A miRNA-based signature detected in primary melanoma tissue predicts development of brain metastasis

Abstract

Purpose

Brain metastasis is the major cause of mortality among melanoma patients. A molecular prognostic test that can reliably stratify patients at initial melanoma diagnosis by risk of developing brain metastasis may inform the clinical management of these patients.

Experimental Design

We performed a retrospective, cohort-based study analyzing genome-wide and targeted microRNA expression profiling of primary melanoma tumors of three patient cohorts (n= 92, n= 119, n= 45) with extensive clinical follow up. We used Cox regression analysis to establish a microRNA-based signature that improves the ability of the current clinicopathologic staging system to predict the development of brain metastasis.

Results

Our analyses identified a 4-microRNA (miR-150–5p, miR-15b-5p, miR-16–5p, and miR-374b-3p) prognostic signature that, in combination with stage, distinguished primary melanomas that metastasized to the brain from non-recurrent and non-brain-metastatic primary tumors (training cohort: C-index=81.4%, validation cohort: C-index=67.4%, independent cohort: C-index=76.9%). Corresponding Kaplan-Meier curves of high- vs. low-risk patients displayed a clear separation in brain-metastasis-free and overall survival (training: p<0.001, p<0.001, validation: p=0.033, p=0.007, independent: p=0.021, p=0.022, respectively). Finally, of the microRNA in the prognostic model, we found that the expression of a key lymphocyte miRNA, miR-150–5p, which is less abundant in primary melanomas metastatic to brain, correlated with presence of CD45⁺ tumor infiltrating lymphocytes.

Conclusions

A prognostic assay based on the described miRNA expression signature combined with the currently used staging criteria may improve accuracy of primary melanoma patient prognoses and aid clinical management of patients, including selection for adjuvant treatment or clinical trials of adjuvant therapies.

Introduction

Melanoma is a prevalent cancer of increasing incidence globally (1), with a cost of nearly a billion dollars annually to the US health care system(2). Most patients are initially diagnosed with localized melanoma (stage I and II) and are effectively cured through surgical management. Despite generally good prognoses, ∼10% and ∼40% of patients diagnosed with localized lesions, respectively, die from melanoma. In contrast, nearly 40% of patients with stage III (nodal spread) cancers are long-term survivors(3). These outcomes indicate that current histopathology-based staging does not fully capture heterogeneity of outcomes for primary melanoma patients. Coupling these staging criteria with additional tools (such as molecular markers) has the potential to vastly improve prognostic accuracy for these patients. For those individuals that develop metastatic melanoma, brain metastasis, in particular, is a major clinical burden. Brain metastasis occurs in 5–15% of all melanoma patients and is the cause of up to 50% of melanoma deaths(4). Brain metastasis is rapidly fatal, with median survival of 4 months after diagnosis(5). Approaches to better stratify primary melanoma patients by their risk of developing brain metastasis may yield substantial benefit to high-risk patients through intensified surveillance, aggressive treatment, and future adjuvant therapies. Moreover, interrogation of molecular alterations within primary tumors of different outcomes may inform our understanding of the biology and mechanisms underlying melanoma progression. Consequently, identification of robust prognostic markers of brain metastasis represents an important clinical challenge in melanoma biology.

Recent studies have explored the potential of microRNA (miRNA), small, non-coding RNA that regulate stability and translation of protein-coding messenger RNA, as biomarkers of disease outcomes(6). In contrast to messenger RNA, microRNA expression levels and patterns are highly stable in formalin-fixed, paraffin-embedded (FFPE) tissue, with expression comparable to those of fresh tissue (7–10). Moreover, in addition to stability in FFPE tissues, their diagnostic and prognostic value resides in their lineage/tumor type specificity, and ease of isolation and quantification(10, 11). Additionally, rather than mere passenger alterations, burgeoning literature has revealed fundamental roles for some miRNA in the tumorigenicity of diverse cancers(12), including melanoma(13–15), providing a mechanistic basis for the prognostic potential of some miRNA. However, global analysis of the value of miRNA as biomarkers of disease outcomes is still underexplored in melanoma, and its investigation may help improve the accuracy of patient prognoses.

We analyzed the prognostic value of miRNA expression in primary melanoma, specifically for their ability to distinguish patients who develop melanoma brain metastasis from those with non-recurrent tumors or tumors metastatic to other sites. Here, we report a molecular signature of miRNA expression detected in primary melanoma tissue at the time of initial diagnosis that, in combination with stage, stratifies patients by their risk of developing brain metastasis. Moreover, we identify a melanoma cell-extrinsic miRNA (the lymphocyte miRNA, miR-150) as a core member of this signature, providing further evidence of the importance of immune response to the control of melanoma.

