Tumor progression and therapy resistance are regulated by the tumor microenvironment (TME), a multicellular ecosystem shaped by complex cellular interactions. Recent advances in single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) have revolutionized our understanding of tumor heterogeneity. While current studies focus on isolated TME components, the clinical significance of coordinated multicellular dynamics remains underexplored. We aim to develop a universal algorithm for decoding clinically relevant multicellular programs (MCPs) in muscle-invasive bladder cancer (MIBC).

We presented SpatioMCPnet, a framework for identifying MCPs from scRNA-seq and ST data. Spatial cell compositions were resolved by Cell-type Assignment by Robust Decomposition (CARD) deconvolution and harmonized via Isometric Log-Ratio transformation (ILR). MCPs were identified through Graph Neural Network-Convolutional Neural Network (GNN-CNN) co-embedding with iso-depth boundaries, and marker genes were extracted using transcriptional saliency mapping. Spatial MCPs were functionally annotated via MsigDB database. Cell-cell communication within spatial niches was inferred via optimal transport (OT)-based ligand-receptor signaling. For bulk RNA-seq, niche abundance was quantified by ν-Support Vector Regression (ν-SVR) deconvolution and linked to prognosis via kernelized Cox regression. The multimodal framework integrated spatial transcriptomic data and histopathological features via Graph Attention Network (GAT)-OT fusion. Clinical translation is achieved by mapping spatial MCPs to H&E whole-slide images via U-Net.

We conducted scRNA-seq on 15 MIBC specimens and ST on 4 paired samples. SpatioMCPnet identified a tumor-invasive interface-enriched MCP, defined as TAM_TIMP1-CAF niche. This niche was functionally enriched in extracellular matrix (ECM) remodeling pathways and active TIMP1-CD63 interaction. We next recovered and quantified the TAM_TIMP1-CAF niche in the TCGA-MIBC and IMvigor210 cohorts. This niche was significantly associated with poor outcomes and anti-PD-L1 therapy resistance. The GAT-OT fusion model achieved high prediction ability of the abundance of TAM_TIMP1-CAF niche on H&E whole-slide images.

This study establishes a novel computational framework to decipher spatial multicellular ecosystems in MIBC. We identified therapeutically targetable niches, such as the TAM_TIMP1-CAF niche driving MIBC therapeutic resistance, and demonstrated their clinical detectability in histopathology. Our work shifts the paradigm from cell-centric to functional tissue unit-based oncology, advancing precision medicine.