Background
Cutaneous melanoma continues to be a significant contributor to cancer morbidity and mortality, with over 90,000 new cases and over 9000 deaths expected in 2018 [
1]. Assessment of survival outcomes is based on the American Joint Committee on Cancer (AJCC) staging [
2]. Stage I and II patients greatly outnumber later stage patients, thus the vast majority of melanoma-related deaths occur in patients belonging to this group at diagnosis [
3]. In the Multicenter Selective Lymphadenectomy Trial (MSLT-1), 13% of node-negative patients had biologically aggressive disease that resulted in metastases and death [
3,
4]. The fact that a substantial proportion of melanoma related deaths occur in patients with thin, T1, melanoma tumors has also been reported [
5‐
7]. Based on current guidelines these patients do not receive the intensive surveillance or adjuvant therapy offered to AJCC high risk patients [
8]. Recent advances in our understanding of tumor biology should enable us to identify high-risk disease based on molecular characteristics of the tumor [
9‐
11].
A 31-gene expression profile (GEP) test that dichotomizes cutaneous melanoma patients as Class 1 (low-risk) or Class 2 (high-risk) has been previously described [
12,
13]. Class 2 results are associated with an increased risk for metastatic disease that is independent of staging factors [
12]. This study evaluates the GEP test in a previously unreported, independent cohort of 523 cutaneous melanoma cases from a multi-center consortium.
Methods
Cohort selection
Following institutional review board approval of the study and waiver of patient consent at each of the 16 participating centers, archival formalin-fixed, paraffin-embedded primary cutaneous melanoma tumor tissue was collected. Inclusion in the study required biopsy confirmed stage I–III cutaneous melanoma diagnosed between 2000 and 2014, with at least 5 years of follow-up, unless there was an earlier documented recurrence or metastatic event. Thus, all cases diagnosed after October 31, 2011 that were included in the study had a documented metastatic event. All cases included in the study that had no documented metastasis event had at least 5 years of follow-up. Clinical, pathological and outcome data were collected by collaborating centers through an electronic case report form, and on-site monitoring of each case was completed prior to data analysis with a censor date of October 31, 2016.
Data collection and class assignment
Expression profiling of the 31 genes (28 class-discriminating and 3 endogenous control genes; Additional file
1: Table S1) was performed via RT-PCR and radial basis machine (RBM) predictive modeling was used to generate a probability score and subsequent class assignment (Class 1 or Class 2) for each sample, as previously described [
12,
13]. Only cases that met preestablished pre- and post-analytic quality control thresholds were included (Table
1).
Table 1
Clinical characteristics of the cohort
Median age (range), years | 59 (18–92) |
Median follow-up for patients without a metastatic event, years (range) | 7.5 (5.0–16.5) |
Recurrence/distant metastasis | 142/111 |
Median time to first recurrence, years (range) | 1.2 (0.0–10.0) |
AJCC stage | |
I (total) | 264 (50%) |
IA | 108 |
IB | 76 |
Unknown substagea | 80 |
II (total) | 93 (18%) |
IIA | 35 |
IIB | 26 |
IIC | 17 |
Unknown substage | 15 |
III (total) | 166 (32%) |
IIIA | 69 |
IIIB | 57 |
IIIC | 35 |
Unknown substage | 5 |
Breslow thickness | |
Median (range), mm | 1.2 (0.1–29.0) |
≤ 1 mm | 223 (43%) |
> 1 mm | 296 (56%) |
Unreported | 4 (1%) |
Mitotic index | |
< 1/mm2 | 99 (19%) |
≥ 1/mm2 | 240 (46%) |
Unreported | 184 (35%) |
Ulceration | |
Absent | 309 (59%) |
Present | 133 (26%) |
Unreported | 81 (15%) |
SLN status | |
Untested | 186 (36%) |
Negative | 180 (34%) |
Positive | 157 (30%) |
GEP Class | |
Class 1 | 314 (60%) |
Class 2 | 209 (40%) |
The RBM model generates a linear probability score from 0 to 1. Within the model, cases with a probability score between 0 and 0.49 are labeled Class 1, with samples within one standard deviation (SD) of the median probability score for Class 1 cases (0–0.41) designated as Class 1A and samples outside of the SD (0.42–0.49) designated as Class 1B (Additional file
2: Supplemental methods). Similarly, Class 2 cases have a score between 0.5 and 1. Samples with a probability score within one SD of the median (0.59–1) are classified as Class 2B, while those with a score outside the SD (0.5–0.58) are labeled Class 2A. In both the Class 1 and Class 2 groups, “A” subclass reflects a better and “B” reflects a worse prognosis within the Class. Results from subclass analysis are reported in the clinical setting.
Primary endpoints were recurrence-free survival (RFS), or time from diagnosis to any local, regional, or distant recurrence, excluding a positive SLN, and distant metastasis-free survival (DMFS), or time from diagnosis to any distant metastasis. Melanoma-specific survival (MSS), or time from diagnosis to death documented as resulting from melanoma, was a secondary endpoint. All survival variables were calculated from documented diagnosis and event (or censor) dates.
Statistical analysis
Kaplan-Meier and Cox proportional hazards survival analyses were performed using R version 3.3.0, with P < 0.05 considered statistically significant by log-rank method or Cox regression analysis. For proportional hazards analysis, Breslow thickness was measured as a continuous variable, while all other factors were dichotomized.