Materials and Methods

Study Population

All patients were enrolled in the Interdisciplinary Melanoma Cooperative Group (IMCG) database of NYU Langone Medical Center (NY, NY). Informed consent was obtained from all patients and approval acquired by the Institutional Review Board of NYU School of Medicine (protocol #10362). IMCG patients are actively, prospectively followed up every 6 months after primary tumor removal and every 3 months after a recurrence. Extensive clinical information is annotated in the IMCG database. Parameters relevant to this study include date of diagnosis, Breslow thickness, ulceration presence, mitotic index, and stage for primary tumors; dates of diagnoses and anatomical locations of metastases, date of last follow-up or death, last clinical status, and cause of death. All patients developing brain metastasis did so during active follow up. Medical oncologists (A.P. and I.O.) review all deceased patients’ histories and last clinical statuses to determine if melanoma was the cause of death. Patients were selected using the following criteria: For our initial study cohort (training), we excluded patients who were non-recurrent/metastatic and whose initial diagnosis occurred within 1 year of the start of the study and thus not expected to reach >3 years of follow-up by the anticipated study conclusion. We excluded all non-cutaneous melanoma patients. From the remaining cases, we randomly selected 92 cases, with a ratio of half non-recurrent to half recurrent patients. Within recurrent patients, to maximize the chance to detect a prognostic signature of the development of brain metastasis, we selected approximately 50% who had already developed brain metastasis. Subsequent validation cohorts were selected similarly. All patients were treated surgically for removal of primary melanoma lesions.

Clinical Specimens

All melanoma specimens were human primary, cutaneous melanoma samples that were collected, formalin fixed, and paraffin embedded at the time of surgery (Discovery cohort: 1989:2007, Validation cohort: 1994:2009, Independent cohort: 1997:2010). All tumors were classified according to the 2009 American Joint Committee on Cancer (AJCC) staging system.

RNA Extraction

5μM sections of formalin-fixed paraffin-embedded (FFPE) samples (4–12 sections per patient depending on the size of the primary tumor) were attached to PEN-Membrane 2.0μM slides (Leica) designed for laser capture micro-dissection. Primary melanoma tissues were macroscopically dissected using disposable scalpels (Feather No. 11) under a dissecting microscope and guided by images of hematoxylin and eosin (H&E) staining of consecutive sections on which a board-certified pathologist (F.D.) marked areas of tumor. Exclusion of surrounding stroma by macro-dissection consistently allowed RNA extraction from greater than 80% tumor. RNA extraction was performed using the miRNeasy FFPE kit (Qiagen, Netherlands) following manufacturer’s recommendations, using the Xylene/Ethanol method for deparaffinization/rehydration.

microRNA microarray expression profiling and data pre-processing

MiRNA expression profiling of RNA extracted from FFPE primary melanomas was performed by Exiqon, Inc, blinded to sample characteristics. Briefly, a reference sample was generated by mixing an equal amount of all samples analyzed, independently for each cohort. RNA from sample and reference were labeled with Hy3™ or Hy5™ fluorescent label, respectively, mixed pair-wise, and hybridized to the miRCURY™ LNA arrays (Exiqon, Denmark). Arrays were scanned using the Agilent G2565BA Microarray Scanner System (Agilent Technologies, Inc., USA). Image analysis was carried out using the ImaGene software (BioDiscovery, Inc., USA). The quantified signals were background corrected (Normexp with offset value 10) and normalized using the global Lowess (LOcally WEighted Scatterplot Smoothing) regression algorithm. Data are represented as the Log2 transformation of the ratio of median intensities per probe of Hy3-labeled sample to Hy5-labeled reference. MIAME-compliant microarray data was deposited into the Gene Expression Omnibus of NCBI (GEO) under the accession GSE62372.

microRNA real-time qPCR and data processing

10ng RNA was reverse transcribed in 10μl reactions using the miRCURY LNA™ Universal RT microRNA PCR, Polyadenylation and cDNA synthesis kit (Exiqon, Denmark). cDNA was diluted 50X and assayed in 10ul PCR reactions according to the protocol for miRCURY LNA™ Universal RT microRNA PCR; each microRNA was assayed once by qPCR on the microRNA Ready-to-Use PCR, Human pick-n-mix panel. Negative controls excluding template from the reverse transcription reaction was performed and profiled like the samples. The amplification was performed in a LightCycler® 480 Real-Time PCR System (Roche, Germany) in 384-well plates. The amplification curves were analyzed using the Roche LC software, both for determination of cycle threshold (Ct) (by the 2nd derivative method) and for melting curve analysis. The amplification efficiency was calculated using algorithms similar to the LinReg software. All assays were inspected for distinct melting curves and the Tm was checked to be within known specifications for the assay. Measurements that did not pass the service provider’s (Exiqon) threshold (Ct < 37 and at least 5 cycles different from no template control measurements) were imputed as the lowest value (highest Ct) measured across samples for each miRNA (miR-16–5p: 1 samples, miR-374b-3p: 6 samples). Using NormFinder(16) we identified three miRNA to be most stably expressed across all samples (hsa-let7e-5p, miR-30d-5p, and miR-423–3p). The average Ct for these miRNA was employed as a normalizer for this data set. A normalized Ct is defined as NormCt = Ct (average of 3 normalizers) – Ct(sample).

