Deep Learning Bridges Histology and Transcriptomics to Predict Molecular Subtypes and Outcomes in Muscle-Invasive Bladder Cancer

Abstract

Muscle-Invasive Bladder Cancer (MIBC) is a heterogeneous disease with distinct molecular subtypes influencing prognosis and therapeutic response. However, molecular profiling through RNA sequencing remains costly, time-consuming and complicated by intratumoral heterogeneity. We developed a Deep Learning (DL) approach to infer molecular subtypes from routine histopathological slides and to evaluate its prognostic value in patients treated with neoadjuvant chemotherapy (NAC). We developed an DL-model predicting the expression of 848 subtype-associated genes from histological images of transurethral resection of bladder tumor, enabling spatial molecular subtyping at tile level. The model was trained on 297 NAC-treated patients from the VESPER clinical trial and evaluated on three independent cohorts (COBLAnCE, n=224; Saint-Louis, n=30 and TCGA, n=315), covering diverse staining protocols and scanner types. Spatial transcriptomics from six VESPER patients confirmed the spatial consistency of the inferred expression profiles. Our approach achieved a ROC AUC of 0.94 for molecular subtype prediction, with 95% of genes significantly predicted, demonstrating its ability to capture transcriptomic dysregulations from histological morphology. Predicted expression maps revealed spatially coherent patterns and intratumoral molecular heterogeneity. Importantly, tumors predicted with basal/squamous features (pure or mixed), were associated with significantly worse progression-free and overall survival after NAC (log-rank p=0.014 and 0.037, respectively). This DL-based framework enables accurate and spatially resolved inference of gene expression and molecular subtypes in MIBC without sequencing. These findings could improve patient stratification in clinical practice and support the design of more targeted clinical trials. Further validation in larger cohorts is needed before routine clinical implementation.