Smoking-Stratified Signal Decomposition and Feature Selection for Never-Smoker Cancer Classification in a Combined Lung–Breast Metabolomics Cohort

Metabolomic cancer classifiers trained on mixed-smoking cohorts may embed tobacco exposure signal within their predictions, degrading performance in never-smokers, a population in which lung adenocarcinoma is frequently diagnosed. We developed a two-stage framework that (i) decomposes a shared 129-metabolite panel from a combined lung–breast cancer cohort (n=1038) into cancer-specific (Signal C), smoking-specific (Signal S), and shared (Signal S∩C) components using two-way analysis of variance with Benjamini–Hochberg correction, and (ii) applies multiple feature-selection strategies to identify the minimal Signal C subset that surpasses the all-metabolite baseline for never-smoker cancer detection. Two-way ANOVA partitioned 54 of 129 metabolites as cancer-specific (Signal C) and 57 as smoking-specific (Signal S), suggesting that nearly half of the shared panel is influenced by tobacco exposure. A model of 19 Signal C metabolites, selected by composite rank aggregation across four feature-selection methods and trained with gradient-boosted trees, achieved a never-smoker area under the receiver operating characteristic curve (AUC) of 0.907 on pooled out-of-fold predictions (0.910 as a mean across folds) against an all-metabolite baseline of 0.895, using 85% fewer metabolite measurements. The signal decomposition is a reproducible and interpretable way to identify metabolites whose case–control differences are not attributable to tobacco exposure, and it permits a substantial reduction in panel size.