Evaluation of tumor-infiltrating lymphocytes (TILs) by histological analysis of primary tumors

The presence or absence of tumor-infiltrating lymphocytes (TILs) was recorded. The cases with TILs were further sub-classified into brisk and non-brisk based on the published criteria(17, 18).

CD45 immunohistochemistry

Immunohistochemistry was performed on 5 micron sections of FFPE primary melanoma tissue using mouse anti-human CD45, (Leukocyte Common Antigen, LCA), clone RP2/18 (Cat. 760–2505, Ventana Medical Systems Tucson, AZ USA). In brief, sections were deparaffinized in xylene (3 changes), rehydrated through graded alcohols (3 changes 100% ethanol, 3 changes 95% ethanol) and rinsed in distilled water. Antibody incubation and detection were carried out at 40°C on a NexES instrument (Ventana Medical Systems Tucson, AZ USA) using Ventana’s reagent buffer and detection kits unless otherwise noted. Endogenous peroxidase activity was blocked with hydrogen peroxide. CD45 antibody was applied neat as directed by the manufacturer and incubated for 30 minutes. Primary antibody was detected with iView biotinylated goat anti-mouse followed by application of streptavidin-horseradish-peroxidase conjugate. The complex was visualized with 3,3 diaminobenzidene and enhanced with copper sulfate. Slides were washed in distilled water, counterstained with hematoxylin, dehydrated and mounted with permanent media. Appropriate positive and negative controls were included with the study sections. An attending pathologist (FD), who was blinded to patients’ clinical data and miR-150 expression measurements, scored CD45 expression as the absolute number of positively-stained immune cells demonstrating characteristic lymphocytic morphology in a representative high-power field (HPF; 0.2 mm²) that was selected by scanning each slide at × 40 to find the field with the highest antibody expression. The CD45 expression was assessed in two locations: intratumoral and peritumoral (within 0.5 mm from the tumor edge). In addition, the percentage of CD45⁺ cells, defined as the ratio of CD45⁺ cells over tumor cells in one high-power field, was calculated in the intratumoral location.

Statistical Analyses

Using the discovery cohort (n = 92), we analyzed results for 1,360 miRNA, of which data for 578 miRNA were available for at least 67% samples. All miRNA were standardized into mean 0 and unit variance variables by subtracting the mean and dividing by the standard deviation (SD) within each cohort for scale consistency across cohorts. MiRNAs were first ranked by univariate association of expression level for each miRNA with brain metastasis-free survival via Cox proportional hazards regression analysis, with adjustment for tumor stage as a continuous variable. The top 50 ranking miRNAs were used as candidates included in the multivariate Cox proportional hazards model with tumor stage in the regression. The 4 miRNA-signature was selected by minimizing Akaike’s information criterion (AIC) of the multivariate Cox proportional hazards model through forward stepwise selections(19). In the training cohort, the four identified miRNAs were present in all samples, with the exception of one patient who had a missing value for miR-15b. The linear coefficients of the miRNAs in the signature were estimated from the remaining 91 training samples with complete data. Subsequent analyses were performed with all samples after data imputation for the 4 microRNA of the signature. Microarray-based measurements (training and validation cohorts) were imputed as the average value of all samples in the cohort (n = 92 and n = 119, respectively) for a given miRNA. RT-qPCR-based measurements (independent cohort) were imputed as the lowest value detected across samples for a given miRNA. The linear combination of model predictors weighted by regression coefficients was defined as the risk score. To test the classifier, regression coefficients of the Cox model were applied to the validation cohort and the independent cohort to obtain their risk scores. The discriminating power of the risk score was evaluated by Harrell’s C index(20) in the three cohorts separately. The risk scores were also evaluated for utility in predicting brain metastasis by identifying the area under the Receiver Operating Characteristic (ROC) curve (AUC) in the three cohorts separately(21). For this purpose, we defined cases as patients with brain metastasis documented by routine imaging (MRI or CT) of high-risk patients and/or imaging of symptomatic patients with no restriction for time to brain metastasis (as defined by the time from initial primary tumor diagnosis to initial brain metastasis documentation). We defined controls as non brain-metastasis patients (non-recurrent/non-metastatic patients and recurrent patients without documented brain metastasis) with ≥3 years of follow-up from initial primary tumor diagnosis. A risk score cutoff using the Youden Index of the ROC curve was chosen to separate patients into high- and low- recurrence risk groups(22). The same cutoff was applied to the validation and independent cohorts to stratify patients. Statistical comparison of ROC curves generated from the miRNA + stage classifier to stage alone was performed with Delong’s test for two correlated ROC curves. To compare the brain metastasis-free and overall survival distributions of the two groups in each cohort, data were plotted in Kaplan-Meier survival curves and analyzed statistically by the Wilcoxon test, as it is a more sensitive measure than the log-rank test to compare differences in survival probability between groups that occur in early points in time (23).

