Abstract PR011: Imaging mass cytometry captures patient heterogeneity enabling BCG response stratification in non-muscle invasive bladder cancer

• Published in: Clinical cancer research Vol. 30; no. 10_Supplement; p. PR011
• Main Authors: Pour, Ali Foroughi, Baker, Dylan, Sivajothi, Santhosh, Choy, Bonnie, Meghani, Khyati, Yu, Yanni, Fall, Leigh A., Ristau, Benjamin, Meeks, Joshua J., Robson, Paul, Chuang, Jeffrey H.
• Format: Journal Article
• Language: English
• Published: 17.05.2024
• ISSN: 1557-3265; 1557-3265
• DOI: 10.1158/1557-3265.BLADDER24-PR011

Abstract BCG immunotherapy is the standard of care for high-risk non-muscle invasive bladder cancer (NMIBC). The mechanisms of response to BCG are not fully understood. Patient response to BCG is highly variable with ~30-40% of individuals experiencing recurring or progressing disease after treatment within 24 months. Prognostic markers of response to treatment are not fully understood and remain insufficient for patient stratification. In order to develop an understanding of BCG response by comprehensive assessment of the tumor context. We performed single cell spatial proteomic profiling via imaging mass cytometry (IMC) on tissue microarrays of tumor biopsies from 53 bladder cancer patients (36 responding and 17 non-responding). We then validated our finding on an external cohort of whole pathology slides from 10 patients (4 responding and 6 non-responding). The IMC panel was developed based on scRNAseq profiles of bladder tumor samples to provide an unbiased cellular context of NMIBC. Additionally, IMC data is accompanied by hematoxylin and eosin (H&E) images of adjacent cuts and annotated by a clinical pathologist to distinguish tumor, immune, and stromal regions. We first developed a convolutional neural network (CNN) to reproduce pathologist annotations of tumor, stroma, and immune regions from H&E data with high accuracy (AUC>93%). We then registered IMC and annotated H&E images to train a random forest model to predict pathologist annotations using IMC data (AUC>95%). To determine the cellular composition of TMA spots, we used sparse non-negative matrix factorization (sNMF) to encode and cluster IMC data within tumor, stroma, and immune regions. NMF analysis revealed that a component defined primarily by KRT5, CD11b, and CD14, specifically within tumor-annotated regions adjoining COL1A1+ stromal regions, is a prognostic marker of poor response (p-value<0.05). This is consistent with previous findings that tumor-associated myeloid cells indicate poor prognosis. Additional NMF components of the immune annotated region showed weak associations with response. Immune regions defined by CD68, CD11b, CD31, and PDPN corresponded with poor BCG response, indicating myeloid-stromal-vascular interactions outside the tumor influence outcome, while vascular immune regions defined by high aSMA presence correspond with better response. Collectively, combining NMF components to generate a single response score for responding and non-responding patients (p-value<0.05) proved more effective at patient stratification than any single NMF component alone. The external cohort validated our random forest model for separating tumor, stroma, and immune regions, our sNMF model for clustering IMC images, and our response score (p-value<0.1). Citation Format: Ali Foroughi Pour, Dylan Baker, Santhosh Sivajothi, Bonnie Choy, Khyati Meghani, Yanni Yu, Leigh A. Fall, Benjamin Ristau, Joshua J. Meeks, Paul Robson, Jeffrey H. Chuang. Imaging mass cytometry captures patient heterogeneity enabling BCG response stratification in non-muscle invasive bladder cancer [abstract]. In: Proceedings of the AACR Special Conference on Bladder Cancer: Transforming the Field; 2024 May 17-20; Charlotte, NC. Philadelphia (PA): AACR; Clin Cancer Res 2024;30(10_Suppl):Abstract nr PR011.