Paper Review — Verify Claims Fast

Key numeric summary (visual-first)

• Cohorts: OA (≥65) n=108; YA (18–65) n=154 — TCGA‑SARC bulk RNA‑seq.
• DEA thresholds: |log2FC| ≥1.5 and FDR <0.005 → 733 sig‑DEGs (197 up, 536 down in OA).
• Downstream: 10 sig‑TFs (DoRothEA/TRRUST filtered), top 10 network hubs (CytoHubba), 16 genes prognostic in OA (univariate Cox p≤0.05).

All claims in the paper derive from the preprint dataset and analysis pipeline described by the authors Nagarajan et al. preprint (2025)

Strengths (concise, evidence-based)

1. Clear use of public TCGA data and a standard pipeline (TCGAbiolinks/edgeR) — enables potential reproducibility if code + processed data are shared Methods reported in preprint.
2. Multiple complementary downstream analyses — functional enrichment, TF‑target filtering (DoRothEA/TRRUST), network hubs (STRING/CytoHubba), and survival analyses — giving a layered biological story rather than single‑test reporting.
3. Focus on an important clinical subgroup (older sarcoma patients) that is underrepresented in clinical studies; linking age with transcriptomic features is valuable clinically and mechanistically.

Major concerns and methodological critiques

1. Heterogeneity (histologic subtypes & sample composition): TCGA‑SARC contains multiple sarcoma histologies with divergent biology (leiomyosarcoma, liposarcoma, undifferentiated pleomorphic sarcoma, etc.). Age distribution often correlates with subtype and stage; the paper does not present stratified analyses controlling for histologic subtype, tumor stage/grade, or primary site in the main DEA/survival models. Without adjusting, observed age‑associated DEGs may reflect subtype composition shifts between OA and YA rather than true age effects. The manuscript notes subtype mix in supplement but does not properly adjust the DEA/design matrix for subtype covariates — a strong potential confounder Preprint methods and supplement.
2. DEA thresholding and effect sizes: Authors use a hard log2FC cutoff of ±1.5 with FDR<0.005. That is stringent and can miss biologically important smaller effects, and they do not present volcano with labelled effect sizes or report numbers of genes removed by filtering. More critically, there is no explicit modeling of potential batch effects (e.g., sequencing center, library protocol) or inclusion of surrogate variables — TCGA data commonly require batch correction/adjustment (TCGAbiolinks helps but must be explicitly reported) TCGAbiolinks reference / best practice.
3. Survival analysis design and multiple testing: The authors ran univariate Cox for ~477 genes (after removing 256 genes with skewed medians) and reported 31 with p≤0.05 and 16 deemed directionally meaningful. However, they do not report correction for multiple testing (FDR) across Cox tests and appear to rely on p≤0.05 — inflating false positives. Also, dichotomizing expression at median loses power and assumes bimodality; continuous Cox models (with assessment for linearity) or penalized multivariate models would be more robust. The paper acknowledges some genes have contradictory literature in other cancers, which increases the need for stringent correction and external validation Preprint survival analysis description.
4. Lack of independent validation: Important prognostic calls (16 genes) and novel lncRNA findings are not validated in any independent sarcoma datasets (e.g., other TCGA splits, GEO, or the recent large sarcoma compendia). Independent validation across cohorts or cross‑validation would substantially increase credibility. The paper states code/data on GitHub but does not present replication; this is the single largest practical weakness.
5. Biological interpretation overreach: The authors link OA DEGs to EMT, Wnt/β‑catenin, calcium signaling, MMP/ECM, and immune changes. While many of these pathways appear in enrichment outputs, pathway overinterpretation is common with bulk enrichment results: enrichment can be driven by cell-type composition differences (e.g., immune/stromal fraction) rather than tumor‑intrinsic signaling. The study does not present deconvolution (CIBERSORT, MCPcounter, or xCell) to quantify immune/stromal fractions between OA and YA groups, which would be necessary to separate composition from cell‑intrinsic regulation (and is recommended in bulk‑RNA studies) TME deconvolution best-practice.

Concrete recommendations to improve the manuscript (short action list)

1. Recompute DEA with a design matrix that includes histologic subtype, tumor purity / ESTIMATE immune/stromal scores, stage/grade, and any sequencing batch variables; report adjusted results and show that age signal remains.
2. Perform immune/stroma deconvolution (e.g., xCell, CIBERSORTx, MCPcounter) and show whether pathway enrichments persist after adjusting for cell‑type fractions.
3. For survival associations: run continuous Cox models (expression per SD) with multivariate adjustment (age as continuous, sex, stage, subtype), apply FDR correction across tested genes, and present forest plots with HRs and 95% CIs.
4. Validate the 16 prognostic genes (and top DEGs) in external datasets (e.g., GEO SARC datasets, recent large sarcoma compendia or the 1300‑sarcoma landscape; see Nature Communications 2025 sarcoma compendium) or by nested cross‑validation to show robustness Sarcoma compendium for validation.
5. Share fully reproducible analysis scripts and processed matrices (counts + metadata) in GitHub and a DOIed archive (Zenodo) to permit external reproduction.

Which claims are well supported vs. tentative?

• Well supported: That there are transcriptomic differences between older and younger TCGA‑SARC samples using standard edgeR pipelines (the paper identifies 733 DEGs under the stated thresholds) — this is an empirical observation given the pipeline used DEG counts reported in preprint.
• Tentative: Assigning mechanistic causality (e.g., that OA patients' worse survival is driven by EMT/Wnt activation) — plausible but confounded by subtype composition and microenvironment differences; needs multivariable modeling, deconvolution, and orthogonal validation (IHC, spatial transcriptomics, or single‑cell data) to be convincing.

Minimal re-analysis plan I would run (bioinformatics blueprint)

1. Obtain TCGA‑SARC raw counts and metadata via TCGAbiolinks (GDC), replicate normalization (TMM/voom) and QC; compute tumor purity/ESTIMATE, record histologic subtype, grade, stage, sequencing batch.
2. DEA with limma‑voom using design = ~ age_group + histology + tumor_purity + batch; produce volcano + effect‑size ranked list; save full table.
3. Deconvolution (xCell/CIBERSORTx) to test whether DEGs remain after adjusting for immune/stromal fractions; re‑run GSEA on residuals/adjusted expression.
4. Survival: for the OA set, run multivariate Cox (continuous expression) controlling for subtype/stage/purity; correct p‑values for multiple testing (Benjamini‑Hochberg) and report only FDR<0.05 hits with HRs and CIs.
5. Validate top genes in at least one external dataset (GEO) or the sarcoma compendium (Nature Communications 2025) and present replication concordance (rank‑rank correlation, direction agreement).

If you want I will run this pipeline on TCGA‑SARC for you (download, DEA with covariates, deconvolution, multivariate survival and external validation) — click below to start a reproducible AI Scientist run.

Bottom-line judgement (balanced)

This is a useful exploratory analysis highlighting age‑related transcriptomic differences in TCGA‑SARC and proposing candidate prognostic genes/pathways for older patients. However, the central biological claims require rigorous control for subtype/tumor composition, correction for multiple testing in survival analyses, and independent validation before clinical or mechanistic conclusions are justified. The study is a hypothesis‑generating resource rather than definitive evidence.

Confidence note: the assessment is grounded on the methods and results reported in the preprint and standard best practices for bulk RNA‑seq analysis (see TCGAbiolinks reference) Primary preprint source