Kruskal-Wallis rank sum test was used to compare the risk scores among non-recurrent/non-metastatic, non-brain metastatic, brain metastatic concomitant with or subsequent to other sites, and isolated, first-site brain metastatic patient groups.

Two-sample unpaired t test was used to compare log transformed miRNA expression between brisk and non-brisk TILs. The association between brisk status and B-Met was assessed by χ² test. Pearson correlation coefficients were used to characterize the correlation between log-transformed miRNA expression and CD45 expression.

All statistical analyses were performed in R 2.14.0. Specifically, functions from packages “survival” and “ROCR” were used.

Results

Brain-metastasis prognostic miRNA signature derived from primary melanoma tissue

To explore the potential prognostic value of microRNA in primary melanoma, we performed global and targeted microRNA expression profiling of total RNA extracted from FFPE primary cutaneous melanoma patient tissues. Table 1 presents a detailed summary of clinicopathological features of patient groups included in this study. The three patient cohorts were comprised of 92 (training), 119 (validation), and 45 (independent) individuals. Localized melanomas (stage I and II) comprise 77.2%, 62.2%, and 95.6% of patients, respectively. For each cohort, approximately 50% non-recurrent and 50% recurrent patients were selected, and among the latter, approximately 50% had already developed a documented brain metastasis. Median follow-up time for control patients (non-recurrent or recurrent without brain metastasis) was approximately 90 months (range: 10–231), 50 months (range: 12–177), and 70 months (range: 16–100), respectively, with >80% exceeding 5 years of follow-up after initial diagnosis. Using global miRNA expression profiling data from our training cohort, we established a prognostic model of the development of brain metastasis. To do so, we identified the top-ranking miRNAs by univariate association of expression with brain metastasis-free survival (BMFS) via Cox proportional hazards (PH) regression analysis with adjustment for tumor stage. We used this candidate set in multivariate Cox PH regression analysis, including tumor stage coded as a continuous variable, to establish a model predictive of BMFS. This analysis revealed a 4-miRNA signature (miR-15b-5p, miR-150–5p, miR-16–5p, and miR-374b-3p) with adjustment for tumor stage (Table 2) that robustly predicts brain metastasis development (Harrell’s C-index of 81.4% (95% CI = 69.6%, 93.2%). In contrast, a model using clinical stage alone achieved Harrell’s C-index of 69.0%. The same set of miRNAs was identified coding stage as a categorical value or using a logistic regression-based approach (Table S1, Table S2).

Table 1.

Summary of clinical characteristics of patients used in this study.

Cohort	Training			Validation			Independent
Characteristic	Non- recurrent	Non-brain met*	Brain met	Non- recurrent	Non-brain met	Brain met	Non- recurrent	Non-brain met	Brain met
All Cases (n)	44	22	26	47	41	31	26	6	13
Ulceration Status (n)
No	27	10	8	27	17	12	24	1	8
Yes	17	12	18	20	24	19	2	5	5
Stage At Diagnosis (n)
I	14	3	2	13	4	4	25	2	7
II	29	10	13	23	16	11	1	4	4
III	1	9	11	11	21	14	0	0	2
IV	0	0	0	0	0	2	0	0	0
Thickness (mm)
Median	2.2	2.9	3.1	2.2	4.2	3.7	1.0	1.8	1.6
Range	0.9–11	1.1–12	0.85–30	0.85–24	1.1–30	0.5–13	0.4–4	0.95–30	0.66–16
Histogical Subtype (n)
Nodular	22	16	20	23	20	19	7	3	6
SSM	19	4	3	21	10	9	17	2	6
Other	3	2	3	3	11	3	2	1	1
Follow Up (months)
Median	90.5	88.0	31.5	53.0	43.0	38.0	65.0	74.5	34.0
Range	31–124	10–231	15–198	12–133	7–177	5–122	36–100	16–97	11–131
Alive	84.1%	40.9%	3.8%	83.0%	34.2%	3.2%	88.5%	16.7%	0.0%
Dead	15.9%	59.1%	96.2%	17.0%	65.9%	96.8%	11.5%	83.3%	100.0%

^*met = metastatic.

