| title | Pathway Subtyping Framework: Validated Molecular Subtype Discovery from Pathway-Level Genomic Data | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| tags |
|
||||||||||
| authors |
|
||||||||||
| affiliations |
|
||||||||||
| date | 20 February 2026 | ||||||||||
| bibliography | paper.bib |
The Pathway Subtyping Framework is an open-source Python package for discovering molecular subtypes in genetically heterogeneous diseases. It shifts subtyping from individual genes to biological pathways — curated gene sets representing shared molecular processes — providing biologically motivated dimensionality reduction that improves statistical power and interpretability. The framework accepts gene expression matrices (bulk RNA-seq, microarray) or variant burden scores (VCF), aggregates them into pathway-level scores using ssGSEA, GSVA, or mean-Z methods, clusters patients using Gaussian Mixture Models (GMM), and enforces mandatory validation gates before accepting results. It is disease-agnostic: the same pipeline operates on any condition for which curated pathway definitions exist.
This paper describes the framework as of v0.2/v0.3 (February 2026) and several of its empirical claims have since been retracted. It has not been rewritten; this notice scopes what is still safe to cite.
Retracted or superseded:
- The adaptive bootstrap-threshold model (R²=0.889) does not reproduce.
- The 47-dataset benchmark contained an empty-input ARI artifact (14 invalid rows), an incorrect independence claim, and count discrepancies. A corrected dataset supersedes it — cite the concept DOI
10.5281/zenodo.19323753, which always resolves to the newest version.- The cross-disease convergence figures quoted below (ARI=0.870, bootstrap ARI=0.923 as a general validation claim, CMS4 recovery ~76%) come from that uncorrected analysis and must not be quoted as current results.
- The associated preprint (Research Square rs-9284565) has been withdrawn from journal review.
- The framing "validated molecular subtype discovery" in the title above overstates what the gates establish. Subsequent work (v0.8.0) found the original bootstrap-stability gate tested pathway independence, not discreteness, and could certify a continuous gradient as a subtype.
Still accurate: the architecture, the scoring methods, the input handling, and the general motivation in Statement of Need.
Full notice and corrected artifacts:
CORRECTION_2026-07/ERRATUM_2026-07-08.md. Current gate design:docs/discreteness_gate.md.
Over 100 risk genes contribute to autism spectrum disorder alone, each explaining a small fraction of cases [@Satterstrom2020]. Similar genetic architectures characterize schizophrenia [@Trubetskoy2022], epilepsy, and intellectual disability. This heterogeneity means that patients sharing a single clinical diagnosis may have fundamentally different molecular disruptions — and may respond to different treatments. Gene-level clustering is underpowered for typical cohort sizes, and existing pathway analysis tools focus on differential enrichment between predefined groups rather than unsupervised subtype discovery.
The Pathway Subtyping Framework addresses this gap by combining pathway-level feature engineering with validated unsupervised clustering. Unlike GSEA [@Subramanian2005] or GSVA [@Hanzelmann2013], which compare predefined groups, this framework discovers groups from the data. Unlike NMF-based approaches [@Brunet2004], it leverages curated biological pathway knowledge. And unlike any existing tool, it enforces mandatory validation gates that distinguish real molecular subtypes from statistical artifacts.
Multi-modal input. The framework accepts bulk RNA-seq expression matrices, microarray data, and VCF files (rare variant burden scoring). All inputs converge on a shared pathway aggregation engine.
Three scoring methods. Pathway-level scores are computed via ssGSEA [@Barbie2009], GSVA [@Hanzelmann2013], or mean-Z (Z-score averaging across member genes).
Gaussian Mixture Model clustering. GMM with full covariance, BIC-based model selection over k=2–8, 10 random restarts, and deterministic seeding.
Five mandatory validation gates:
- Label shuffle (negative control): Verifies clusters are not random artifacts
- Random gene sets (negative control): Verifies curated pathways drive clustering, not arbitrary gene groupings
- Bootstrap stability: Verifies clusters are robust to resampling
- Ancestry independence: Verifies clusters do not reflect population stratification
- Cross-modal concordance: Verifies subtypes replicate across data modalities
Correction modules. ComBat batch correction, ancestry regression (PCA-based), and sensitivity analysis across clustering algorithms, feature subsets, and normalization methods.
Benchmark suite. Built-in comparison against PCA+K-means, NMF+K-means, gene-level K-means, and random baseline.
Disease templates. Pre-curated GMT pathway files for autism, schizophrenia, epilepsy, intellectual disability, Parkinson's disease, and bipolar disorder.
Multi-omic integration (v0.3.0). Single-cell pathway scoring (h5ad support), bulk deconvolution (NNLS), signaling pathway databases (CellPhoneDB, CellChatDB), and multi-omic fusion (concatenate, weighted average, intersection strategies).
The framework has been validated on ten independent transcriptome datasets spanning brain, blood, and lymphoblastoid tissue, five microarray platforms, RNA-seq, and three diseases (autism, schizophrenia, colorectal cancer):
Brain tissue (4 datasets, 649 samples, ASD + SCZ):
- GSE28521 [@Voineagu2011]: 32 frontal cortex samples (16 ASD, 16 control). Discovered a GABA-Collapsed subtype (n=9, 100% ASD, Cohen's d=3.21 for GABA signaling).
