Decoding Dendritic Cell Subtypes via Integrated Radiogenomics: A Stacked Ensemble Model for Predicting Immunotherapy Response in <scp>NSCLC</scp>

We pioneer a multimodal framework integrating single-cell RNA sequencing (scRNA-seq), radiomics, and deep learning to decipher dendritic cell (DC)-mediated mechanisms underlying anti-PD-1 response in non-small cell lung cancer (NSCLC). Single-cell RNA sequencing of tumor samples from responders and non-responders identified nine immune cell types, among which DCs displayed significant differences between groups. Cell-cell communication and pseudotime analyses highlighted conventional DC2 (cDC2) and tolerogenic DC (tDC) as key subsets linked to therapeutic outcomes. Four cDC2 marker genes (FAM3B, TFAP2A, RTKN2, and XCL2) and two tDC marker genes (KRT6A and RAB27B) showed predictive value, and experimental validation confirmed reduced cDC2 and tDC abundance in responders, with upregulation of FAM3B, RTKN2, and XCL2, and downregulation of TFAP2A, KRT6A, and RAB27B. DCs also modulated CD8⁺ T cell activity via OSM-IL6ST and IL15-IL15RA signaling. These six genes were incorporated into machine learning models combining transcriptomic (LSTM), clinical (EnhancedClinicalRNA), and radiomic (ResNet50) data. A stacked ensemble learning approach integrating all three modalities achieved superior performance, with an accuracy of 0.97 and an AUC of 0.99. In summary, our results demonstrate that combining single-cell transcriptomics and radiomics through ensemble deep learning enables accurate prediction of immunotherapy response in NSCLC, and identifies six DC-associated marker genes with potential as prognostic biomarkers.