Table 2.

Members, hazard ratios, and p values of the 4-miRNA signature predicting development of melanoma brain metastasis derived from Cox proportional hazards modeling.

	HR (95% CI)*	P-value
hsa-miR-150–5p	0.31 (0.14, 0.68)	0.003
hsa-miR-15b-5p	4.50 (1.48, 13.68)	0.008
hsa-miR-16–5p	0.11 (0.03, 0.45)	0.002
hsa-miR-374b-3p	0.57 (0.39, 0.84)	0.004
stage	2.95 (1.43, 6.06)	0.003

^*Hazard ratios (HR) and 95% confidence intervals (CI) of the standardized miRNA expressions obtained from the discovery cohort.

From the training cohort, we defined risk scores for individual patients as the linear combination of the 4 miRNAs and tumor stage weighted by their regression coefficients in the Cox model. The same set of coefficients was applied to microarray or targeted RT-qPCR data, respectively, of two additional patient cohorts (validation (n = 119) and independent (n = 45)) to obtain risk scores. In combination with stage, this brain metastasis-distinguishing miRNA classifier displayed Harrell’s C-index of 67.4% (95% CI = (56.3%, 78.5%)) and 76.9% (95% CI = (58.4%, 95.4%) in the validation and independent cohorts cohort respectively (Table 3). Using the same patient cohorts, a model using clinical stage alone achieved Harrell’s C-index of 65.8% and 72.1%, respectively.

Table 3.

4-miRNA signature in combination with stage improves accuracy of prognosis of brain metastasis development.

		Training	Validation	Independent
All samples (n)		92	119	45
C-index (95% Confidence Interval)	miRNA + stage	81.4% (69.6%, 93.2%)	67.4% (56.3%, 78.5%)	76.9% (58.4%, 95.4%)
	stage	69.0% (58.7%, 79.3%)	65.8% (55.4%, 76.1%)	72.1% (59.9%, 84.4%)
≥ 3 years follow-up (n)		85	92	44
AUC (95% Confidence Interval)	miRNA + stage	81.5% (70.3%, 92.6%)	64.1% (52.5%, 75.8%)	78.7% (63.1%, 94.2%)
	stage	70.4% (59.6%, 81.2%)	59.2% (47.8%, 70.6%)	67.6% (52.0%, 83.2%)

ROC curves support prognostic potential of miRNA signature

To further evaluate the prognostic performance of the miRNA signature obtained by Cox analysis, we calculated the area under the curve (AUC) of the receiver operating characteristic (ROC) curves for development of brain metastasis. For this purpose, we excluded patients for whom we did not have at least 3 years of follow-up, unless they had a documented brain metastasis. We defined cases as patients with documented brain metastasis regardless of time to brain metastasis (n = 26, 31, 13), and controls as patients who did not have documented brain metastasis and for whom we had ≥3 years of follow-up since initial melanoma diagnosis (n = 59, 61, 31). In combination with stage, this brain metastasis-distinguishing miRNA classifier displayed an AUC of 81.5% (95% CI = (70.3%, 92.6%)), 64.1% (95% CI = (52.5%, 75.8%)) and 78.7% (95% CI = (63.1%, 94.2%) in the training, validation and independent cohorts, respectively (Table 3). In contrast, using the same patient cohorts, a model using clinical stage alone yielded AUCs of 70.4% (95% CI = (59.6%, 81.2%) p = 0.042), 59.2% (95% CI = (47.8% 70.6%) p = 0.318), and 67.6% (95% CI = (52.0%, 83.2%) p = 0.215), respectively. For patient subsets restricted for ≥3 years follow-up since initial melanoma diagnosis or development of brain metastasis, the negative and positive predictive values of this classifier are 87.5%, 75.0%, and 87.5%, and 46.7%, 41.7%, and 50.0%, respectively, in the three cohorts (Table S3, Figure S1). While low, PPV using this classifier is higher than the percentage of brain metastasis cases within each cohort (26/85 (30.6%), 31/92 (33.7%), and 13/44 (29.5%), respectively). We selected the Youden index of the ROC curve of the discovery cohort as a risk-score cutoff to stratify patients into high- and low-risk groups for all patients in each cohort. High- and low-risk groups plotted in Kaplan-Meier survival curves displayed significant differences in the three patient cohorts for BMFS (p < 0.001, p = 0.033, and p = 0.021, respectively) and overall survival (OS) (p < 0.001, p = 0.007, and p = 0.022, respectively) (Figure 1).

