Ensemble machine learning using unmatched pre-treated multi-omic data can classify response to immune checkpoint inhibition in non-small cell lung cancer

Introduction: Immune checkpoint inhibitors (ICI) improve survival in non-small cell lung cancer (NSCLC), yet predictive biomarkers remain limited. Multi-omic data can improve classification, but incomplete multi-omic overlap poses challenges for integrative analysis. Ensemble methods can leverage partially overlapping multi-omic data for biological insight and predictive power. We developed an ensemble machine learning (ML) framework integrating unmatched exome and transcriptome datasets to classify ICI response, enabling model training when only a single data modality is available for a patient. Methods: Using the SU2C-Mark Foundation NSCLC dataset (SU2C NSCLC), separate exome- (n=241) and transcriptome-based (n=84) models were trained using cross-validation, and validated via ensemble learning on overlapping multi-omic data (n=68). Models were tested in an independent SU2C NSCLC cohort (MDA), and externally validated in melanoma and urothelial datasets. Results: In SU2C NSCLC, separate models were trained in the unmatched exome and unmatched transcriptome data, and results were blended to predict response in the 68 matched, multi-omic cases (46% responders). The baseline ensemble model using the clinically assessed PD-L1 tumor proportion score (TPS) achieved an area under the ROC curve (AUC) of 0.60. Incorporating domain knowledge and blending the two model predictions (tumor mutation burden [TMB] & CXCL9 expression) improved AUC to 0.76, while a simple ensemble including TMB, CXCL9 , PD-L1, first-line therapy and dual ICI reached AUC 0.80. LASSO models adding Amp5p15.33 ( TERT locus), SBS7b COSMIC mutational signature, and clonal neoantigens achieved AUC 0.79; non-linear models (RF/XGBoost) reached 0.73. In the MDA cohort (n=22), baseline PD-L1 AUC was 0.68, domain knowledge ensemble was AUC 0.88, simple ensemble AUC was 0.82, and the previously selected test model of LASSO achieved 0.79. The domain knowledge ensemble generalized well on external data: in urothelial cancer (Mariathasan et al. Nature 2018; train n=105; val n=45), AUC was 0.94, and in melanoma (Freeman et al., Cell Rep Med. 2022; unmatched exome n=99, unmatched RNA n=98, matched multi-omic validation n=55), AUC was 0.65. Differential expression and pathway analysis in misclassified tumors highlighted stromal and histologic heterogeneity as residual error sources. CXCL9 expression and TMB were not co-linear in TCGA, helping to explain their additive power. Conclusion: A pragmatic ensemble ML framework can integrate unmatched exome and transcriptome data to classify ICI response in NSCLC. Blending immune ( CXCL9 expression) and genomic (TMB) features yields robust, interpretable, generalizable, and biologically meaningful models. This approach supports translation of clinically actionable, integrative biomarkers for precision immunotherapy.