A multiscale framework for hypothesis generation from sex-stratified X-linked gene networks

Stratified gene co-expression network analyses compare molecular organisation across biological groups; however, their interpretation is challenged by cohort imbalance, network instability, correlation-method dependence, and multiplicity. These issues are particularly relevant when comparing network structures across heterogeneous cohorts, where apparent patterns may not be reproducible. This study presents a reproducible framework for stratified gene co-expression network analysis. It integrates cohort balancing, bootstrap-stabilised consensus networks, topology-based statistics, paired label permutations, multiplicity control, and correlation-method benchmarking. Its design distinguishes robust statistical evidence from exploratory signals and their biological interpretation. The framework is illustrated through a multi-cancer case study of sex-associated co-expression network organisation among X-linked genes classified as escape, inactive, or variably escaping from X-chromosome inactivation. The analysis included 15 TCGA cancer cohorts and 90 sex-stratified networks, enabling systematic comparison of network structure across cancer types and X-inactivation gene classes. Across 315 paired permutation tests, none survived Benjamini–Hochberg correction, providing no multiplicity-adjusted evidence. Nevertheless, exploratory graphlet profiles revealed cancer- and gene-class-specific patterns of network organisation, highlighting candidate structures for independent replication. The case study illustrates how the framework can distinguish nominal or visually compelling differences from multiplicity-adjusted evidence while preserving informative exploratory patterns for structured hypothesis generation.