Figure 1. A 4-miRNA signature accurately stratifies patients by brain metastasis-free and overall survival.

Risk scores for individual patients were defined as the linear combination of model predictors weighted by regression coefficients. An optimal risk score cutoff selected by using the Youden Index of the discovery cohort ROC curve was chosen to classify patients into high- and low- risk groups. Group data were plotted in Kaplan-Meier survival curves and statistical analysis performed by Wilcoxon tests for brain metastasis-free and overall survival of training (A, n = 92), validation (B, n = 119), and independent (C, n = 45) patient cohorts.

Finally, to further examine the prognostic value of the miRNA signature beyond stage alone, we estimated the signature’s performance within individual at-diagnosis stages for each cohort. We observed that the model performed best for patients with stage II tumors (Figure S2, Table S4). These results may suggest that the miRNA signature would be more beneficial to melanoma patients whose tumors have not advanced to clinically evident nodal disease, though it may also reflect that these cohorts contained greater numbers of these tumors.

Collectively, our results demonstrate that expression of a small set of miRNA, measured from primary melanoma tissues at initial melanoma diagnosis, may improve the prognostic capacity of AJCC stage for the development of brain metastasis.

miRNA signature reflects brain-specific tropism for some melanomas

Our group recently identified a subgroup of 36% of melanoma brain metastasis patients whose tumors spread to the brain as the isolated, first site of visceral metastasis(24). This subset of patients suggests that some melanomas may be clinicopathologically and/or molecularly distinct and may develop molecular alterations that dictate tissue tropism of metastatic cells. To examine if the described miRNA signature is reflective of this possibility, we further stratified the studied melanoma patients from all cohorts into four groups: non-recurrent/non-metastatic (n = 115), non-brain metastatic (n = 67), brain metastatic concomitant with or subsequent to other sites of metastasis (n = 51) and brain metastatic as the isolated, first site of visceral metastasis (n = 16). We found that the median signature risk scores reflect the difference among these four groups (p < 0.001, Kruskal-Wallis rank sum test; Figure 2). This sub-analysis suggests that the described signature may be reflective of brain-specific tropism for some melanomas.

Figure 2. Median risk scores support that the 4-miRNA classifier is partially reflective of brain tropism.

Patients from all cohorts were divided into four groups (non-recurrent (n = 115), non-brain recurrent (n = 67), concomitant brain metastatic (n = 51), and isolated, first-site brain metastatic (n = 16)) and risk scores plotted in box & whiskers format (A). Whiskers are min-max, excluding outliers (defined as above 1.5 IQR away from the 25% and 75% quantiles). Statistical analysis was performed by Kruskal-Wallis rank-sum test.

miR-150–5p levels correlate with the presence of tumor infiltrating lymphocytes

We further explored the prognostic value of miR-150–5p, because it is not expressed in melanoma cell lines and short-term cultures (25, 26), but well detected in melanoma tissues measured in this study and others(27, 28), which likely contain mixed cell-type populations. Since miR-150–5p is highly expressed in mature B and T-cells(29, 30), we hypothesized that differences in its expression between primary melanomas of different outcomes may be reflective of tumor infiltrating lymphocytes (TILs), rather than of a tumor-cell intrinsic property. Indeed, we found that miR-150–5p levels were significantly lower in samples classified as ‘non-brisk’ (n = 27) compared to ‘brisk’ (n = 55) TIL response (p = 0.006, Figure 3A) in a set of 82 primary melanoma samples (23 metastatic to brain, 20 metastatic to other sites, and 39 non-metastatic), with lower miR-150–5p detection associated to higher brain-metastasis occurrence (p = 0.025, Figure 3B). To independently confirm the correlation between miR-150–5p levels and hematopoietic cell infiltrates, we stained 48 primary melanomas, which we had previously analyzed for miR-150–5p expression and TILs, for the leukocyte marker, CD45. We found that miR-150–5p positively correlates with the absolute and relative number of intratumoral CD45⁺ cells (p = 0.007 and p = 0.016 respectively, Figure 3C,D and Figure S3). Collectively, these results indicate that miR-150–5p expression detected in primary melanoma tissues is likely to derive from TILs. Moreover, these data suggest that defective immune response to a primary melanoma or active immune evasion by melanoma cells may enable increased brain-metastasis potential for some patients’ tumors.

Figure 3. MiR-150–5pexpression correlates with CD45⁺ TILs in primary melanoma tissues.

