Using global gene expression to discriminate thin melanomas with poor outcomes

ABSTRACT

Most melanomas present as thin lesions (≤1.0 mm) with a good prognosis; however, a small percentage of patients with thin lesions experience recurrence or metastasis. The aim of our study was to identify a distinct pattern of gene expression within thin melanomas known to have eventually metastasized to regional lymph nodes or distant sites compared with those that followed the typical course with good response to wide local excision alone. Patients who were disease-free for a minimum of 10 y served as controls (n = 10) to the experimental group who developed metastasis (n = 9). Laser capture microdissection was used to specifically isolate cancer cells from formalin-fixed paraffin-embedded tissue with subsequent gene expression analysis on Affymetrix Human Transcriptome Array 2.0 Arrays. Although gene expression differences were observed between the patients with thin melanoma with poor clinical outcome and those with good clinical outcome, neither the number of genes nor the magnitude of the fold difference was very substantial or significant. Cluster analysis with this subset of genes could definitively separate a subset of the poor responders from the good responders, but there remained a mixed group of tumors that could not be predicted from gene expression alone. Pathway analysis identified cellular processes that were regulated based on the response, including categories commonly associated with melanoma progression. Ultimately, we concluded that there were very few differences between these groups. Future research will be required and investigation of the mutational landscape may be another strategy to uncover genomic changes that drive recurrence and metastasis in thin melanoma.

Introduction

Melanoma of the skin is the fifth most common cancer diagnosed in the United States with an estimated 76,380 new cases in 2016.¹ The incidence rates of melanoma have increased relentlessly at a faster rate than those of any other malignancy over the past 2 decades.² It is well established that there has been a significant increase in the number of thin melanomas (<1.0 mm in thickness), which may be attributed to increased skin cancer awareness and skin cancer screening in many Western countries.^3,4 As a result, melanocytic lesions are increasingly being sampled or removed at an early stage. Although the vast majority have a favorable outcome, the American Joint Committee on Cancer (AJCC) has emphasized the variability (85% to 99%) of 10-year survival for thin melanoma,⁵ and estimates suggest that up to 20% of thin melanomas can be associated with metastasis, among which 5% are fatal.⁶ Given the increasing number of cases of thin melanoma, it is clearly important to identify factors that influence prognosis.

Despite greater molecular characterization of melanoma and other cancers,⁷ Breslow thickness is still one of the most powerful prognostic factors in both the AJCC and UICC schemes. Within the group of thin melanomas, Breslow thickness ≥0.75 mm has been associated with poorer prognosis as well as several other patient and histologic features including patient age >65 years, male sex, tumor site on the scalp or neck, tumor invasion of the entire papillary dermis, ulceration, high mitotic rate (≥1/mm²), and nodular or acral lentiginous histologic subtype.^8,9 Other researchers have investigated whether these clinicopathologic features can be improved upon to refine prognostic information for thin melanomas. Thareja and colleagues examined the utility of anti-phosphohistone 3 staining of mitotic cells as a strategy to improve upon mitotic index (stained by H&E) as a predictor of positive sentinel nodes in thin melanomas¹⁰ but were not able to show any significance. In a small exploratory study, Glazer and colleagues refined histologic criteria using an algorithm based on the spatial and statistical distribution of the nuclear chromatin and were able to identify those patients at high risk for metastatic disease.¹¹ Examination of prognostic factors in thin melanomas that recur late suggests that ulceration may be the most important feature associated with risk for metastases.¹²

Many candidate biomarkers have been studied in melanoma in an attempt to improve upon clinicopathologic features as prognostic indicators.^13-19 Several of these biomarkers have been linked to growth, invasion, or immunologic processes based on the concept that during malignant transformation and progression melanoma cells must acquire the ability to proliferate indefinitely and escape their local environment. However, very few of these studies specifically address the problem of identifying high-risk thin melanomas. One factor that has been studied in this context has been the cell proliferation–associated antigen Ki-67,²⁰ which appears to have prognostic value in separating high-risk thin melanomas from those at low risk.²⁰ However, this biomarker has not displaced mitotic count as an indicator of proliferative activity in melanoma.

An alternative to biomarkers based on biological considerations is to use gene expression profiling to identify differentially expressed genes that may be involved in melanoma progression and development of metastases.^21-23 Although studies have identified potential expression signatures, few have specifically studied thin melanoma. Jaeger and colleagues studied differential gene expression across a variety of melanomas of different stages of progression, subtypes, and tumor thickness.²⁴ In a comparison of melanomas ≤1.0 mm with those ≥2.0 mm, 199 genes were identified as differentially expressed and the Gene Ontology terms of cell communication, cell adhesion, and ectoderm development were significantly enriched in terms of enhanced expression in the thin melanomas.²⁴ However, this study did not examine the impact of gene expression on prognosis within the thin melanoma cohort. To our knowledge, the present study is the first to specifically study differential gene expression in a cohort of thin melanomas with poor clinical outcome compared with a matched group with expected good outcome.

