Uveal melanoma and the GEP test: Uveal (ocular) melanoma is the most common primary intraocular malignancy in adults. Unlike cutaneous melanoma, which is surgically resected, uveal melanoma is often treated with radiation while the tumor remains in the eye. A validated 15-gene expression profile (GEP) test - using SVM and weighted voting classification algorithms on fine needle aspiration biopsy (FNAB) specimens - classifies tumors as Class 1 (low metastatic risk) or Class 2 (high metastatic risk, approximately 50% 5-year metastatic death rate). This test directly guides systemic surveillance intensity and clinical trial enrollment for uveal melanoma patients.
The intratumoral heterogeneity question: The GEP test typically samples a single tumor site by FNAB, under the assumption that this single sample is representative of the entire tumor's transcriptional phenotype. However, intratumoral genetic and transcriptional heterogeneity has been demonstrated in other tumor types (notably glioblastoma, where multiple transcriptomic subtypes co-exist within a single tumor). For uveal melanoma - which grows as a single solid intraocular mass - the question of whether different sampling sites within the same tumor yield the same GEP class had not been systematically studied.
Study design: Researchers at the University of Cincinnati (Augsburger, Correa) prospectively sampled 80 uveal melanomas at three or more tumor sites by FNAB, submitting aspirates from two distinct sites to the Harbour laboratory for independent GEP testing and classification. The primary outcome was the frequency of discordant (different class at the two sampled sites) GEP results, using both the SVM algorithm (commercially deployed) and the weighted voting (WV) algorithm.
Key finding: Using the SVM algorithm, 9 of 80 cases (11.3%, 95% CI 9.0-13.6%) were clearly discordant - different GEP classes at the two sampled sites. If low-confidence classifications and one failed test are also counted as discordant, the rate rises to 13-15 of 80 cases (16.3-18.8%). Discordant tumors showed survival curves resembling Class 2 (high-risk) patients more than Class 1, suggesting that single-site Class 1 assignments in heterogeneous tumors may underestimate metastatic risk.
FNAB sampling procedure: All biopsies were performed by two experienced ocular oncologists (JJA, ZMC) using 25-gauge needles at three to four tumor sites selected based on tumor thickness, location, and clinical features. Aspirates from the second and fourth sampled sites were suspended in tissue culture medium, snap-frozen, and shipped overnight to the Harbour laboratory at Washington University for GEP testing. The remaining aspirates were submitted for cytopathological analysis to assess cell type (spindle, epithelioid, or mixed).
GEP testing - 15-gene assay: The GEP test analyzes RNA by microarray technology using 12 classifying genes (CDH1, ECM1, EIF1B, FXR1, HTR2B, ID2, LMCD1, LTA4H, MTUS1, RAB31, ROBO1, SATB1) and 3 control genes (MRPS21, RBM23, SAP130). Control genes were consistently expressed across all cases. A sample was considered a technical failure if any control gene or at least 3 of the 12 classifying genes were undetectable after 40 PCR cycles. In this series, 2 of 160 aspirates (1.25%) failed.
Classification algorithms: Two machine learning algorithms were applied to each GEP profile. The weighted voting (WV) algorithm produced a probability score from -1.0 (strong Class 1) to +1.0 (strong Class 2). The support vector machine (SVM) algorithm - favored by the test developer and used in commercial deployment (DecisionDx-UM) - produced a discriminant function score from -1.53 to +1.57 in this series. Both algorithms were independently applied to each of the two aspirates from each patient.
Concordance and discordance definitions: Concordant classification required the same binary class assignment (Class 1 or Class 2) for both the 'a' and 'b' specimens. Discordant classification was defined as different class assignments between the two sites, or a failed GEP test on one aspirate with a valid result on the other. Failed aspirates were assigned score 0.0 for both algorithms. Survival analysis used Kaplan-Meier curves to compare metastasis rates in concordant Class 1, concordant Class 2, and discordant subgroups.
Overall GEP class distribution: Using the SVM algorithm on the first-site aspirates, 52 tumors (65.0%) were Class 1, 26 (32.5%) were Class 2, and 2 failed. On the second-site aspirates: 55 tumors (68.8%) were Class 1 and 25 (31.3%) were Class 2. Using the WV algorithm: 53/80 (66.3%) were Class 1, 25 (31.3%) were Class 2, and 2 failed on first-site aspirates; 53 were again Class 1 and 27 (33.8%) were Class 2 on second-site aspirates. The aggregate class distributions were consistent across sites and algorithms.
Discordant cases by SVM algorithm: Each scatter plot (WV and SVM) showed 48 cases (60.0%) concordant Class 1, 21 cases (26.3%) concordant Class 2, and 9 cases (11.3%) clearly discordant. Among the 9 clearly discordant cases, 4 were Class 1 at site 'a' and Class 2 at site 'b', while 5 showed the reverse pattern. Interestingly, 2 cases were discordant by one algorithm and concordant by the other, highlighting that algorithm choice affects discordance detection.
Broader discordance estimate: If cases with low-confidence class assignments (plotted near the 0.0 threshold in either algorithm's scatter plot) and the 2 cases with a failed GEP test on one aspirate are also counted as discordant, the frequency rises to 13/80 (16.3%) by WV and 15/80 (18.8%) by SVM. The authors note that if a single-site biopsy had been performed in all cases, single-site sampling would have misclassified 7.5% of tumors compared to the two-site consensus result.
