Background
Uveal melanoma (UM) is a rare, intraocular cancer that affects approximately 1600 patients per year in the United States [
1]. Most patients are treated with eye-sparing radiotherapy, primarily through plaque brachytherapy or proton beam radiation, while only a small proportion (~10%) will undergo enucleation. Despite the high rate of primary tumor control, ~50% of patients will develop metastatic disease, primarily to the liver, after which prognosis is poor [
2]. While the clinicopathologic features assessed in AJCC staging, including tumor size, ciliary body involvement, and extraocular extension, are important factors when assessing metastatic risk, even Stage I-II patients have a 12-30% UM-related mortality rate by 10 years [
3,
4]. Because of this, frequent surveillance was generally recommended to monitor all UM patients for disease spread [
5,
6].
Gene expression profiles that are associated with low-risk (Class 1A), intermediate-risk (Class 1B), or high-risk (Class 2) outcomes have been shown to provide information useful for risk-tailored surveillance plans [
7‐
9]. Because treatment of the primary tumor is highly effective with plaque radiotherapy, enucleation is not common, and the amount of tumor tissue available for molecular prognostic testing is limited. During the development of the gene expression profiling test for UM prognostication (also known as DecisionDx®-UM), focus was placed on providing patients with a test that was robust and could be run successfully and reproducibly on a very small amount of tissue obtained through a fine needle aspirate biopsy (FNAB), as well as formalin-fixed paraffin embedded (FFPE) tissue from enucleations [
8‐
10].
Recommendations for the development of clinically useful molecular biomarkers have suggested that a biomarker test needs to demonstrate clinical validity, utility, and analytic validity, and that supporting data for each must be transparent and readily available for both physicians and patients [
11]. Several widely-used genomic tests for different types of cancer have achieved high levels of evidence for each of these criteria [
12‐
17]. The clinical validity and clinical utility of the 15-gene expression profile test for UM was reported by the Cooperative Ocular Oncology Group (COOG) [
9], and the prognostic accuracy of the test has been confirmed in multiple single- and multi-center studies [
18‐
23]. The test’s impact on clinical decision-making for UM patients has also been demonstrated [
23,
24]. The focus of this study was to evaluate performance metrics of the test in a CLIA-certified laboratory setting, describe the rate of technical success on both FNAB and FFPE tissue from enucleations, and report correlations with pathological variables.
Discussion
Given the advances in molecular cancer diagnostics in the last decade, recommendations have been established to guide careful, thorough assessments of these tests [
11,
27]. One of the major criteria for a clinically useful, well-validated assay is that the test must demonstrate analytic validity, which includes reliability and reproducibility of the test to measure the intended molecular analytes [
11,
27]. In this study, we report on the analytic validity of the 15-gene expression profile test (DecisionDx-UM) as performed in a CLIA-certified clinical laboratory. Reproducibility and reliability were demonstrated through i) inter-assay concordance, ii) inter-assay concordance of samples used for proficiency testing and instrument verification throughout one year, iii) intra-assay concordance, iv) inter-lab concordance between two CLIA-certified laboratories, and v) inter-instrument/operator concordance. Molecular classifications by DecisionDx-UM were 100% concordant for four of these assessments, with only 3 discordant specimens out of the total 143 samples in the analytic experiments. The reported discriminant scores for the same tumor sample were highly correlated, as reflected by R
2 values above 0.96 for the inter-assay, intra-assay, inter-instrument, and inter-lab experiments. Furthermore, the run-to-run, lab-to-lab, instrument-instrument, and intra-run variabilities between discriminant scores were all within acceptable limits that would not impact patient care. These results demonstrate that the DecisionDx-UM class assignment for the same tumor specimen is consistent within the same PCR card and when run on separate days in different PCR cards, even with several weeks in between runs. To our knowledge, this is the only report of analytic validity for any prognostic test for UM. Comparable analytic validation has not yet been reported for fluorescence in situ hybridization (FISH), microsatellite analysis (MSA), single nucleotide polymorphism (SNP) arrays, or multiplex-ligation probe amplification (MLPA), all of which can be used to detect monosomy 3 and other chromosomal copy number changes that have been associated with UM metastasis, or next-generation sequencing to detect mutations and chromosomal aberrations to estimate UM prognosis.