MiR-150–5p levels were measured by RT-qPCR in a subset of primary melanoma samples (n = 82; 23 brain metastasis, 20 extracranial metastases, and 39 non-recurrences). Differences in miR-150–5p expression were evaluated by two-tailed unpaired T-test for samples with ‘non-brisk’ (n=27) vs. ‘brisk’ (n=55) TIL responses (p = 0.006) (A) and brain metastasis (n=23) vs. Non brain-metastasis (n=59) (p = 0.025) (B). A subset of 48 primary melanomas (brain metastasis (n = 12), non-brain metastasis (n = 36)) previously analyzed for miR-150–5p expression and TILs were stained for the leukocyte marker, CD45. MiR-150–5p levels were plotted against the absolute (C) and relative (D) number of intratumoral CD45⁺ cells (p = 0.007 and p = 0.016 respectively).

Discussion

In this study, we evaluated the prognostic capacity of miRNA expression from primary cutaneous melanoma tissues, particularly for its ability to predict the development of brain metastasis. We analyzed miRNA expression for approximately 250 primary melanoma patient tumors, which to our knowledge, represents one of the largest molecular profiling studies of primary melanoma to date, a tumor type for which access to large numbers of tissues with extensive clinical follow-up is rare. Using these datasets, we developed and validated a 4-miRNA (miR-15b-5p, miR-150–5p, miR-16–5p, miR-374b-3p) prognostic model that, combined with tumor stage, classifies primary melanoma patients by risk of developing brain metastasis.

Defining an accurate prognosis for primary melanoma patients is a key clinical goal. Currently, management of these patients is guided exclusively by gross and histopathological criteria, which clinicians use to select patients for more extensive staging (including sentinel lymph node evaluation), intensified surveillance, and/or adjuvant therapy (e.g. interferon-alpha). However, low- and moderate-risk primary melanoma patients with histopathologically similar tumors can have vastly different outcomes, supporting the notion that current staging can be improved to better capture this heterogeneity. In supplement of histopathological features, molecular changes within tumors of similar staging hold immense promise to better understand, diagnose, prognosticate, and develop treatments for melanoma.

The utility of miRNA as biomarkers of cancer outcomes has been increasingly explored. Studies of a variety of cancers have identified associations between miRNA and clinical outcomes(31–34) with limited use in prognostic modeling(35, 36). MiRNA as prognostic biomarkers have not been comprehensively explored in primary melanoma tissues. Using a candidate approach, Satzger and colleagues found that high abundance of miR-15b-5p, consistent with the signature herein, correlates with poor recurrence-free and overall survival and is an independent prognostic factor in primary melanoma tissues(37). In addition, our group and others found that expression in primary melanomas of pro-metastatic miRNA correlates with various tumor progression parameters(14, 15). Moreover, our group previously identified a miRNA signature, of which miR-15b-5p and miR-150–5p are members, measured in sera of primary melanoma patients at initial diagnosis that is prognostic for disease recurrence(38, 39). In contrast to these reports, the current study focused on miRNA in primary melanoma tissues as prognostic biomarkers for the development of site-specific (brain) metastasis. We speculate that certain molecular alterations in some primary melanomas may be indicative of organ-specific tropism as opposed to general metastatic capacity. In support of this notion, we found that the miRNA signature risk scores reflect the differences between primary tumors that metastasized to brain as the isolated, first visceral site of metastasis compared to those that metastasized to brain concomitant with or subsequent to other sites of metastasis. This correlation is consistent with the concept that brain tropism may be encoded in some primary melanomas.

Interestingly, signature member miR-150–5p was well detected in primary melanoma tissues, however its expression is not detectable in isolated melanoma cultures(25). We reasoned that this paradox might be explained by an extrinsic non-melanoma-cell source for miR-150–5p detected in melanoma tissues. MiR-150–5p is highly expressed in and a key regulator of mature B and T-cells(29, 30, 40), suggesting TILs may be a key source of detected miR-150–5p. Supporting this hypothesis, we found that miR-150–5p levels significantly correlated with CD45⁺ TILs in primary melanoma patient samples. The immune system has long been known to restrain some melanomas, and recent studies have found TIL response associates with various parameters of melanoma progression and outcome(41–44). Moreover, the remarkable efficacy achieved for some patients using immune checkpoint inhibitors (ipilimumab, nivolumab, pembrolizumab) clearly shows the intricate relationship between melanoma cells and host immune response(45, 46). Our findings support this concept of immune cells as key repressors of melanoma progression generally, and perhaps to brain metastasis, specifically. As a biomarker, miR-150–5p expression may represent a quantitative way to measure TIL response to aid prognostic stratification of high- and low-risk patients. Moreover, if TIL infiltrate represents a good predictor of response to immunotherapies, then miR-150–5p expression may also represent a valuable predictive biomarker.

