PitNET Reproducibility Benchmark

Version 0.1.0 (2026-10-22)

This repository contains the analysis code, harmonized metadata, summary results, and final figures for a multi-cohort benchmark of public pituitary neuroendocrine tumor (PitNET) transcriptomes.

The benchmark evaluates:

1. GH-versus-NFPA lineage transfer across RNA-seq and microarray cohorts;
2. NFPA invasion-signature transfer across independent cohorts;
3. transfer of an autosomal sex-associated program across NFPA, ACTH tumor, and normal pituitary contexts;
4. phenotype-lineage confounding in a public invasive-versus-non-invasive comparison.

Key results

• GH-versus-NFPA lineage: log fold-change correlation 0.843, direction concordance 0.932, external AUC 1.000.
• NFPA invasion: correlation -0.005, external AUC 0.500.
• Autosomal sex program: external AUC 0.472 in ACTH tumors and 0.467 in normal pituitary.
• GSE260487: invasion and lineage are completely confounded; PC1 R2 = 0.982 for both.

Repository layout

• scripts/: processed-data analysis and figure-generation code.
• config/: cohort inventory and recommended PitNET metadata schema.
• data/metadata/: public GEO sample-level metadata.
• data/raw_manifest.csv: file names, sizes, SHA256 values, and expected raw data locations.
• results/: differential-expression, signature, and benchmark result tables.
• figures/: final vector PDF and 300 dpi PNG figures.
• environment/: software versions and Python requirements.
• reports/: benchmark summary and novelty audit.

Data

Raw GEO matrices are not redistributed. data/README.md lists the source accessions and expected processed files. The hash manifest identifies the exact files used in version 0.1.0.

Requirements

• R 4.5.1 or a compatible recent R version.
• R packages listed in environment/R-packages.txt.
• Python 3.11 with packages listed in environment/requirements-pituitary.txt.
• BiocManager access for platform annotation packages.

The original analysis was run in an isolated environment with R 4.5.1, limma 3.66.0, edgeR 4.8.2, MCPcounter 1.2.0, Python 3.11.9, Scanpy 1.11.5, and scvi-tools 1.4.2.

Running

1. Download the processed GEO files listed in data/README.md into the matching data/raw/<accession>/ directory.
2. Set PITNET_PYTHON and PITNET_RSCRIPT if the executables are not on PATH.
3. Run:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/run-analysis.ps1

The scripts reproduce the result tables and Figure 2-5. Figure 1 is produced by scripts/11_metadata_completeness.R.

Citation

See CITATION.cff. The public release is available at https://github.com/626619622/PitNET-ReproducibilityBenchmark. A Zenodo DOI will be added after the repository is archived.

License

Code is released under the MIT License. GEO data remain subject to the terms of their original repositories.