Gene expression profiling is one of the clinically significant variables for disease prognostication recommended by the 2017 AJCC staging guidelines [
28], and as such, the DecisionDx-UM test is routinely used across the United States and has been clinically available since 2010. The prognostic test was developed after two distinct molecular subtypes of UM were identified and shown to have correlation with outcomes [
7,
10,
29,
30]. A 15-gene RT-PCR test was developed to identify these UM molecular subtypes and prospectively validated in multiple studies [
8,
9,
20,
21,
23]. As eye-sparing treatments are frequently utilized in the contemporary management of UM and enucleations are less common, the standard practice for physicians utilizing DecisionDx-UM is to collect a biopsy prior to or at the time of radioactive plaque or proton beam clip placement. Due to the lack of residual tissue aside from that obtained during biopsy, successful gene expression profiling on a single biopsy is critical. While the technical success of the 15-gene expression profile assay has been previously reported as ranging from 95 to 99% [
8,
9,
19], many of the samples in those studies were tested in the research laboratory that developed the assay, which was then licensed by Castle Biosciences, Inc. in 2009. In this report, we demonstrate that 96.4% of 5516 samples that have been clinically tested at Castle Biosciences’ laboratory generated successful molecular classification reports, establishing the consistency of technical success from the test. Of the unsuccessful tests, 39 out of 201 (19%) were due to samples outside of quality control specifications (for FNABs, incorrect volume and/or not frozen,
n = 18; for FFPE tissues, insufficient tumor volume,
n = 21). Thawed samples and those with insufficient/excessive volume have been previously reported to be associated with technical failures of gene expression profiling [
8]. The technical success rate of DecisionDx-UM testing is substantially higher than what has been previously reported for other molecular methods used in UM prognostication, including FISH, MSA, and array CGH, which range from 50 to 87% in much smaller sample sets [
31‐
34].
Of the 5315 successful tests reported, 43.4% were Class 1A, 22.4% were Class 1B, and 34.2% were Class 2. Overall, these proportions are similar to class those reported in the Cooperative Ocular Oncology Group (COOG) study, which were 47% Class 1A, 13%, Class 1B and 40% Class 2 [
9]. Likewise, other single- and multi-center studies that analyzed subsets of the clinically tested patients reported here have shown similar proportions of Class 1A, 1B, and 2 results [
21,
23,
35].
Of the enucleation specimens tested for which there were associated pathology data, the proportion of Class 1 patients shifted to be predominately Class 1B vs. Class 1A, and the proportion of Class 2 tumors was also increased compared to the total clinical population. Given the significantly increased LBDs and thicknesses of enucleation specimens (i.e. those that necessitated removal of the globe), an increase in Class 2 and Class 1B tumor classification is not unexpected. Importantly, there were no significant differences between Class 1A and 1B tumors in terms of cell morphology, ciliary body involvement, or extraocular extension, underscoring the utility of molecular testing to delineate risk in these tumors that otherwise share similar pathologic features. Overall, the pathology and molecular class data in our clinically tested cohort reflect published reports that greater LBD and tumor thickness tend to be clinicopathologic features associated with more aggressive tumors [
9,
20,
22,
23], and these riskier phenotypes are most frequently seen in Class 1B or Class 2 tumors. An advantage of gene expression profiling is that it reflects objective, intrinsic tumor biology, whereas measurements of LBD in particular can be subjective due to variation between observers and in techniques used for size measurement [
36]. Additionally, cytopathologic analysis can be impaired by a high rate of insufficient cellularity from FNABs [
19]. Several studies have shown that GEP is the most significant independent prognostic factor for metastatic risk when compared to clinicopathologic features, including LBD [
9,
18‐
20,
22,
23,
37].