Cancer metabolomic classifiers trained on mixed-smoking cohorts can absorb tobacco-exposure signal and lose accuracy in never-smokers, a group in which lung adenocarcinoma is often diagnosed. We used a two-stage approach on a shared 129-metabolite panel from a combined lung–breast cancer cohort (n = 1038). Two-way ANOVA with Benjamini–Hochberg correction split the panel into cancer-specific (Signal C), smoking-specific (Signal S), and shared components. Of 129 metabolites, 54 were Signal C and 57 were Signal S, so nearly half of the panel tracked tobacco exposure. Rank aggregation across four feature-selection methods then isolated a 19-metabolite Signal C model, trained with gradient-boosted trees. In never-smokers it reached an out-of-fold AUC of 0.907 (mean fold AUC 0.910) versus an all-metabolite baseline of 0.895, with 85% fewer measurements. The decomposition flags metabolites whose case–control differences are not due to smoking and supports a much smaller panel.