Multimodal AI and single-cell transcriptomics integrate to construct a histopathological prognostic model for bladder cancer, revealing the RTN3-glycolysis axis in chemoresistance

Bladder cancer is a common malignant tumor of the urinary system, and gemcitabine-based chemotherapy often fails because of drug resistance. Tumor heterogeneity, particularly copy number variations (CNVs) heterogeneity at the genomic level, is a core factor that drives disease progression and drug resistance. However, the relationship between CNV heterogeneity at the single-cell level and the histomorphological features that are relied upon in a routine pathological diagnosis remains unclear, limiting its application in clinical prognostic assessments and precision therapy. This study integrated six public single-cell RNA sequencing (scRNA-seq) datasets as discovery cohorts to systematically analyze the cellular landscape of bladder cancer. The inferCNV algorithm was used to infer CNVs at the single-cell level and identify subpopulations with high CNV (HCNV). A multimodal artificial intelligence (AI) strategy combining deep learning and image analysis was employed to construct a prognostic model based on routine H&E-stained pathological sections, with HCNV activity serving as a biological anchor for pathology feature selection. Hub genes were screened using multiomics data, and in vitro and in vivo functional experiments using bladder cancer cell lines were conducted to validate the functions and mechanisms of these hub genes. The single-cell analysis revealed that HCNV cells constitute a malignant subpopulation with a unique molecular phenotype characterized by widespread activation of oncogenic pathways (such as TGF-β, JAK–STAT and PI3K), active communication with the tumor microenvironment (e.g., cancer-associated fibroblasts and M2 macrophages), and a poor prognosis for patients. The multimodal AI-based pathological imaging model effectively predicted patient survival. The model-derived risk score showed a significant but modest correlation with transcriptomic HCNV activity. Through multi-algorithm feature selection, we identified RTN3 as the core gene driving the HCNV phenotype. Our study confirms that RTN3 not only orchestrates an immunosuppressive microenvironment but also directly mediates resistance to gemcitabine. Mechanistically, RTN3 activates the JAK2/STAT3 signaling pathway, which transcriptionally upregulates key glycolytic enzymes (PKM2, GLUT1, and LDHA), thereby driving glycolytic metabolic reprogramming and ultimately conferring gemcitabine resistance in bladder tumors. This study reveals a malignant cell subtype defined by HCNV in bladder cancer at the single-cell level and elucidates the molecular and microenvironmental basis of the poor prognosis of bladder cancer. A clinically translatable multimodal AI pathological prediction model was successfully developed. Furthermore, our findings establish the RTN3/JAK2/STAT3/glycolysis axis as a critical mechanism of gemcitabine resistance, highlighting RTN3 as a promising therapeutic target for overcoming chemoresistance in bladder cancer.