Results

Patient demographics

Table 1 lists the clinicopathologic features of patients with good and poor response. Because of the small numbers of patients, there were no statistical differences between the 2 groups; however, there were some trends that were consistent with previous studies. Although the mean Breslow thickness was similar in the 2 groups, there were 3 lesions greater than 0.75 mm in the poor response group compared with only one among the good responders. There were more males in the group with poor response compared with those with a good response (66.7% vs. 40%, respectively). The poor response group was also more likely to be Clark's level 4 than the good responders (44.4% vs. 30.0%). Although the data on mitotic index were incomplete, there were more lesions with mitotic index greater than 1.0 in the poor responders. Interestingly, none of the lesions were noted to have ulceration or had evidence of angiolymphatic/perineural invasion.

Table 1.

Clinicopathologic characteristics of study patients.

Response	Patient ID	Age	Sex	Primary Site	Time to met	Histology	Clark's Level	Breslow depth (mm)	Ulcer.
Poor	AX4169	55	M	Thigh	SLN+	SSM	4	0.87	No
Poor	AX4171	59	M	Scalp	1 yr	SSM	4	0.80	No
Poor	AX4170	70	M	Leg	5 yr	SSM	3	0.48	No
Poor	AX4182	51	F	Scalp	3 yr	SSM	2	0.32	No
Poor	AX4684	68	F	Arm	SLN+	NM	4	0.45	No
Poor	AX4704	34	F	Chest	SLN+	SSM	3	0.91	No
Poor	AX4803	41	M	Chest	5 mo.	SSM	3	0.35	No
Poor	AX5332	74	M	Back	3 yr	SSM	4	0.70	No
Poor	AX5487	66	M	Scalp	3 yr	LMM	2	0.25	No
Good	AX4957	79	F	Arm	>10 yr	SSM	4	0.67	No
Good	AX4958	22	F	Chest	>10 yr	SSM	3	0.35	No
Good	AX4959	55	M	Back	>10 yr	SSM	2	0.32	No
Good	AX4960	78	M	Scalp	>10 yr	SSM	4	0.93	No
Good	AX4961	44	F	Leg	>10 yr	SSM	3	0.65	No
Good	AX4962	85	F	Arm	>10 yr	SSM	2	0.48	No
Good	AX4963	46	M	Thigh	>10 yr	SSM	2	0.40	No
Good	AX4964	79	M	Scalp	>10 yr	LMM	3	0.51	No
Good	AX4965	50	F	Arm	>10 yr	SSM	4	0.64	No
Good	AX4966	30	F	Arm	>10 yr	SSM	3	0.57	No

Time to met, time to metastasis; SLN+, positive SLN node at time of resection; mo, months; yr, years; Ulcer., ulceration.

Histology: SSM, superficial spreading melanoma; NM, nodular melanoma; LMM, lentigo maligna melanoma.

Differential gene expression analysis

Global gene expression analysis was performed on Affymetrix Transcriptome arrays with differentially expressed genes identified using ANOVA based on good or poor response. The initial analysis identified 153 differentially expressed genes (P ≤ 0.01) between the good and poor responders, but only a single gene met a 1.5-fold cutoff: small nucleolar RNA, C/D box 69 (SNORD69) was downregulated 1.7-fold in the poor responders.

To detect signaling changes between the groups, we used Elsevier's Pathway Studio for sub-network enrichment analysis (SNEA). Similar to gene set enrichment analysis (GSEA), SNEA determines if a sub-network of genes is highly regulated between conditions. Table 2 lists the sub-networks of genes associated with regulating cell processes that are highly regulated between the patients with poor response and those with a good response. The associated sub-networks of genes regulating cell differentiation, apoptosis, and cell proliferation were different according to prognosis (Fig. 1). While the actual analysis was not limited to a subset of genes, for ease of illustration these figures include only genes that meet relaxed cutoffs that are appropriate for pathway analysis.^25,26

Table 2.

SNEA results of highly regulated sub-networks between poor and good responders.

Regulating Cellular Processes of:	P-value
Differentiation	0.00E+00
Apoptosis	0.00E+00
Proliferation	0.00E+00
Keratinization	1.70E-09
mRNA metabolism	4.98E-05
Vesicle-mediated transport	7.69E-05
Keratinocyte differentiation	9.00E-05
Translation initiation	1.69E-04
RNA recruitment	1.94E-04
RNA replication	2.60E-04

Figure 1.

Sub-networks of genes regulating the cellular processes of differentiation, apoptosis, and proliferation. Figure was limited to include only genes that met relaxed cutoffs of P ≤ 0.10 and 1.2-fold cutoff. Genes shown in gray are upregulated in poor responders compared with good responders; genes shown in white are downregulated.

