SF1-lineage identity, not the highest-risk k4 cluster, defines non-functioning pituitary neuroendocrine tumors: independent cross-cohort and cross-platform methylation validation
This repository contains the full analysis code and processed results for an independent, cross-cohort validation of the NF-PitNET methylation taxonomy (target journal: Neuro-Oncology).
NF-PitNETs reproducibly map to the GSE283928 tumor cluster 0 (c0, an SF1-lineage methylation identity) across four independent cohorts spanning two microarray platforms (mean 84.3% c0 mapping; Neou 81.2%, GSE207937 75.6%, Moller-2025 raw-idat 86.3%, GSE115783 HM450 94.1%). Moller's highest-risk k4 cluster (TPIT/PIT1, HR 3.63) maps to the secretory clusters 3/4, not to c0 (probe-level p = 1.6×10⁻³⁰), correcting the common conflation of "highest-risk cluster" with "aggressive NF-PitNET." The within-lineage k-subdivision is recoverable from independent minfi processing (leave-one-out 99.1% nearest-centroid / 93.7% random forest), and the k3/k4/k5 high-risk axis shows a trend toward increased regrowth risk (HR 2.14, p = 0.072; adjusted HR 2.09, p = 0.088).
All raw data are third-party and publicly available; no raw data are redistributed here (only processed results and code). See docs/data-availability.md for the full data-source notes, download links, and local storage layout:
| Cohort | Platform | N (composition) | Source | Accession |
|---|---|---|---|---|
| Neou 2020 | EPIC-850K | 86 (32 NF-PitNET; 54 functioning) | ArrayExpress/BioStudies | E-MTAB-7762 |
| GSE283928 | EPIC-850K | 118 (4 tumor clusters: c0 n=70, c2 n=2, c3 n=22, c4 n=24) | GEO (Belakhoua et al., 2025) | GSE283928 |
| GSE207937 | EPIC-850K | 77 (45 NS/NF-PitNET; 32 functioning) | GEO (Silva-Junior et al., 2023) | GSE207937 |
| GSE115783 | HM450 | 40 (34 NFPA tumors; 6 healthy-pituitary controls) | GEO (Kober/Bujko et al., 2018/2019) | GSE115783 |
| Moller 2025 idat | EPIC-850K | 117 (silent gonadotroph NF-PitNET, k1-k5, longitudinal regrowth follow-up) | Mendeley Data | 10.17632/vr2wb38r6t.1 |
| GSE209903 | RNA (transcriptome) | differential expression (NF-PitNET vs functioning) | GEO (Silva-Junior et al., 2023) | GSE209903 |
Roles: GSE283928 supplies the reference tumor clusters (c0/c2/c3/c4) used as nearest-centroid targets; Neou and GSE207937 are independent EPIC-850K validation cohorts; GSE115783 is the cross-platform (HM450) external validation cohort with embedded healthy-pituitary controls; the Moller-2025 raw idat are used for the fully de novo, k-label-free lineage projection and the direct prognostic test; GSE209903 provides the transcriptomic concordance check.
scripts/ All Python (54) and R (17) analysis scripts, fully runnable
results/ Processed beta matrices, centroids, permutation outputs, figures, tables
docs/ Manuscript (manuscript-v2.md), data-availability notes, submission docs, review reports
experience-archive/ Documented analysis pitfalls and reusable workflows
- Download raw data from the sources above into a local
DATASET_PATH(paths are read from a local.env, which is not committed; seedocs/data-availability.mdfor the local layout and theDATASET_PATH/DATASET_PATH_TEMPconvention). - Follow
docs/supplementary-methods.mdfor software versions and random seeds (minfi v1.48.0 / R 4.3.3; scikit-learn 1.5.1, random_state=42; 20,000 permutations, seed=0). - Run scripts in dependency order (see headers of each script; key pipeline entry points:
moller_idat_to_beta.R,neou_crosscohort.py,nn_mixing_permutation.py,build_supplementary_tables.py).
results/Neou/joint_pca_coords.csv— 281-sample joint PCA (PC1-20)results/Neou/nn_mixing_permutation.json— NN-mixing label-permutation p-value (p < 0.001)results/Moller2025/strict_projection/— 117-idat de novo projection (86.3% → c0)results/Moller2025/within_lineage_562dmp/— within-lineage k recovery (99.1%/93.7% LOO)results/Moller2025/prognostic_validation/— Cox regrowth models (HR 2.14 / adjusted 2.09)results/Neou/supplementary_table1-3_*.csv— supplementary tablesresults/Moller2025/figure4_merged.png— merged multi-panel figure
This is a secondary analysis of publicly available, de-identified genomic datasets obtained from GEO, ArrayExpress/BioStudies, and Mendeley Data. No new human-subjects data were generated; the original studies obtained ethical approval and informed consent as described in their respective publications. No additional ethics approval was required for this in-silico reanalysis.
Analysis code: for academic use; see manuscript for citation. Raw data remain under the original sources' terms.
This repository is under active development. Versioned DOI will be minted via Zenodo upon acceptance.