- GSE64018 [@Gupta2014]: 24 temporal cortex RNA-seq samples. Independent replication: disease-enriched subtype (83% ASD) with the same top 3 disrupted pathways (GABA, glutamate, cell adhesion).
- GSE80655 [@Ramaker2017]: 141 schizophrenia + control samples, 3 brain regions. Three subtypes identified (Dopamine-Hyperactive, Neurodevelopment-Activated, Synaptic-Collapsed).
⚠️ Superseded — see the correction notice above. This was originally reported as the first real-data analysis to pass all 3 validation gates (bootstrap ARI=0.923), with cross-disease convergence of ASD and SCZ pathway sets at ARI=0.870. Re-analysis found the partition tracks brain region, not diagnosis (region Cramér's V 0.66; diagnosis permutation p=0.234 at donor level, i.e. indistinguishable from chance), and that the 141 samples come from only 48 donors, so sample-level tests overstated significance. Do not cite the 0.870 convergence figure. - GSE53987 [@Lanz2019]: 205 samples across 4 diagnoses (SCZ, BD, MDD, CTL) and 3 brain regions (Affymetrix microarray). Cross-platform replication from GSE80655 RNA-seq with projection ARI=0.319 and cross-disease ARI=0.792, confirming pathway convergence across platforms.
Blood and peripheral tissue (4 datasets, 880 samples, ASD + SCZ):
- GSE111175 [@Gazestani2019]: 141 blood leukocyte samples (28 ASD, 113 control). Pathway subtypes detectable in peripheral tissue; synaptic transmission scores correlate with ADOS Social Affect severity (rho=-0.52, FDR p=0.032).
- GSE18123 [@Kong2012]: 285 blood samples (72 ASD, 213 control) across two Affymetrix platforms. Cross-cohort projection from GSE111175 achieves ARI=0.374, confirming blood-based subtype replicability.
- GSE15402 [@Hu2009]: 116 lymphoblastoid cell lines (87 ASD, 29 controls). ADI-R severity subgroup association (chi-square p=0.001); framework subtypes capture clinically meaningful variation.
- Multi-cohort SCZ blood (5 GEO datasets, 407 samples, 177 SCZ): Multi-platform replication across 5 microarray platforms. Merged k=7 subtypes with Hertzberg concordance mapping (4 Immune-like + 3 Neuro-like subtypes).
Cancer tissue (1 dataset, 452 samples, CRC):
- TCGA-COAD: 452 primary colorectal adenocarcinoma tumors from NCI GDC (RNA-seq). 50 MSigDB Hallmark pathways, k=3 subtypes. CMS external validation: Subtype 0→CMS4 at 76% (Fisher p=1.4e-25). Kaplan-Meier survival analysis with log-rank test and Cox proportional hazards regression.
Across 1,934 samples and 10 independent datasets, the framework demonstrated that molecular subtypes are diagnosis-independent (chi-squared p=0.72) but brain-region-dependent (p=3×10⁻¹⁶), that pathway subtypes are detectable in peripheral blood and lymphoblastoid cells, that subtypes replicate across platforms (microarray vs RNA-seq) and cohorts (cross-cohort projection ARI>0.3), and that the disease-agnostic pipeline recovers established cancer subtypes (CMS4). A knowledge graph analysis integrating STRING protein-protein interactions (4,378 edges among 336 pathway genes, 97.9% coverage) with DGIdb drug-gene interactions identified 20 hub genes (11 cross-disease bridges including AKT1, CTNNB1, GSK3B, PTEN), 6 Louvain communities (all cross-disease), and 1,546 drug repurposing candidates ranked by target gene network centrality.
The framework is implemented in Python (≥3.8) with core dependencies limited to NumPy, pandas, scikit-learn, and SciPy. Optional extras include [vcf] (pysam for VCF parsing), [viz] (Plotly for interactive visualization), and [sc] (anndata for single-cell data); the current release ships seventeen extras in total. It is distributed via PyPI (pip install pathway-subtyping, RRID:SCR_028051, bio.tools:pathway-subtyping), archived on Zenodo (concept DOI: 10.5281/zenodo.18638048, which resolves to the latest release), and available as Docker images (rohitdataops/pathway-subtyping on Docker Hub). Jupyter notebooks are provided spanning synthetic tutorials and real-data validation across autism, schizophrenia, and cancer datasets.
(Figures in this section were current at v0.2/v0.3. The test count, extras list, notebook count and DOI have all moved since; see CHANGELOG.md and README.md for current values. The Zenodo DOI originally printed here, 10.5281/zenodo.18442426, is a retired concept DOI that no longer resolves.)
The GSE28521 dataset was generated by Voineagu et al. (2011), GSE64018 by Gupta et al. (2014), GSE80655 by Ramaker et al. (2017), GSE111175 by Gazestani et al. (2019), GSE18123 by Kong et al. (2012), and GSE53987 by Lanz et al. (2019), all publicly available via NCBI GEO. Pathway definitions are curated from SFARI Gene, KEGG, Reactome, and MSigDB.