Clustering and subset analysis

Hierarchical clustering was performed based on the 153 differentially expressed genes (P ≤ 0.01, Fig. 2). The clustering resulted in the identification of 3 main branches: (1) a population of 6 good responders, (2) a population of 7 poor responders, and (3) a mixed population of 4 good responders and 2 poor responders. It should be noted that further investigation did not produce any clinical or outcome measures, nor were there any analytical conditions (e.g., RNA quality) that would contribute to differentiating this mixed group. Nevertheless, to examine whether this selected subset of patients provided any useful information, the data were filtered to exclude the mixed population. When reduced to this limited number of samples, many more genes (1,914) were identified as differentially expressed (P ≤ 0.01), but only 12 genes showed a magnitude of change greater than 2-fold. This includes genes such as ZNF764, CMTM8, and TRADD, which are upregulated in the poor response patients, and SCARNA9, which is downregulated. We then looked at the sub-networks of genes involved in regulating cellular processes that are highly regulated between poor and good responders. The top rated sub-networks are commonly seen in cancer studies (cell proliferation, cell differentiation, apoptosis) but others may be more indicative of melanoma, such as the cell processes of keratinization and skin barrier (Fig. 3). While the process of identifying the highly regulated sub-networks utilizes all of the data on the array, Fig. 3 only includes genes that meet the criteria of P ≤ 0.10 and 1.2-fold. These are very loose restrictions to enable better visualization of the pathway and no conclusion should be drawn on the individual genes. [Note: only APOC1 has a fold change greater than 1.5-fold.]

Figure 2.

Hierarchical clustering of 19 samples of thin melanoma. Samples were clustered based on the expression pattern of 153 differentially expressed genes (P ≤ 0.01) between poor and good responders. Individual genes are represented on the X-axis and the different tumor samples are shown on the Y-axis. Poor responders are shown in blue; good responders are shown in green.

Figure 3.

Sub-networks of genes regulating the cell processes of skin barrier and keratinization. Figure was limited to include only genes that met relaxed cutoffs of P ≤ 0.10 and 1.2-fold cutoff. Genes shown in gray are upregulated in poor responders compared with good responders; gene in white are downregulated.

Discussion

Only a small percentage of patients with thin melanomas will undergo local recurrence or develop metastases, but the increasing incidence of melanoma and the success of early detection strategies combine to elevate this high-risk subset of patients to a significant number. It is therefore of increasing importance to be able to identify and target these poor responders so that more aggressive treatments and follow-up can be offered. However, it is also important that identification of this sub-set of patients is robust as excessive initial therapy may result in unnecessary complications. In a recent randomized trial of 900 patients, wider local excision of 3 cm as opposed to 1 cm in high-risk primary melanoma greater than 2 mm in thickness was associated with a doubling of surgical complications.²⁷ In addition, sentinel lymph node biopsy (SLNB) is a controversial issue in melanoma as randomized trial data show modest benefit at the expense of complications in one-third of patients.²⁸ Clearly, the decision to proceed with intensive therapy in patients with thin melanoma is not trivial, and how best to identify patients most likely to benefit from intensive therapy remains a clinical dilemma.

By definition, thin melanomas (≤1 mm thickness) fall into stage pT1a or b, with the latter defined by the presence of ulceration and/or mitotic rate ≥1 per mm². In general, SLNB is not recommended for primary melanomas ≤0.75 mm thick. Currently, besides thickness, there is little consensus regarding high-risk features for finding a positive sentinel lymph node. Historically, risk factors include ulceration, high mitotic rate, lymphovascular invasion, regression, growth, sex, age, and site.⁹ A recent international workshop on thin melanoma suggested that 0.75 mm should be the upper limit for the definition of a thin melanoma⁹ as almost all thin melanomas that develop metastases show tumor thickness between 0.75 and 1.0 mm.^29,30 However, in our small cohort of poor responders, 6 of 9 patients presented with a Breslow thickness less than 0.75 mm.

In this study we sought to identify a distinct pattern of gene expression within thin melanomas that had eventually metastasized to regional lymph nodes or distant sites. Previous gene expression studies have identified molecular signatures related to metastasis in melanoma^21,24,31 and have suggested that alteration of genes involved in cell cycle regulation, mitosis, cell communication, and cell adhesion were the principal cellular processes associated with the metastatic lesions. This present study showed a remarkable lack of discriminating genes in the primary tumors between the good responders and patients with eventual poor response. Our hope was that through different gene expression patterns we could tease out which thin melanomas had a higher propensity to metastasize to regional lymph nodes or distant sites. Recently, a 15-gene PCR-based assay (derived from DNA microarray analysis) that was developed for uveal melanoma³² was applied to cutaneous melanoma.³³ The signature contained 12 discriminating genes (HTR2B, GNAQ, ID2, MTUS1, ECM1, ROBO1, SATB1, LTA4H, EIF1B, RAB31, FXR1, and LMCD1) and 3 control genes (MRPS21, RBM23, and SAP130). In combination with SLNB, this signature was able to improve identification of high-risk SLNB-negative patients with aggressive disease and patients classified as high-risk by conventional parameters who are unlikely to have progression of their disease. This signature contains genes of various functions with no obvious relationship to known melanoma biology or previous gene signatures,^21,24,31 but it will be interesting to study its utility in thin melanomas.

