The scale of this activity is striking. Using reproducible PubMed searches covering January 2019 through June 2026, the authors retrieved 4,416 records matching their query for public-omics or bioinformatic secondary analysis in gastric cancer, and 1,812 records for prognostic-signature and risk-model studies. Searches targeting single-cell and spatial gastric cancer omics returned 876 and 148 records respectively, while artificial intelligence and machine-learning prognosis studies returned 636.

The authors are careful to frame these counts as query-dependent descriptive signals rather than systematic prevalence estimates, and they explicitly state that the scans were not formal quality appraisals of any individual study. Because PubMed continues to index records within the fixed date window, re-running the same queries at the revision date retrieved slightly more records — 4,460, 1,840, 904, 159, and 662 respectively — and both sets of counts, with the complete PMID-level record sets behind them, are reported in the supplementary materials. The review catalogs the resource landscape in detail.

TCGA-STAD remains the most frequently reused multi-layer resource, linking gene expression, mutation, copy-number alteration, DNA methylation, and clinical annotation, and underpinning the widely cited molecular classification into Epstein-Barr virus-positive, microsatellite instability, genomically stable, and chromosomal instability subtypes. The Gene Expression Omnibus and related repositories provide many independent gastric cancer datasets for external replication and cross-platform comparison, though with heterogeneous preprocessing and metadata quality. The Asian Cancer Research Group cohorts add a clinically informative subtype framework, while ICGC and cBioPortal support exploratory cross-study browsing of mutations, copy-number changes, and survival associations.

Proteogenomic resources add protein-level triangulation, and single-cell repositories resolve malignant epithelial states, fibroblast subsets, myeloid programs, and T-cell exhaustion at cellular resolution. Spatial platforms such as Visium, Xenium, CosMx, MERFISH, and Stereo-seq preserve tissue architecture, allowing analysts to test whether bulk-derived signals are anatomically plausible. Prognostic modeling serves as the review's central stress test because it is common, clinically attractive, and methodologically fragile.

The authors dissect a recognizable template: choose a biological theme such as hypoxia, pyroptosis, or metabolism; collect candidate genes; screen features by univariate Cox regression; select variables with LASSO-Cox or another machine-learning method; construct a risk score; divide patients into high- and low-risk groups; and report Kaplan-Meier curves, time-dependent area under the receiver operating characteristic curve, concordance indices, and nomograms. Each checkpoint can introduce optimism. If feature filtering occurs before train-test separation, validation information leaks into model development.

If cutoffs are chosen after inspecting survival separation, the resulting curves exaggerate performance. If the external cohort is used to re-estimate coefficients, reselect features, or recalibrate thresholds, validation quietly becomes redevelopment. The statistical hazards run deeper than pipeline hygiene.

Low events per variable increase overfitting in high-dimensional omics data, where feature selection is unstable and batch effects can make a gene appear prognostic on one platform but not another. The proportional-hazards assumption underlying Cox models is frequently untested in signature studies. Discrimination, calibration, and clinical utility are separate properties: a high AUC does not guarantee accurate risk estimation, and accurate risk estimation does not guarantee decision-relevant benefit.

Publication bias then preferentially displays successful signatures while failed replications remain unpublished, creating what the authors describe as discovery abundance without cumulative knowledge. Public cohorts also differ in survival endpoints, censoring mechanisms, follow-up duration, and treatment exposure, so validation should apply the same endpoint definition whenever possible rather than treating all survival variables as interchangeable. A title-and-abstract screen of the 100 most recent records returned by the prognostic-signature query quantifies some of these reporting gaps.

External validation was mentioned in 46 of 100 records, calibration in 36, decision-curve or net-benefit analysis in 33, clinical comparators in 67, and experimental validation in 30. Code availability was mentioned in zero of 100 — though the authors caution this is an abstract-level lower bound, since code statements typically appear in full-text declarations rather than abstracts. The screen was coded by a single author without duplicate coding, and the authors stress that these are illustrative text-mining signals about what titles and abstracts mention, not evidence that the corresponding analyses were performed adequately.

They also note that the screened corpus represents broad prognosis literature in gastric cancer, including clinical, imaging, and cost-effectiveness studies, rather than the prevalence of methodological practices within public-omics research specifically. Beyond prognostic models, the review scrutinizes the interpretive shortcuts that pervade secondary analysis. Immune-deconvolution tools such as ESTIMATE, CIBERSORT, xCell, and TIMER produce model-based estimates, not direct measurements, and different tools can yield different estimates from the same data because they use different reference signatures and assumptions.

