Spatial Immune Profiling and AI-Based Classifiers Identify Predictors of BCG Therapy Outcomes in High-Risk Non-Muscle-Invasive Bladder Cancer

TL;DR

This study uses advanced imaging and AI to identify immune and stromal features that predict whether BCG therapy will work for high-risk bladder cancer patients.

Contribution

The study introduces a novel AI-based model (IMC-GA-MIL) and spatial profiling to predict BCG response in bladder cancer.

Findings

BCG nonresponse is linked to a tumor microenvironment rich in fibroblasts and plasma cells.

Immune cells within tumor regions are associated with BCG response and better survival.

The IMC-GA-MIL model predicted BCG response with 90% accuracy using immune and myeloid markers.

Abstract

High-risk non-muscle-invasive bladder cancer (NMIBC) frequently recurs and may progress, requiring intensive surveillance. Bacillus Calmette–Guérin (BCG) immunotherapy is the standard treatment for high-risk disease; however, reliable biomarkers of treatment failure are lacking. We analyzed 82 high-risk NMIBC tumors treated with BCG using Hyperion imaging mass cytometry (IMC) and evaluated the data using two complementary approaches: spatial single-cell phenotyping and an IMC-specific AI-based gated attention multiple instance learning model (IMC-GA-MIL) to predict BCG response. Using single-cell IMC data, we characterized tumor, immune, and stromal phenotypes and examined their associations with BCG response and survival. BCG nonresponse was associated with a tumor microenvironment enriched in fibroblasts and plasma cells, whereas BCG response was associated with immune cells localized within tumor regions. The IMC-GA-MIL model identified marker patterns consistent with immunosuppressive biology as key predictive features of BCG response.

Background/Objectives: High recurrence rates and intensive lifelong surveillance make bladder cancer among the costliest malignancies to treat. Although Bacillus Calmette–Guérin (BCG) immunotherapy is the standard treatment for high-risk non-muscle-invasive bladder cancer (NMIBC), up to 50% of patients fail to respond, and predictive biomarkers are lacking. Molecular profiling has established three BCG response subtypes (BRS1–3), with BRS3 characterized by an immunosuppressive, BCG-resistant phenotype; however, these features have not been validated at single-cell spatial resolution. Methods: We applied imaging mass cytometry (IMC) to 82 BCG-treated high-risk NMIBC samples and performed (i) single-cell IMC with unsupervised clustering to identify phenotypic cell clusters and quantify cluster abundances and (ii) a convolutional neural network-based gated attention multiple instance learning model trained on IMC images (IMC-GA-MIL) to predict BCG response. Cluster abundances were summarized using II (immune composition within the immune compartment), TT (tumor phenotypic composition), and IT (immune/stromal abundance relative to tumor cells) indices. Results: Single-cell IMC identified 18 distinct phenotypic cell clusters. In BCG responders, immune cells localized within the tumor compartment were enriched and independently protective (HR 0.67, 95% CI 0.49–0.92). BCG nonresponse was associated with a higher abundance of fibroblast-dominant clusters relative to tumor cells (IT index). Plasma cell-dominant clusters were the strongest predictors of progression (II index HR 2.28, 95% CI 1.37–3.79; IT index HR 1.25, 95% CI 1.06–1.48). The IMC-GA-MIL model predicted BCG response with 90% accuracy (9/10) and identified myeloid- and T-cell-associated marker patterns involving CD14, CD11b, CD68, CD8, and FOXP3 as the most informative contributors. Conclusions: Spatial single-cell profiling and IMC-GA-MIL identify spatial immune and stromal features associated with BCG failure. However, findings from both analyses should be considered exploratory and will require validation in larger, independent cohorts.

Figures