Multi-Platform Transcriptomic Classification of LUAD Driver Mutation Subtypes via Heterogeneous Ensemble Learning

Background: Accurate identification of EGFR-mutant, KRAS-mutant, and triple-negative (TN) LUAD driver mutation subtypes is essential for guiding targeted therapy, yet standard molecular testing is invasive and not always feasible. Gene expression-based machine learning offers an alternative that avoids additional invasive sampling; however, existing models are largely restricted to binary classification and lack cross-platform generalizability. Methods: We developed a multi-cohort, ComBat-harmonized pipeline integrating SHAP-driven feature selection and a Bayesian-optimized heterogeneous ensemble of five learners, extended with NSGA-II Pareto-optimal weight optimization under leak-free cross-validation. External validation used an independent RNA-seq cohort with reference-batch harmonization and threshold calibration. Results: ParetoVoting achieved the best internal performance (F1-macro = 0.736, AUROC = 0.877), statistically indistinguishable from SoftVoting (F1-macro = 0.735; Cliff's δ=-0.12), which was carried forward for external validation. On TCGA-LUAD (n=508), SoftVoting attained F1-macro = 0.700 and AUROC = 0.850, though this reflects an out-of-range KRAS threshold (representative value ∼0.66-0.68). Ablation analysis showed components contributed unequally within internal cross-validation, with ComBat's critical role instead emerging during cross-platform external generalization. SHAP analysis identified biologically coherent signatures consistent with known LUAD oncogenic signaling. Conclusions: This framework provides a reproducible, cross-platform approach to three-class driver mutation subtyping from transcriptomic data, offering benchmarks for this underexplored task and potential clinical utility where tissue-based profiling is infeasible.