary melanoma tumors but not metastatic tumors.

We understand that current TIL scoring systems have limitations, including a lack of complete standardization. Lymphocytic infiltration in TCGA was coded as both distribution and density distribution data. The Dermatopathology service at MD Anderson has used a standardized approach to classifying TIL for the last 19 years as “minimal”, “non-brisk”, and “brisk,” based on the distribution data. Due to differences in scoring systems, TIL scores reported in the TCGA cohort cannot be directly translated to those of the MD Anderson cohort; for the purposes of this study, we combined distribution and density TIL scoring data in TCGA into a single score to allow comparison to MDA TIL scoring data.

Another potential limitation of this study was the small number of observations relative to the highly dimensional expression data. In addition, it is recognized that the CD8+ T-cells population itself is very heterogeneous; examination of the relative roles of CD8+ T-cell subsets is important but beyond the scope of the current investigation.

We estimated immune-cell subsets based on RNA-seq from whole tissues at a single time point in TIMER; the approach is unable to formally distinguish stromal or intra-tumor immune-cell localization or take into account tumor heterogeneity or different anatomic sites of metastasis (for example, solid organ vs. lymph node). Future studies that included data with improved spatial and temporal information could help resolve these challenges. Finally, although we identified SNPs contributing to TIL infiltration associated with risk of melanoma recurrence in MD Anderson cohort, further mechanistic analysis and validation is beyond the scope of this study. We acknowledge therefore that the findings reported herein, including those related to association between tumor immune microenvironment cytokines and melanoma patient survival, require further investigation and external validation.

In conclusion, our results suggest that expression of specific tumor cytokines represent important biomarkers of melanoma immune response, and immune response contributes independently to melanoma patient outcome.

MATERIALS AND METHODS

TCGA dataset patients

We used the melanoma TCGA dataset to estimate immune cell subsets and identify biomarkers of immune response and patient outcome. We downloaded demographic, clinical phenotype, RNA-seq, and follow-up information of 479 TCGA melanoma patients (https://portal.gdc.cancer.gov/projects/TCGA-SKCM). Clinical prognostic factors [2009 American Joint Committee on Cancer stage (AJCC), Breslow tumor thickness, ulceration, and mitosis], tumor location, mutation load, tumor molecular features, and TIL score data. In TCGA, TIL was coded as both distribution and density as follows: lymphocyte distribution, 0=0%; 1≤25%; 2=25%−50%; 3≥50% [minimal or absent=0, non-brisk=1 and 2, brisk=3]; lymphocyte density, 0=absent; 1=mild; 2=moderate; 3=severe. This data was re-coded for the purposes of this study as distribution + density [range 0-6]. Complete clinical data were available for 470 patients; 290 patients were men, 437 were white, 231 had stage I or II disease, 194 had stage III or IV disease, and 45 had unknown stage. The median age at diagnosis was 58 years. Complete follow-up information was available for 462 of the 470 patients; the median follow-up time was 28.9 months, and 222 patients died during the follow-up period. mRNA gene expression was measured using the IlluminaHiSeq_RNASeqV2 platform.

MDACC patient cohort

Our Melanoma SPORE Project 5 investigation was a hospital-based case-control study in which patients with cutaneous melanoma and cancer-free controls were recruited at MDACC from March 1998- August 2008. Samples and data were available for 931 patients and 1026 controls, who were frequency-matched according to age and sex. Patients and controls completed a comprehensive skin lifestyle questionnaire and passed quality control filters for genotyping. Data from an additional case series comprising 873 individuals presenting for treatment of cutaneous melanoma at MDACC were also included, bringing the total number of melanoma patients to 1804. We excluded those with melanoma in situ or atypical melanocytic proliferation, as well as nonwhite patients, in whom the incidence of melanoma is very low. All individuals gave written informed consent under an MDACC Institutional Review Board-approved protocol. European-descent melanoma patients with all stages of disease evaluated at MDACC were eligible for inclusion. Demographic information (sex, age, weight, height, and race), pigmentation characteristics (skin, hair, and eye color), sun exposure, and clinical prognostic factors (2009 AJCC stage, Breslow tumor thickness, ulceration, and mitosis) were determined for each participant. Of the 1804 melanoma patients in our published GWAS(Amos et al., 2011), 1396 had received pathologist evaluation of lymphocyte infiltration of the tumor. 13% of the 1396 patients had a brisk immune score (lymphocyte distribution: 0=0%, 1≤25%, 2=25%-50%, 3≥50% [absent=0, only scattered lymphocytes; non-brisk=1 and 2, anything that is neither minimal nor brisk; brisk=3, band-like infiltrate along the entire deep edge of the invasive component and interacting with the tumor cells]). The current study used high-density genotype data obtained from DNA of our melanoma patient and control populations. The samples were genotyped by using the Illumina HumanQmnil-Quad_v1-0_B array and details of genotyping and quality control were described previously (Amos et al., 2011). Finally, 818,237 genotyped SNPs remained for the primary analysis. Imputation of ungenotyped SNPs was performed using MACH (Li et al., 2010) applied to genotype data from all subjects. In total, we had 2,649,586 imputed or directly genotyped SNPs eligible for an association study. (Fang et al., 2013). SNP analysis was performed through ProbABEL software (Aulchenko et al., 2010).