An inferred T-cell fraction does not reveal exhaustion, spatial exclusion, or suppressive myeloid context, and checkpoint expression does not equal checkpoint blockade response. Weighted gene co-expression network analysis identifies modules, but correlation does not establish regulation. Competing endogenous RNA networks built from expression correlations and target databases rarely establish stoichiometry, subcellular localization, or direct binding.

Cell-cell communication tools infer potential interactions from ligand-receptor co-expression but do not measure signaling. Single-cell reanalysis carries its own pitfalls: dissociation bias, ambient RNA, doublets, subjective annotation, and pseudoreplication, where thousands of cells are treated as independent biological replicates despite coming from few patients. To convert these observations into practice, the authors propose GC-PROVE, a five-domain appraisal framework covering problem anchoring, resource provenance, omics workflow transparency, validation hierarchy, and evidence restraint.

Problem anchoring demands that an analysis begin with a clinical or biological question rather than a fashionable gene set. Resource provenance requires auditable accessions, sample filters, endpoint definitions, and cohort-overlap checks — particularly important because the same patients can appear across TCGA and GEO repositories. Validation hierarchy orders evidence from public discovery through external replication, tissue or protein support, and functional perturbation to clinical utility, so that the strength of a claim matches the evidence behind it.

The authors emphasize that GC-PROVE is an author-proposed heuristic with no validation study or inter-rater reliability estimates, functioning as a screening and routing layer that points to fuller standards such as TRIPOD+AI and PROBAST rather than replacing them. Five published case studies demonstrate how the framework produces traceable judgments. A 2023 mitochondrial-gene risk score by Chang and colleagues anchors a stated biological question, reports a locked LASSO-Cox formula, applies it to two independent external cohorts, and reports calibration and decision-curve analysis, scoring favorably on most domains, though it is not compared against a molecular subtype or full clinical model.

By contrast, a 2025 study by Jiang validates through a random training-testing split of the same TCGA cohort rather than an independent cohort, placing it at the discovery end of the hierarchy where hypothesis prioritization, not external validation, is the appropriate claim. A 2025 deep-learning model by Wang integrating TCGA-STAD, CPTAC, and single-cell resources reports only internal held-out validation, with no calibration, decision-curve analysis, or clinical comparator. The authors also map how far genuine clinical biomarkers sit from public-omics candidates: HER2 amplification, MSI-high status, PD-L1 combined positive score, and Claudin 18.2 each require defined assays, cutoffs, and regulatory approval, none of which can be inferred from a database association alone.

The review closes with a forward-looking agenda for 2026 through 2030. Single-cell data should identify the cellular origin of bulk signals; spatial data should test anatomical plausibility; and foundation models such as Geneformer, scGPT, and scFoundation should be benchmarked against transparent, task-matched baselines on data not used during pretraining — a caveat the authors consider essential because pretraining corpora may overlap evaluation atlases, inflating apparent transferability. Perturbation-aware approaches, including Perturb-seq and CRISPR screens, should be treated as hypothesis-prioritization systems requiring experimental confirmation.

The authors argue that unsuccessful external validation and failed replication are legitimate results that can redirect attention to better-supported biology, and that public-omics secondary analyses are most valuable when they make the next experiment, validation cohort, or clinical question more precise. For a field producing thousands of statistically significant survival curves, the message is blunt: credibility, reproducibility, and clinical anchoring — not signature count — should define progress.

Subject of Research: Methodological quality of public omics data reuse in gastric cancer research

Article Title: Public omics data reuse in gastric cancer: from discovery abundance to translational evidence

Article References: Su, S., Huang, C., Yang, B., Huang, Y., Yang, P., & Liu, A. (2026). Public omics data reuse in gastric cancer: from discovery abundance to translational evidence. Frontiers in Oncology, 16, Article 1974260. https://doi.org/10.3389/fonc.2026.1974260

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1974260

Keywords: gastric cancer, public omics data, TCGA-STAD, prognostic signatures, GC-PROVE, data leakage, overfitting, single-cell RNA sequencing, spatial transcriptomics, external validation, machine learning, reproducibility