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

A malignant bladder cancer subpopulation defined by high copy number variation (HCNV) was identified at the single-cell level, showing widespread activation of oncogenic pathways such as TGF-β, JAK-STAT and PI3K, active communication with tumor microenvironment components including cancer-associated fibroblasts and M2 macrophages, and association with poor patient prognosis. The relationship between single-cell CNV heterogeneity and the histomorphological features relied upon in routine pathological diagnosis was previously unclear; this work uses HCNV activity as a biological anchor linking molecular heterogeneity to pathology morphology. Based on six public scRNA-seq datasets as discovery cohorts, with CNVs inferred by the inferCNV algorithm, representing computational inference and association-level evidence.

A multimodal AI prognostic model based on routine H&E-stained pathological sections was constructed and effectively predicted patient survival, with the model-derived risk score showing a significant but modest correlation with transcriptomic HCNV activity. By combining deep learning with image analysis and using HCNV activity to guide pathology feature selection, molecular-level heterogeneity information can be channeled through routinely available clinical slides. Model performance and correlation come from the study's own modeling and validation analyses, with the correlation described in the text as significant but modest.

Multi-algorithm feature selection identified RTN3 as the core gene driving the HCNV phenotype, and the study confirms that RTN3 both orchestrates an immunosuppressive microenvironment and directly mediates gemcitabine resistance. It converges the HCNV phenotype onto a concrete, actionable molecular target, offering a candidate therapeutic target for chemoresistance. Combines multiomics data screening with in vitro and in vivo functional experiments using bladder cancer cell lines.

Mechanistically, RTN3 activates the JAK2/STAT3 signaling pathway, which transcriptionally upregulates key glycolytic enzymes PKM2, GLUT1 and LDHA, driving glycolytic metabolic reprogramming and ultimately conferring gemcitabine resistance in bladder tumors, establishing the RTN3/JAK2/STAT3/glycolysis axis as a critical resistance mechanism. It links the HCNV-associated malignant phenotype to metabolic reprogramming and chemoresistance as a testable pathway. Mechanistic conclusions are supported by in vitro and in vivo functional experiments, constituting experimental mechanistic evidence.

This work addresses prognostic assessment and chemoresistance mechanisms in bladder cancer: for researchers, it offers a single-cell-to-pathology integration framework anchored on HCNV activity and a mechanistic lead, the RTN3/JAK2/STAT3/glycolysis axis, that can be further tested; for clinical translation, its multimodal AI model is built on routine H&E-stained sections, suited to settings where existing pathology resources are used for risk stratification. Results rest mainly on public single-cell datasets, multiomics screening, and in vitro and in vivo experiments in bladder cancer cell lines, so the applicable setting should be understood as the bladder cancer cohorts and experimental systems defined by this study.

Readers may still note: the correlation between the model risk score and transcriptomic HCNV activity is significant but modest, so how far the model can serve as a proxy for HCNV remains to be observed; inferences about HCNV subpopulation communication with the microenvironment rely on computational algorithms and their biological meaning awaits further experimental support; and validation of the RTN3/JAK2/STAT3/glycolysis axis comes mainly from bladder cancer cell-line experiments, so extension to patient tumors and clinical resistance remains an open question. In addition, the available text is summary-level and lacks figures and specific statistical details, so readers seeking the concrete magnitude of model performance and experimental effects should consult the original figures and supplementary materials.