Spatial architecture contributes to failure of bulk biomarker-guided neoadjuvant immunotherapy selection in bladder...

Predictive biomarkers for immune checkpoint inhibitors (ICIs) are largely identified retrospectively, but their prospective clinical utility remains unproven. Here, we report DUTRENEO, a prospective randomized phase 2 trial testing whether a retrospectively validated 18-gene bulk tumor inflammation signature (TIS) can guide neoadjuvant ICI therapy in muscle-invasive bladder cancer. The trial does not meet its primary endpoint, and this demonstrates that bulk gene-expression stratification does not sufficiently enrich for responders. To define the biology underlying this failure, we generate single-cell spatial transcriptomic profiles of 377 genes in ∼5.4 million cells across large tissue areas. Response is governed by spatial architectures invisible to bulk assays, including CD8⁺ T cell proximity to cancer cells, localized checkpoint co-expression within epithelial cancer-rich neighborhoods, and fibroblast-rich immune-excluded communities in non-responders. We provide a quantitative framework showing that ≥77 genes and ≥3-mm tissue diameter regions preserve predictive spatial signal at scalable throughput. The registration details of the trial are EudraCT 2017-002246-68.