Machine learning model predicts melanoma prognosis using RNA modifications

Melanoma remains one of the most aggressive and treatment-resistant forms of human cancer, and although immune checkpoint inhibitors have transformed outcomes for a subset of patients, clinicians still lack reliable tools to predict who will benefit from these expensive and sometimes toxic therapies. A new study published in the Journal of Cancer Research and Clinical Oncology offers a potential advance: a machine learning–derived risk score built from genes that regulate RNA chemical modifications, which the researchers say can forecast both patient survival and the likelihood of response to immunotherapy. The work, led by Bailu Wu and Zhen Li of the First Affiliated Hospital of Zhengzhou University together with collaborators at Simcere Diagnostic Technology in Nanjing, distills the activity of dozens of RNA modification regulators into a compact ten-gene signature the authors call the RMODscore.

The biological premise underlying the study rests on a rapidly expanding field known as epitranscriptomics, the study of reversible chemical tags placed on RNA molecules after they are transcribed from DNA. The most famous of these tags, N6-methyladenosine or m6A, is deposited by writer enzymes, removed by erasers, and interpreted by reader proteins, collectively forming a regulatory layer that controls RNA stability, splicing, export, and translation. The researchers in this study did not limit themselves to m6A; they also included regulators of N1-methyladenosine (m1A), 5-methylcytosine (m5C), and 7-methylguanosine (m7G), assembling a comprehensive panel of 84 such regulators. Each of these modification types has been implicated individually in tumor initiation, progression, and immune evasion, but the authors argue that studying them in isolation misses the coordinated dysregulation that likely drives malignant behavior in melanoma.

To build their model, the team performed unsupervised clustering on the expression patterns of all 84 regulators across multiple melanoma cohorts, drawing heavily on data from The Cancer Genome Atlas (TCGA). This analysis revealed three distinct RNA modification subtypes of melanoma, and the clinical stakes of this molecular stratification became immediately apparent: survival differed significantly among the three groups, demonstrating that the collective activity of RNA modification machinery carries genuine prognostic information rather than mere molecular noise. Because clustering alone does not identify which individual genes matter most, the researchers next turned to weighted gene co-expression network analysis, or WGCNA, a technique that organizes thousands of genes into modules based on correlated expression patterns and then links those modules to clinical traits of interest. This step pinpointed one module, designated ME13, as most significantly associated with the RNA modification subtypes, providing a focused set of candidate genes for predictive modeling.

The core of the study is the construction of the RMODscore itself. Using least absolute shrinkage and selection operator regression, known as LASSO, together with multivariate Cox proportional hazards regression, the team compressed the candidate gene list into a ten-gene signature. LASSO regression is particularly well suited to this kind of problem because it penalizes model complexity, effectively shrinking the coefficients of less informative genes to zero and guarding against the overfitting that plagues many high-dimensional genomic studies. Multivariate Cox regression then ensured that each retained gene contributed independent prognostic information, adjusting for the influence of the others. The resulting score assigns each melanoma patient a continuous risk value computed from the weighted expression levels of the ten genes, with higher scores indicating a molecular profile associated with worse outcomes.

Validation was where the model earned its credibility. The RMODscore showed strong predictive performance not only in the TCGA discovery cohort but also, critically, in four independent melanoma immunotherapy datasets drawn from patients treated with checkpoint blockade. Across these cohorts, patients with high RMODscore values experienced significantly poorer survival than those with low scores, and the high-score group was consistently associated with clinical resistance to immunotherapy. The model’s ability to generalize across cohorts generated by different institutions and treatment protocols is an important benchmark, as many published genomic signatures fail precisely this test of external validation.

Mechanistically, the researchers found that the score divided melanoma tumors into biologically recognizable states. Tumors with high RMODscore values exhibited what oncologists describe as immune-cold phenotypes: they carried fewer infiltrating immune cells, showed dampened expression of immune activation signatures, and displayed upregulation of proliferation-related pathways that drive unchecked cell division. These are exactly the tumors that tend to shrug off checkpoint inhibitors, which work by unleashing pre-existing antitumor T cells and therefore require an inflamed tumor microenvironment to function. Conversely, tumors with low RMODscore values showed greater immune activation, the inflamed, T cell–rich milieu in which antibodies targeting PD-1 and CTLA-4 achieve their most durable responses. In other words, the RNA modification signature appears to capture, at the level of transcript regulation, the immunological architecture of the tumor that ultimately determines whether immunotherapy can succeed.

Beyond prognosis and immunotherapy prediction, the study ventured into the territory of drug repurposing. By correlating RMODscore values with pharmacogenomic data, the team found that the score predicted sensitivity to inhibitors of the ERK and JNK signaling pathways, both components of the mitogen-activated protein kinase cascade that is famously hyperactivated in melanoma through BRAF and NRAS mutations. ERK inhibitors are currently in clinical development as a strategy to overcome resistance to BRAF-targeted therapy, and JNK inhibitors have been explored in various oncology contexts. The suggestion that a high RMODscore might flag tumors susceptible to these drugs raises the possibility of using the score not merely as a passive prognostic marker but as an active guide to combination or sequenced treatment strategies, pairing immunotherapy with pathway inhibition in patients whose scores indicate a poor likelihood of checkpoint response.

The clinical significance of the work lies partly in what it adds to an increasingly crowded field of prognostic signatures for melanoma. Numerous gene expression models have been proposed over the past decade, including signatures based on immune genes, metabolic pathways, and broader multi-omics integrations, and several have advanced toward clinical use in assessing recurrence risk. What distinguishes the RMODscore approach is its grounding in RNA modification biology, a mechanistic layer that sits upstream of both tumor-intrinsic proliferation programs and tumor-immune crosstalk. Because RNA modification regulators are enzymes and binding proteins with defined activities, they represent not just biomarkers but potential therapeutic targets in their own right; dysregulated m6A machinery, for example, has already been shown in preclinical studies to influence PD-L1 expression and T cell-mediated killing. A score derived from this machinery could therefore track biology that is itself druggable.

The authors are appropriately measured in their conclusions, framing the RMODscore as a tool for prognostic stratification that may provide preliminary insights for future therapeutic exploration rather than a ready-made clinical test. Significant hurdles remain before such a signature could reach the clinic. The study relies on retrospective bulk transcriptomic data, which cannot resolve how RNA modification regulators behave in individual cell types within the tumor microenvironment; single-cell and spatial profiling would be needed to confirm the cellular origins of the signal. Prospective validation in randomized immunotherapy trials, standardization of the measurement assay, and demonstration that the score improves clinical decisions beyond established factors such as tumor stage, lactate dehydrogenase levels, and PD-L1 immunohistochemistry would all be required. Nevertheless, the consistency of the score across five independent cohorts and its dual performance in both survival prediction and immunotherapy response forecasting give it a stronger evidentiary footing than many signatures of its kind.

The research was funded by the Henan Province Natural Science Foundation Key Science Fund Project and the Central Plains Science and Technology Innovation Leadership Talent Program, and the corresponding author is Zhen Li of the Interventional Radiology Department at the First Affiliated Hospital of Zhengzhou University. Published as an open-access article under a Creative Commons license, the study arrives at a moment when the epitranscriptomics of cancer is moving from descriptive cataloging toward predictive, clinically actionable science. If subsequent studies confirm its performance, the RMODscore could join a growing arsenal of molecular tests that help clinicians decide which melanoma patients should receive immunotherapy first-line, which might benefit from earlier combination strategies, and which experimental agents targeting the RNA modification machinery itself deserve accelerated development.