Cross-platform, random-effects meta-analysis of differentially expressed genes in diabetic nephropathy (DN) across three independent human glomerular microarray cohorts from GEO. Implemented in pure Python (numpy/scipy/pandas) — no R/limma dependency; all statistics are re-implemented from the primary literature so every step is transparent.
📄 Full write-ups: results/REPORT.md (glomerular microarray) ·
results/RNASEQ_REPORT.md (whole-kidney RNA-seq companion + ARCHS4 re-examination)
RNA-seq companion / ARCHS4 comparison. A second meta-analysis applies the same study-matched effect-size framework to three human whole-kidney RNA-seq cohorts (GSE142025, GSE162830, GSE166239) and compares to the SuLab ARCHS4 result. Key finding: the ARCHS4 top genes (immediate-early factors FOS/FOSB/EGR1/NR4A1/DUSP1) show I²≈95% heterogeneity and fail FDR here — a tissue-procurement artifact that naïve pooling surfaces but random-effects demotes. The reproducible cross-modality DN signal is podocyte injury (NPHS1, NPHS2, PTPRO, MAGI2), significant in both compartments; the fibrosis/complement program is glomerular-compartment-specific. See
fig5–fig7.
| Dataset | DN | Control | Platform |
|---|---|---|---|
| GSE30528 (Woroniecka 2011) | 9 | 13 | GPL571 · HG-U133A_2 |
| GSE96804 (Pan 2018) | 41 | 20 | GPL17586 · HTA-2.0 |
| GSE104948 / GPL22945 (ERCB) | 7 | 18 (living donors) | GPL22945 · HG-U133+2 ENTREZG CDF |
Three biological cohorts on three Affymetrix platforms (57 DN vs 51 control).
- Download normalized GEO series matrices; assign DN vs control from sample metadata.
- Map probes → HGNC gene symbols per platform; collapse to genes (max-mean probe).
- Per-study DE via limma-style empirical-Bayes moderated t-test.
- Per-study effect sizes as Hedges' g (standardized mean difference).
- Pool with DerSimonian–Laird random-effects meta-analysis (+ Cochran's Q / I²); BH-FDR. Cross-checked with a weighted Stouffer p-value combination.
- 11,472 genes measured in all three studies.
- 1,714 high-confidence DEGs (present in all 3, FDR < 0.05, direction-consistent): 1,022 up / 692 down in DN.
- Recovers the canonical DN glomerular signature — podocyte loss (NPHS1↓, TJP1, MPP5, GJA1) and fibrosis/complement/macrophage activation (TGFBI, COL1A1, LUM, MMP2, C1QA, VSIG4↑).
pip install numpy pandas scipy matplotlib GEOparse
python scripts/run_de.py # Stage 1: download + per-study DE + effect sizes
python scripts/meta.py # Stage 2: random-effects meta-analysis (+ Stouffer)
python scripts/figures.py # Stage 3: figures + ranked tablesRaw GEO downloads are cached under data/ (git-ignored; re-created on first run).
scripts/ recon.py · lib.py · run_de.py · meta.py · figures.py
results/ REPORT.md, meta_results.csv, robust_meta_DEGs.csv, figures (png), per-study tables
- Woroniecka KI et al. Diabetes 2011 — GSE30528
- Pan Y et al. 2018 — GSE96804
- European Renal cDNA Bank (ERCB); Ju W / Grayson PC et al. — GSE104948 (GPL22945 glomerular subseries)
Built with Claude Code.