Discussion
The use of molecular classification of disease is now routine in clinical practice [
10,
14]. For any new molecular clinical test it is critical to evaluate whether the test i) accurately predicts its intended outcome; ii) has consistent, sustainable accuracy across multiple independent studies, and iii) adds value beyond existing clinical tools [
15‐
17]. Here we report that the 31-gene expression profile test is able to predict metastatic risk in an independent cohort of 523 melanoma patients with results that are consistent with those reported in prior studies [
12,
13]. In this cohort, we observed a 5-year DMFS rate of 93% for Class 1 cases and 62% for Class 2 cases (compared to 100% and 58%, respectively, in the smaller, initial study). We previously reported that this test could identify the majority of SLN-negative patients with an elevated risk of metastasis [
12]. In this study, the majority (70%) of the node-negative patients who had a distant metastasis were Class 2, as well as the majority (78%) of SLN-negative patients who died from melanoma (7 of 9 patients).
This study is based on a cohort of melanoma patients with clinical characteristics that align with those of the general cutaneous melanoma population. While the SLN positivity rate is higher than the 15–20% reported in previous studies, the 5-year survival rates for the SLN-negative and SLN-positive groups (95% vs. 75%, respectively) are similar to those reported in the MSLT-1 study (90% vs. 70%, respectively) [
3,
4]. Breslow thickness, ulceration and mitotic rate were all important in univariate models of risk prediction (Table
3), supporting similarity with previous cohorts used to identify relevant staging factors. SLN status is currently regarded as the gold standard for prognosticating cutaneous melanoma, as a positive SLNB is associated with a significantly increased risk of metastasis [
4] and our results confirm this. Compared to the SLNB procedure, the GEP test performed with better sensitivity across all endpoints studied. The results suggest that the GEP could enhance current prognostic accuracy by identifying clinically and pathologically SLN-negative patients who harbor an elevated risk of metastasis. Thus, highest sensitivity for detecting patients at high risk for recurrence, distant metastasis or melanoma-specific death can be achieved when the test is used in combination with current staging criteria. Importantly, this is coupled with high negative predictive values across endpoints, reflecting a substantially low risk associated with the Class 1 result. While the positive predictive values are lower, this accuracy metric may be impacted by 1) a favorable host immune response to metastatic tumor cells; and 2) follow-up time that is not long enough to observe the metastatic event. Importantly, the positive predictive values observed for the GEP are similar to those observed SLN status in this cohort (Table
5).
Considering that approximately two thirds of melanoma-related deaths in patients originally diagnosed without distant metastatic disease (stage I-III) occur in SLN negative patients (stage I-II) [
3], the identification of patients in this group with biologically aggressive disease is a clinically significant unmet need. The current study demonstrates that implementing the GEP test after initial staging of melanoma tumors adds value by further stratifying the risk associated with stage I and stage II patients. That value is illustrated by a risk of recurrence that is three times higher for the stage I/Class 2 group compared to the stage I/Class 1 group (15% vs. 5%), and nearly seven times higher when comparing the stage I/Class 2B group to the stage I/Class 1A group (27% vs. 4%). The stage II/Class 2 group has nearly twice the risk of recurrence compared to the stage II/Class 1 group (49% vs. 28%), however, it should be noted that five of the nine events in the Class 1 group were regional recurrences. By comparison, the stage II/Class 2 group has three times the risk of developing distant metastasis compared to the stage II/Class 1 group (43% vs. 13%) and five times the risk in the stage II/Class 2B group compared to the stage II/Class 1A group (47% vs. 9%). The ability to subdivide stage II patients into groups with as high as 43% chance of developing distant metastasis and alternatively groups with as low as 5% risk at 5-years could significantly impact management decisions and clinical care. The results suggest that the GEP offers the opportunity to personalize risk assessment within each of these population-based AJCC stages.
The identification of high risk early stage patients is especially relevant considering current advances in melanoma therapies, which require us to improve risk evaluation in order to better weigh benefit versus harm of adjuvant therapy [
18]. These findings suggest that new tools are necessary to supplement current staging approaches, even as we achieve better outcomes for melanoma patients overall. Early stage patients could potentially benefit from adjuvant therapy but may not be recognized as high risk by the current staging system, and even among stage III patients there is often a dilemma as to whether systemic treatment is appropriate. The results of this study suggest that the GEP test should be evaluated in the context of new adjuvant therapy trials and trials evaluating the benefit of management approaches in stage III patients.
One of the limitations of this study is the inclusion of samples in the cohort that were diagnosed prior to widespread standardization of reporting for pathological variables such as Breslow thickness, ulceration and mitosis and therefore some pathology reports did not specify all features. However, the Cox regression models assessing the association between GEP and those factors account for this limitation and only patients with all factors specified were included in this analysis. Another limitation is the retrospective nature of the study and thus does not take into account recent advances in management of patients with advanced melanoma in the adjuvant and metastatic settings. However, recently published results of an interim analysis of the GEP test in a prospective cohort show consistency of results with this another retrospective cohorts [
12,
13,
19].
Current guidelines indicate that management should ultimately be tailored to an individual’s probability of recurrence [
20]. The risk classification provided by this test, along with current prognostic factors, can be used to better estimate an individual’s risk for recurrence and therefore aid in determining the most appropriate surveillance methodology and frequency. As illustrated in Fig.
4, the clinical utility of the test in conjunction with SLNB can identify as many as 89% of the patients who will experience a distant metastasis, and over 70% of those patients who are SLNB-negative. Several recent studies have demonstrated that modern therapies for melanoma are more effective when disease burden is low [
21,
22]. Thus, the need to accurately predict risk in melanoma patients is more critical than ever to enable risk-tailored surveillance and management of early staged patients with biologically aggressive tumors.
Acknowledgements
The authors wish to thank the physicians and clinical staff at each of the contributing institutions, and recognize the significant contributions to clinical data review and tissue processing by Jeff Wilkinson, PhD, Clare Johnson, RN, Natalie Lassen, PhD, John Stone, PhD, and Kristen Oelschlager, RN.