Although we were restricted to a small sample size due to the nature of the patient population and the availability of sufficient tissue, we attempted to overcome some of the limitations by minimizing potential sources of variation, for example by using laser capture microdissection to select only pathologist-guided melanoma cells for analysis. Nevertheless, the lack of any differences in gene expression highlights the fact that predicting which thin melanomas will eventually lead to a poor clinical outcome is not straightforward and is more subtle than using gene expression to predict disease stage, subtype, or discriminating primary tumors from metastases. One possible explanation that needs to be investigated is that individual differences in immune response play a more crucial role than molecular tumor composition in predicting melanoma recurrence and metastasis. Regardless of the underlying cause, this patient population remains challenging, and multi-institutional efforts will be required to advance genomic knowledge and develop prognostic information for thin melanomas.

Methods

Patient selection

We used a case-control study design. Cases were selected from patients initially diagnosed with thin melanoma (Breslow thickness ≤ 1.0 mm) who underwent primary excision with adequate margins and eventually developed regional or distant metastasis on follow-up. Patients were excluded if available tissue was not sufficient for analysis. Controls were selected from patients with thin melanomas who underwent primary excision with adequate margins and remained disease-free on a minimum 10-year follow-up. The clinicopathologic features of the cohorts are described in Table 1. The use of patient samples and data was approved by the Beaumont Hospital Institutional Review Board.

Laser capture microdissection (LCM)

Paraffin embedded thin melanoma tissue samples were cut into 5-µm sections and mounted onto polyethylene naphthalate (PEN) membrane glass slides. Regions of interest were identified by pathology on corresponding H&E slides of the tissue sections. The stained slides were microdissected using the ArcturusXT™ Microdissection System (Molecular Devices, Sunnyvale, CA) onto CapSure® HS LCM Caps (Molecular Devices).

RNA isolation

RNA was isolated using a RNeasy FFPE Kit (Qiagen, Hilden, DE). The transfer film with the attached dissected material was separated from each CapSure HS LCM Cap and placed in lysis buffer. RNA was purified using spin cartridge technology following the manufacturer's protocol, quantified (Nanodrop 8000, Thermo Scientific), and stored at −80°C.

Whole transcriptome (WT) amplification, purification, fragmentation, and hybridization

RNA was amplified and labeled using the GeneChip™ WT Pico Reagent Kit (Affymetrix, Santa Clara, CA), which enables amplification and target preparation for whole-transcriptome expression analysis. Amplification was performed with 10 ng of total RNA input following procedures described in the WT Pico Reagent Kit user manual. The amplified cDNA was quantified, fragmented, and labeled in preparation for hybridization to GeneChip™ Human Transcriptome 2.0 Arrays (Affymetrix) using 5.5 μg of single-stranded cDNA product and following protocols outlined in the user manual. Washing, staining (GeneChip® Fluidics Station 450, Affymetrix), and scanning (GeneChip® Scanner 3000, Affymetrix) were performed following protocols outlined in the GeneChip™ Expression Wash, Stain, and Scan User Manual for Cartridge Arrays. Array scans from this process yielded signal intensities comparable to arrays prepared from high-quality RNA from cell lines (data not shown).

Gene expression analysis

CEL files containing the raw intensity data from the Affymetrix GeneChip arrays were imported into Partek® Genomics Suite™ (6.6 version 6.15.0730) and normalized using the robust multichip average with a guanine-cytosine content background correction, quantile normalization, log2-transformation, and median polish probeset summarization. Exons were then summarized to genes using the average of the probesets. Differentially expressed genes were identified using ANOVA with 2 factors: prognosis and scan date (random variable). Unsupervised hierarchical clustering was performed using Partek® software (Partek Inc., St. Louis, MO, USA). Hierarchical clustering analysis was performed using Euclidean distance as a similarity measure and average linkage for the agglomerative method. Gene set enrichment analysis (GSEA) and pathway analysis was performed using Pathway Studio (version 10.5.05; Elsevier, Philadelphia, PA). The data discussed in this publication have been deposited in the NCBI Gene Expression Omnibus³⁴ and are accessible through GEO Series accession number GSE79662.

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.

Funding

This work was supported by the generous philanthropic support of the Zafarana family.