De-Leaked Pancreatic Cancer GEO Atlas

Audited PDAC GEO meta-analysis uniting biomarkers, survival, and five PhD tracks.

1,791
Human pancreatic cancer GEO series cataloged (2006 to 2026)
74.9%
Patient overlap caught inside RNA-seq cohort GSE224564 (131/175)
0.89 to 0.95
Blind external ROC AUC for the locked 10-gene diagnostic panel
r = 0.918
Matched Microarray to RNA-Seq risk concordance across 128 tumors

1. NCBI GEO Landscape, GEO2R Critique & Patient-Leakage Audit

Before statistical modeling, we benchmarked NCBI’s default GEO2R workflow and audited patient identifiers across consortium datasets.

Annual Human Pancreatic Cancer GEO Deposits (2006 to 2026)

Hover bars for yearly counts; orange highlights 2023 to 2026 acceleration.

Four Failure Modes Caught in Default GEO2R & Cohorts

  • Cross-Study Patient Leakage: GSE62452 (90/130) and GSE28735 (45 pairs) overlap Discovery GSE183795. Inside bulk RNA-seq GSE224564, 131/175 tumors (German-1/2/3 + UMD-1, 74.9%) match GSE183795, leaving 44 independent patients (UMD-2/3).
  • Unlogged Intensity Trap: GSE16515 distributes linear MAS5 intensities up to 65,045 (median 18.17), fixed by automated log2(x + 1) scaling.
  • Replicate & Paired Power: Averaging 6 technical _rep arrays in GSE15471 (78 to 72 arrays = 36 pairs) and running patient-blocked limma recovers +26 DEGs (1,584 vs 1,558 unpaired).
  • Multi-Probe Noise: Collapsing redundant probes by max IQR removes conflicting symbol calls across GPL6244, GPL570, and GPL15048.

2. Interactive Consensus Biomarker Explorer (565 DEGs)

Across 13,556 shared genes in four cohorts (r = 0.709 to 0.847), weighted Stouffer integration isolated 433 upregulated and 132 downregulated consensus markers paired with acinar enzyme collapse (17/18 genes, OR = 1,984.3, FDR = 2.74e-33).

Showing 565 genes Download Meta CSV
Gene Dir Mean log2FC GSE183795 GSE15471 GSE16515 GSE71729 Meta-Z Meta-FDR Basal d Program Role

Enriched PDAC Biological Modules (Fisher ORA)

Cross-Platform Differential Expression Concordance

Pairwise log2FC correlations vs Discovery GSE183795 (13,556 shared genes):

  • GSE16515 (GPL570 Mayo, log2-fixed): Pearson r = 0.8466
  • GSE71729 (Agilent 2-Channel Moffitt): Pearson r = 0.7513
  • GSE15471 (GPL570 Paired Blocking): Pearson r = 0.7092

3. Blind Cross-Platform 10-Gene Diagnostic Classifier (559 Tissues)

Trained strictly on GSE183795 via single-sample within-transcriptome standardization, then frozen prior to unseen validation.

Interactive Cross-Platform ROC Curves

Two-Channel Platform Calibration Insight

On two-channel Agilent GSE71729, ranking performance stays high (AUC = 0.8883), whereas fixed p = 0.50 collapses Specificity to 0.000 due to Channel-1 reference shift. Youden calibration restores Specificity to 0.9348.

4. Kaplan-Meier Survival, Leakage Distortion & Cross-Tech Bridge

Compare leaked vs strictly de-leaked external survival generalization alongside 128 matched Microarray and RNA-seq patient tumors.

Pooled Independent External (N = 230 Stratified)

Zero Patient Leakage

Matched Cross-Technology Bridge (n = 128)

Exact tumors profiled by Affymetrix GPL6244 and Bulk RNA-Seq achieve r = 0.9180 (p = 1.87e-52).

Multivariable Cox Proportional Hazards Models

5. Basal-Squamous Plasticity & Metastatic Escalation

The locked risk axis separates aggressive Basal-like tumors orthogonally from stromal fibroblasts and increases stepwise toward metastasis.

Basal vs Classical Split

RNA-Seq (GSE224564): Basal (n=48, +0.768) vs Classical (n=95, -0.273), AUC = 0.8096 (p = 8.00e-10).

Array (GSE71729): Basal (n=89, +0.747) vs Classical (n=132, -0.193), AUC = 0.8341 (p = 1.91e-17), independent of Stroma (p = 0.234).