Survival impact of discordance: Kaplan-Meier curves for cumulative uveal melanoma metastasis and metastatic death showed, as expected, that concordant Class 2 patients had substantially higher metastatic rates than concordant Class 1 patients. Critically, the curve for the 9 clearly discordant cases closely resembled that of concordant Class 2 patients rather than Class 1 patients - suggesting that heterogeneous tumors with any site testing Class 2 carry Class 2-level risk, and that a single Class 1 result at one site does not exclude high-risk biology elsewhere in the tumor.
The misclassification risk in practice: Uveal melanoma patients classified as GEP Class 1 typically undergo less intensive surveillance for metastasis (liver imaging every 6-12 months versus every 3-6 months for Class 2). If 11.3% of tumors are heterogeneous and one site tests Class 1 while the other tests Class 2, single-site sampling will underclassify some high-risk patients, leading to inadequate surveillance and potentially delayed detection of metastatic disease at a stage when treatment is less effective.
The transformation versus heterogeneity confusion: The study provides a cautionary framework for interpreting sequential biopsies that yield different GEP classes. If a tumor is biopsied as Class 1, observed to grow, and then rebiopsied as Class 2, the reflexive clinical interpretation is Class 1 to Class 2 transformation over time. The authors argue that this apparent transformation may often reflect sampling heterogeneity at the initial biopsy rather than genuine tumor evolution - particularly since 3 rebiopsy cases in their experience showed stable concordant Class 1 classification despite tumor growth.
Two-site sampling recommendation: The study advocates for two-site FNAB of posterior uveal melanomas with independent GEP testing of each specimen. When both sites yield the same class, confidence in the classification is high. When sites are discordant, the authors recommend treating the patient as Class 2 (high-risk), consistent with the survival data showing discordant tumors behave like Class 2. The added time for two-site sampling is approximately 5 minutes, and cytology costs are similar for one or three specimens.
Intratumoral heterogeneity across tumor types: The finding of transcriptional heterogeneity in uveal melanoma mirrors similar findings in glioblastoma (where multiple transcriptomic subtypes co-exist within a single tumor) and has implications beyond this specific GEP test. Any molecular prognostic test based on a single biopsy from a heterogeneous tumor risks systematic misclassification. The uveal melanoma GEP test is unusual in specifically enabling feasible two-site sampling through the FNAB approach.
Cytopathological classification: Aspirates from one or more additional sites were submitted for cytopathological analysis by pathologists at Cincinnati Children's Hospital. The pathologist classified cells as malignant melanoma (epithelioid, spindle, or mixed), benign uveal nevus, or borderline. Immunocytochemistry with HMB-45 was used to confirm melanocytic origin in non-pigmented cells and borderline cases. In this series, 18.8% of aspirates were insufficiently cellular for cytopathological classification.
GEP failure rate and technical considerations: Of 160 aspirates (2 per patient), only 2 (1.25%) failed GEP testing - demonstrating the technical robustness of the assay even with small FNAB aspirates. The low failure rate reflects both the efficiency of the PCR-based RNA analysis and the careful aspiration technique that preserved adequate cell numbers. Failed aspirates (scored as class 0) contributed to the broader discordance estimate.
Tumor characteristics and discordance association: Analysis of whether tumor size, thickness, or other clinical features predicted GEP discordance showed only a non-significant trend - discordance occurred in 23.8% of thinner tumors (under 3.5mm) versus 4.3% of thicker tumors (above 7mm), possibly because thinner tumors sample a smaller proportion of total tumor mass per biopsy. No significant associations with patient age, largest basal diameter, or intraocular location were found.
FNAB complications in context: Of 53 patients treated with radioactive plaque implantation, 32 (60.4%) had some vitreous hemorrhage - most clearing spontaneously. Only 8 patients (15.1%) required pars plana vitrectomy for non-clearing hemorrhage. Two patients eventually required enucleation due to radiation vasculopathy complications rather than FNAB. The authors note no tract seeding (implantation tumor) complications in their experience, countering concerns about tumor cell dissemination from FNAB.
Prospective two-site GEP implementation: This study provides the evidentiary basis for adopting two-site FNAB as standard practice for uveal melanoma GEP testing. Prospective implementation at multiple centers, with tracking of metastasis outcomes, would provide outcome data to quantify whether two-site testing improves overall prognostic accuracy and reduces the rate of patients incorrectly managed as low-risk who subsequently develop early metastasis.
Understanding the molecular basis of intratumoral heterogeneity: The 11-16% discordance rate raises mechanistic questions about how a single continuous intraocular tumor develops spatially distinct transcriptional programs. Studies comparing chromosome 3 status (monosomy 3 is the strongest genomic predictor of Class 2 in uveal melanoma) at the two sampled sites could determine whether GEP discordance reflects heterogeneous chromosome 3 loss within the tumor or post-chromosomal transcriptional variation.
Comparison with genetic testing approaches: The GEP test competes with chromosome copy number analysis (FISH, SNP arrays, MLPA) as prognostic approaches for uveal melanoma. Parallel two-site sampling for both GEP and chromosomal analysis could determine whether GEP discordance correlates with chromosomal heterogeneity or is an independent phenomenon - providing mechanistic insight and potentially suggesting which test is more robust to sampling location variability.
Cost-effectiveness analysis: While two-site GEP testing would approximately double the cost of prognostic testing, the cost must be weighed against the downstream costs of inadequate surveillance for under-classified high-risk patients who develop metastatic disease. A formal cost-effectiveness analysis incorporating the 11.3% discordance rate, the Class 2 metastatic death rate, and the potential survival benefit from earlier metastasis detection would provide a rigorous economic argument for two-site sampling.