A tissue-aware computational framework for confounding-controlledco-expression network analysis: Context-dependent utility in drug sensitivitymodelling

Abstract

Pan-cancer pharmacogenomic modelling using co-expression networks faces a fundamental challenge: topological metrics predominantly encode tissue identity rather than genuine pharmacological signal, inflating apparent predictive performance and undermining biological interpretability. We introduce a tissue-aware computational framework integrating tissue-specific WGCNA construction, within-tissue edge disruption profiling, and per-gene -score standardisation, with tissue prediction accuracy as a quantitative acceptance criterion for confounder control. The framework was applied to 660 cancer cell lines across three targeted therapies (osimertinib, crizotinib, KRAS G12C Inhibitor-12) and evaluated under repeated stratified cross-validation with FDR correction. Within-tissue standardisation reduced tissue identity encoding in topological features from 88.9% to 12.9%. Within-tissue WGCNA modules outperformed a confounder-free PCA baseline across all three drugs, confirming genuine co-regulatory structure beyond dimensionality reduction. Osimertinib met the pre-specified primary improvement criterion ( , versus tissue-residualised expression), while crizotinib and KRAS G12C Inhibitor-12 showed no meaningful improvement, establishing that the predictive benefit of co-expression topology is highly drug-context-dependent. A proof-of-concept feasibility check in an independent cohort (GSE255958) showed that baseline expression features discriminated drug-tolerant persister cells from responders in the EGFR context (AUC ), though the small sample size precludes strong inferential claims and the network component could not be evaluated externally. This work establishes a reproducible computational standard for confounder-controlled co-expression network analysis in pan-cancer pharmacogenomics, with direct applicability to any CCLE- or GDSC-scale study.