Stepwise Tissue Escalation

Mean Risk Score Z in GSE71729 rises cleanly across stages:

  • Normal Pancreas (n=46): -0.641
  • Primary PDAC (n=145): -0.076 (p = 3.43e-06)
  • Metastasis (n=61): +0.368 (p = 2.56e-05)

ITGA3 & Subtype Drivers

Top Basal drivers in RNA-Seq: LY6D (d=+1.73), IL20RB (+1.69), KRT6A (+1.59), and ITGA3 (+1.49), converging on Oct 2026 CRISPR screen GSE349252.

Classical anchors: LYZ (-1.78), CLDN18 (-1.50), TSPAN8 (-1.45), LGALS4 (-1.45).

6. Publication Figures Gallery (300 DPI)

Click any manuscript panel to inspect full-resolution charts.

Figure 1: NCBI GEO Landscape and Workflow

Figure 1. GEO PDAC Landscape & Study Design

1,791 indexed series (2006 to 2026), technology breakdown, and non-overlapping validation split.

Figure 2: PCA, Leakage Audit, and Log2FC Concordance

Figure 2. PCA Separation, Leakage Audit & Concordance

Tumor vs Normal PCA across 4 cohorts, overlap rates vs Discovery, and log2FC correlation.

Figure 3: Consensus Volcano, Top 24 Heatmap, and ORA

Figure 3. Meta-Volcano, 24 Biomarkers & Pathway ORA

433 Consensus UP / 132 DOWN genes, Desmoplasia/EMT activation, and Acinar enzyme collapse.

Figure 4: Cross-Platform Diagnostic ROC and Calibration

Figure 4. Blind 10-Gene Diagnostic ROC & Calibration

AUCs 0.888 to 0.955 across 559 tissues plus fixed p=0.5 vs Youden two-channel calibration proof.

Figure 5: Kaplan-Meier Curves, Leakage Proof, and Bridge

Figure 5. Kaplan-Meier Survival & Cross-Tech Bridge

Discovery N=134, Zero-Leakage Pooled N=230, Leaked GSE224564 comparison, and r=0.918 bridge.

Figure 6: Basal Plasticity and Metastatic Escalation

Figure 6. Basal-Squamous Plasticity & Metastasis

Basal vs Classical risk separation, Normal-Primary-Metastasis escalation, and ITGA3 convergence.

7. Five Presentation-Ready PhD Proposals

Concrete dissertation tracks anchored directly on audited NCBI GEO cohorts including 2026 experimental releases.

Track 1 / GSE183795, GSE15471, GSE16515, GSE71729 + GSE310252 (n=118)

De-Leaked Multi-Platform PDAC Diagnostic & Basal Stratification

Extends our locked diagnostic and Basal-plasticity pipeline onto freshly resected adjuvant-treated Dutch Parel Biobank RNA-seq (GSE310252, Apr 2026).

Track 2 / GSE353716 (n=94), GSE353717 (n=47), GSE314073, GSE239310

Acquired Resistance to Oncogenic KRAS-G12C & KRAS-G12D Inhibition

Tests whether our EMT/Basal consensus program drives adaptive escape under MRTX1133, sotorasib, and adagrasib across Nov 2026 resistance cohorts.

Track 3 / GSE314577 (Feb 2026, N=2,614) + GSE332591 (n=96)

Longitudinal Pre-Diagnostic & Exosomal miRNA Liquid Biopsy

Models pre-diagnostic circulating miRNA trajectories years prior to onset across 2,614 subjects and validates in serum exosomal miRNA-seq.

Track 4 / GSE349252 (Oct 2026, n=45), GSE283850 (n=44), GSE285406 (n=48)

ITGA3 Dependency, Hypoxia Organoids & ZNF274 Plasticity Checkpoints

Links our computational ITGA3 Basal finding (d=+1.49) to CRISPR dependency screens, hypoxic organoids, and CDK7-inhibitor resistance.

Track 5 / GSE336517 (GeoMx DSP, n=149), GSE343070 (scRNA, n=66), GSE102238 (PNI, n=100)

Spatial CD68+ Macrophage Giant Cells, SEMA4A+ Cachexia & Perineural Invasion

Maps our TAM chemokine and stromal modules onto NanoString GeoMx digital spatial neighborhoods and single-cell cachexia trajectories.

8. Audited NCBI GEO PDAC Directory (Top 80 Cohorts)

Filter cataloged human PDAC GEO studies and click any accession ID to open the official NCBI record.

Showing 80 cohorts Download Catalog CSV
Accession Samples Date Assay Type Study Title