A molecular classifier capable of reliably ascribing risk at the time of initial melanoma diagnosis of a future disease outcome could be a valuable tool to improve melanoma patient management. This concept is particularly important for prediction of brain metastasis, which causes the majority of deaths from melanoma. Adjuvant therapies effective at preventing development of brain metastasis are currently lacking, thus the clinical utility of the described signature is unclear at present. However, ongoing (NCT01667419, NCT01682083, NCT00636168, NCT01682213) and future clinical trials will determine the usefulness of BRAF and MEK inhibitors (vemurafenib, dabrafenib, and trametinib) and immunotherapies (anti-CTLA4, -PD1/PDL1) as adjuvant therapies for primary melanoma patients. Moreover, some of these therapeutics have already benefitted patients with existing brain metastases(47–49), thus treatment options efficacious in the adjuvant setting for high-risk primary melanoma patients may soon be available. We believe that development of a prognostic assay that informs selection of patients who would benefit most from effective adjuvant therapies and identify high-risk patients for clinical trials would ideally be ready for use when such therapies become available. However, translation of this finding into a clinically useful prognostic assay requires a number of future steps. Independent replication of our results with additional patient sets from other institutions would strengthen the reliability of our findings. Further, development of a technical platform amenable to reproducible and precise quantification (such as RT-qPCR) are required to advance this assay toward clinical use. In support of this translation, we find a good correlation between risk scores for identical samples derived from microarray and RT-qPCR data (r = 0.412, p < 0.001). Finally, a large, multi-center prospective study to assess the robustness of this prognostic signature in the general melanoma population would solidify feasibility of conversion of our findings to a clinical prognostic test.

In this study, we elected to develop a model incorporating miRNA expression in combination with tumor staging. Staging incorporates known prognostic factors for primary melanoma (tumor thickness, ulceration, and mitotic index) into a unified prognostic tool. Moreover, its inclusion allowed us to assess the significance of miRNA signatures in predicting brain metastasis survivals for patients with the same stage. However, the significance of stage in the multivariate model may reflect a lead-time bias associated with detecting stage I of II patients before they reach a more advanced stage. In addition, while our model qualitatively improved upon the predictive capacity of clinical stage alone in ROC/AUC analyses of BMFS, these differences were only statistically significant in the training cohort. Our data suggests that the described classifier partially reflects tumor aggressiveness, thus it is a possibility that inclusion of more patients with highly aggressive tumors (but who have not yet developed brain metastasis) in our control groups than would be expected in the general melanoma population lessened the statistical power of the classifier in the described analyses. This possibility implies that the accuracy of the signature may improve when applied on expanded patient cohorts that better resemble the general primary melanoma patient population. Alternatively, inclusion of additional non-recurrent stage I and II tumors in our analyses may have yielded different and/or additional microRNA with brain metastasis prognostic potential. Finally, while our model is a basis for the making of a useful clinical assay, the PPV is low currently. We opted to identify a risk score cut-off to maximize the NPV, as ‘false negatives’ are a much less desirable outcome of a clinical prognostic assay. The low PPV may also partially reflect the need for longer follow-up for control patients, who may still develop brain metastasis, which occurred as late as over 10 years after initial melanoma diagnosis in the patient populations studied here. We expect to continue to monitor these patients and examine the classifier’s accuracy over time based on their risk scores defined here.

In summary, we developed a prognostic miRNA classifier that, in combination with stage, stratifies early stage primary melanoma patients by their risk of developing brain metastasis in three patient cohorts. The described miRNA signature represents a first step to developing a useful molecular, brain-metastasis prognostic assay for melanoma. We believe such a test has the potential to improve clinical care and outcomes of primary melanoma patients.

Statement of translational relevance.

Despite recent therapeutic advances, metastatic melanoma is a devastating disease. Stratification of primary melanoma patients by risk of developing metastatic disease is an ongoing clinical challenge. Relevant to this study, brain metastasis accounts for the majority of melanoma patient deaths, thus identification of early-stage melanoma patients who are at greatest risk of developing brain metastasis is of paramount importance. In this study, we identified a primary melanoma tissue-based miRNA expression signature that, in combination with an existing histopathology-based prognostic measure (AJCC stage), improves prediction of the development of brain metastasis. Moreover, we document that the source of miR-150–5p, a component of the described signature, is predominantly tumor infiltrating lymphocytes, adding to the emerging evidence linking immune response to patient outcomes in melanoma. A clinical assay derived from this signature could inform the clinical management of melanoma patients, including selection for more advanced staging, increased surveillance, or clinical trials of adjuvant therapies.