The original study that prompted the commentary was an ambitious exercise in machine learning applied to cancer genomics. Wu and colleagues set out to build a predictive biomarker for melanoma prognosis and immunotherapy response by exploiting a layer of biology that has attracted enormous attention in recent years: RNA modifications. Rather than looking only at which genes are expressed, the researchers focused on the regulators that govern how RNA molecules are chemically altered after transcription, a process known to influence tumor behavior, immune evasion, and treatment sensitivity. From their data they identified 84 regulators of RNA modification and subjected them to a rigorous feature-selection pipeline combining least absolute shrinkage and selection operator regression, commonly called LASSO, with multivariate Cox regression, a statistical technique designed to isolate the variables that independently predict survival outcomes.

The end product of that pipeline was a compact ten-gene signature, the RMODscore, which the authors validated in four independent immunotherapy datasets drawn from public repositories. The commentators acknowledge that this external validation using Gene Expression Omnibus cohorts demonstrates a degree of transportability to historical data, and they describe the results as a promising basis for biomarker development. But in their view, the leap from public datasets to the clinic is far larger than it might appear. Public repositories, they argue, are a poor proxy for the messy operational reality of hospitals, where genomic and clinical records are stored in-house under wildly varying conditions and where the practical performance of a risk score may diverge sharply from its performance on curated, preprocessed data.

The heart of the commentary is a detailed technical argument about why retrospective validation on shared public data is insufficient. Hospitals differ in their sequencing technology, in the preprocessing workflows applied to raw transcriptomic reads, in the therapeutic regimens their patients receive, in the demographic composition of their populations, in the methods used to censor follow-up data, and in their follow-up schedules. Each of these differences can degrade either the calibration of a predictive model, meaning the agreement between predicted and observed risks, or its discrimination, meaning its ability to separate patients who will do well from those who will not. A score that looks excellent when applied to harmonized public data may quietly lose accuracy when confronted with the idiosyncrasies of a single center’s laboratory and record-keeping practices.

Compounding the scientific problem is a legal and ethical one. Transcriptomic profiles, the commentators note, are sensitive and potentially re-identifiable. Even though gene expression data do not carry names or identification numbers, high-dimensional molecular profiles can sometimes be matched back to individuals, particularly when combined with other information. This re-identification risk limits the willingness of institutions to share patient-level transcriptomic data centrally, which in turn prevents researchers from evaluating otherwise promising signatures at the scale that clinical translation demands. The commentators frame this as a structural bottleneck: the very data needed to prove a biomarker’s worth across institutions are the data that institutions are increasingly reluctant, or legally unable, to pool.

Their proposed solution is a federated validation framework, an approach in which the model travels to the data rather than the other way around. Under this design, each participating melanoma center would keep patient-level data entirely within its own walls. Every center would apply a common data-harmonization protocol to ensure that the inputs to the model are comparable, compute the fixed RMODscore locally on its own patients, and then transmit only privacy-protected summary statistics to a coordinating hub. Those summary statistics would be sufficient to assess discrimination, calibration, clinical utility, and subgroup performance across the whole network, without any single patient’s molecular profile ever leaving its home institution. If the model needs recalibration or coefficient updating, secure aggregation techniques could be used to combine protected parameter updates from all centers into a single revised model.

Crucially, the commentators warn that federated learning alone does not eliminate privacy risk. Model updates themselves, such as gradients shared during training, can leak information about the individuals whose data produced them. To guard against this, they recommend layering differential privacy on top of the federated architecture, using techniques such as differentially private gradient clipping and calibrated noise addition. Differential privacy provides a mathematically quantifiable guarantee that the contribution of any single individual to the final output is bounded, but it comes at a cost: the noise that protects privacy can degrade predictive performance. The commentators insist that the two privacy parameters, epsilon and delta, which quantify the strength of the guarantee, should be reported explicitly in any such study, along with their measured impact on predictive performance and clinical utility. This transparency, they argue, is essential for the clinical community to judge the privacy-utility trade-off for itself.

The commentary goes further than most methodological critiques by prescribing an unusually detailed evaluation protocol. The authors propose a tiered comparison across four settings: local-only models trained and tested within each center, standard federated models, federated models augmented with differential privacy, and, where institutional governance permits, centrally pooled models that serve as a reference point for the maximum achievable performance. Performance metrics should be prespecified, they argue, and should include Harrell’s C-index for survival discrimination, time-dependent area under the curve, the integrated Brier score for prediction accuracy, calibration slope, and decision-curve net benefit, a measure of whether using the model to make decisions actually benefits patients. Beyond the headline metrics, they call for leave-one-center-out validation, in which each center is held out in turn, as a stronger test of geographic transportability than holding out a single selected center. Subgroup analyses by age, sex, disease stage, ancestry, and immunotherapy regimen should probe whether the score performs unevenly across patient populations, and privacy robustness should be stress-tested with established membership-inference and gradient-reconstruction attacks.

One of the most quietly important points in the commentary concerns transparency at the level of individual institutions. Aggregate performance estimates, the authors observe, can conceal poor performance in smaller participating centers. A federated study that reports only a network-wide C-index might mask the fact that the score fails badly at one or two sites, a failure that would matter enormously to the patients treated there. The commentators therefore call for confidence intervals and center-specific performance distributions to be reported alongside pooled results, ensuring that the burden of proof is met not just on average but everywhere the biomarker might be deployed.

The commentators marshal recent evidence to show that their proposal is practical rather than hypothetical. Sheller and colleagues demonstrated in 2020 that federated learning in medicine can achieve model quality close to that obtained through centralized data sharing, a finding that helped launch the field of privacy-preserving medical machine learning. More recently, Liu and colleagues in 2024 applied federated learning to predict treatment response in multicenter non-small cell lung cancer patients, and Ogier du Terrail and colleagues in 2023 used federated approaches to predict histological response to neoadjuvant chemotherapy in triple-negative breast cancer, publishing their results in Nature Medicine. On the privacy side, Wen and Li in 2025 investigated differential privacy in federated multi-omics survival analysis, directly relevant to the kind of transcriptomic risk models at issue here. Taken together, the commentators argue, these studies establish that a privacy-preserving federated assessment of the RMODscore is technically feasible and could determine whether the score remains reliable across genuine institutional differences while offering a governance-conscious pathway toward prospective clinical validation. Their conclusion is measured but firm: the RMODscore is a promising research biomarker that warrants further prospective evaluation, and multicenter federated validation, jointly assessing discrimination, calibration, clinical utility, subgroup performance, privacy risks, and the privacy-utility trade-off, should precede any claim of clinical readiness or routine implementation.

Subject of Research: Privacy-preserving federated validation of an RNA modification-based melanoma prognostic biomarker

Article Title: Comment on “Privacy-preserving multicenter validation of the RMODscore”

Article References: Sengan, S., Mamazulunov, N., & Khujamkulova, I. (2026). Comment on “Privacy-preserving multicenter validation of the RMODscore”. Journal of Cancer Research and Clinical Oncology, 152(9), Article 172. https://doi.org/10.1007/s00432-026-06610-w

Image Credits: AI Generated

DOI: 10.1007/s00432-026-06610-w

Keywords: melanoma, RMODscore, RNA modifications, federated learning, differential privacy, biomarker validation, immunotherapy response, machine learning, survival analysis, data privacy, oncology, multicenter studies