Chauhan, Rohit

Abstract Unsupervised molecular clustering is fundamental to precision medicine, yet most published subtyping studies lack standardized validation criteria, making it impossible to distinguish biological signal from statistical artifact. We present the Pathway Subtyping Framework, which enforces four mandatory validation gates—label shuffle, random gene set, bootstrap stability with adaptive thresholds, and ancestry independence—before any clustering result is reported. On TCGA-COAD (n=452), the framework independently recovered CMS4 (Mesenchymal) with 75.9% CMS4 enrichment (Fisher OR=16.71, p=1.4×10⁻²⁵), outperforming PCA+k-means (67%), gene-level k-means (50%), and NMF (32%) on the same data. Applied to autism (n=56) and schizophrenia (n=281) as pilot demonstrations, the framework produced replicable subtypes across platforms (ARI=0.870) and achieved the strongest gate performance across all datasets (bootstrap ARI=0.923). Cross-disease analysis reveals that molecular subtypes are independent of DSM diagnosis (Cramér's V=0.036) but strongly dependent on brain region (V=0.528); within-region concordance is 84% (ARI=0.862), decomposing the 87% global concordance into disease-agnostic and anatomical components. The adaptive bootstrap threshold model, calibrated on 47 independent benchmarks (R²=0.889, LOOCV RMSE=0.051), replaces arbitrary fixed thresholds with data-driven, silhouette-calibrated gates. The framework is open-source (PyPI, Docker, Codeberg; 1,054 tests; DOI:10.5281/zenodo.18867165).