Deconvolution analysis

We used the validated deconvolution approach TIMER (Li et al., 2016) to estimate proportions of infiltrating immune cell types (B cells, CD4+ T cells, CD8+ T cells, natural killer cells, macrophages, neutrophils, and dendritic cells) from bulk tumor tissue based on RNA-seq data in TCGA samples. These approaches rely on an input of reference gene expression to estimate the proportions of the cell types represented in the sample. The deconvolution of cell proportions can be solved through a heuristic algorithm. We first median-centered and unit-normalized data measurements to bring the variables to the same scale.

Statistical analysis

We selected several cytokines and chemokines that in previous studies showed correlation with tumor immune response or patient outcomes (Carreno et al., 2013, Fridman et al., 2012, Harlin et al., 2009, Huang et al., 1999, Mlecnik et al., 2010, Taube et al., 2012, Vuoristo et al., 2001), and assessed the relationship of these cytokines and chemokines with T-cell infiltration and MSS and OS. The relationship between cytokine and chemokine gene expression levels and tumor immune response was tested using Spearman correlation. Follow-up data and clinical phenotypes available from TCGA are described above. For the MDACC GWAS dataset, patient demographic and clinical data, including body mass index, primary tumor status, tumor thickness, ulceration, mitosis, sentinel lymph node status, tumor sites, and disease stage, OS, MSS, and DFS, were extracted from the MDACC Melanoma Informatics, Tissue Resource, and Pathology Core database.

We performed the following analyses using TCGA data, for which only OS data were available. We first performed a univariate analysis of the effect of CD8+ T cell subset on OS using a Cox regression model. We then adjusted for other covariates, including sex, age at diagnosis, race, disease stage, and tumor purity (the percentage of malignant cells in a tumor tissue specimen). Finally, we used a multivariable biomarker model that included the variables in the above model plus the gene expression levels of selected cytokines and chemokines. In the MDACC dataset, we investigated the relationship between pathologist-assessed TIL score and MSS, with or without adjustment for age at diagnosis, sex, and disease stage, using a Cox regression model. All data analyses were performed using R language (Team, 2014) and the SAS Enterprise Guide 4.3 software program (SAS Institute Inc.). All P values were two-sided, and a P < 0.05 was considered statistically significant. Multiple testing problems were considered for the relationship between multiple chemokine/cytokine-related genes and immune response or survival using Bonferroni correction. P<0.0056 [=0.05/9] was considered statistically significant after considering multiple testing problems.

Data availability

Datasets related to this article can be found at [https://portal.gdc.cancer.gov/projects/TCGA-SKCM], hosted at [National Cancer Institute GDC Data Portal] (The Cancer Genome Atlas Network.Genomic Classification of Cutaneous Melanoma. Cell 2015, 161(7): 1681-1696)

Abbreviations:

AJCC

American Joint Committee on Cancer stage (AJCC)

CI

confidence interval

DFS

disease-free survival

HR

hazard ratio

IFN

interferon

IL

interleukin

OS

overall survival

PD-L1

programmed death-ligand 1

CTLA-4

Cytotoxic T-Lymphocyte Associated Protein 4

RNA-seq

RNA sequencing

TCGA

The Cancer Genome Atlas

MDACC

The University of Texas MD Anderson Cancer Center

MSS

melanoma-specific survival

TILs

tumor-infiltrating lymphocytes

TIMER

Tumor Immune Estimation Resource software