Flawed Benchmarks and Biased Cohorts: Reassessing the Evidence for <scp>CP</scp> ‐ <scp>GEP</scp> (Merlin) Versus 31‐ <scp>GEP</scp> ( <scp>DecisionDx</scp> ‐Melanoma)

The role of genomic testing in guiding sentinel lymph node biopsy (SLNB) decisions for patients with cutaneous melanoma remains an area of active investigation and debate. In a recent article, Prieto et al. compared the performance of two molecular assays—clinicopathologic gene expression profile (CP-GEP or Merlin) and 31-gene expression profile (31-GEP)—for identifying patients at low risk for nodal metastasis who may consider forgoing SLNB [1]. While the authors address a clinically important topic, we believe that key methodological differences and selective reporting of metrics limit the interpretability of their conclusions. First, the authors selectively focus only on the false omission rate (FOR), which they refer to as the false negative rate (FNR). The FOR is calculated as one minus the negative predictive value (NPV), whereas the FNR is defined as one minus sensitivity. Despite the FOR's well-known and direct dependence on disease prevalence, the authors use it as the primary benchmark for assessing clinical utility. In doing so, they overlook other essential diagnostic parameters—such as sensitivity, specificity, and the proportion of low-risk predictions—that better characterize a diagnostic test and directly affect clinical decision-making. By evaluating CP-GEP and 31-GEP primarily through the FOR, without accounting for differences in test characteristics and study design, the comparison risks drawing biased conclusions. Second, the cohorts compared in the Prieto analysis differ substantially in their inclusion and exclusion criteria, which limits the validity of direct comparisons. To allow a more balanced evaluation, we focused on the largest, most representative study from each group: MERLIN_001 for CP-GEP (prospective, blinded; n = 1761) [2] and the Whitman study for 31-GEP (retrospective; n = 1674) [3]. The Whitman study includes 106 patients with T4 tumors and 416 who did not undergo SLNB—groups likely to skew sensitivity, specificity, and prevalence estimates. Including these patients inflates performance metrics and artificially lowers prevalence, thereby increasing the NPV. To address this, we re-analyzed the Whitman data by excluding SLNB-naïve and T4 patients based on published clinical details. This allowed us to reconstruct the confusion matrix (Table 1), recalculate sensitivity and specificity, and use the latter to derive FOR and PPV across a range of simulated prevalence values (Figure 1). We also retained 30 very low-risk T1a patients who underwent SLNB, despite their ambiguous classification status, giving them the benefit of the doubt. Notably, no comparable T1a patients without adverse features were included in the MERLIN_001 study. These adjustments were necessary to ensure a fair, harmonized comparison of diagnostic performance. In the re-analyzed Whitman dataset, only 4.9% of patients were classified as low risk by 31-GEP, compared to 37% in MERLIN_001—highlighting the greater potential of CP-GEP to reduce SLNB procedures. This broader low-risk classification did not compromise performance: across all simulated prevalence levels, MERLIN_001 consistently showed a lower FOR than 31-GEP (Figure 1), indicating a reduced risk of missing node-positive cases and supporting the superior clinical utility of CP-GEP. Moreover, CP-GEP has demonstrated stable sensitivity and specificity across multiple prospective validation studies. In contrast, 31-GEP and its i31-SLNB variant have shown variable performance and rely on post hoc modeling, which raises concerns about overfitting and limited generalizability. Prieto et al. assert that CP-GEP offers no improvement over the American Joint Committee on Cancer (AJCC) staging [1], a claim not supported by the available evidence. In our multivariable analyses, the Merlin assay consistently outperformed AJCC staging and established nomograms from MSKCC and MIA within T-stage subgroups. While Zakria et al. [4] reported that i31-GEP surpassed the MIA nomogram, this conclusion was not sustained when appropriate statistical methods were applied. Lo et al. [5] later demonstrated that the reported advantage of 31-GEP over standard of care disappears once confidence intervals are correctly calculated. In summary, caution is warranted when interpreting cross-study comparisons. Harmonized cohorts, comprehensive metrics, and transparent study design are essential for valid conclusions. D.B., C.G., and J.D. are employees of SkylineDx. A.M. has received research funding from SkylineDx and is listed as an inventor on skin cancer-related technologies owned by Mayo Clinic. C.P. has provided consulting services to SkylineDx. A.P. declares no conflicts of interest. The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.