Introduction:

Muscle-invasive bladder cancer (MIBC) is a heterogenous disease that can be broadly categorized into five molecular subtypes associated with varying response to chemotherapy and immunotherapy and clinical outcomes. While significant advances have been made in comprehensive genomic profiling of MIBC, there is an unmet need for high throughput clinical grade assays to classify MIBC using formalin-fixed paraffin-embedded (FFPE) patient biopsy tissue. We hypothesized that MIBC subtypes can be accurately predicted with a custom NanoString codeset based on the TCGA single patient classifier.

Methods:

We designed a custom NanoString codeset based on the reduced gene set used for the TCGA RNA-based single patient classifier for MIBC subtyping (Kim et al. 2019). We ran the codeset aligned to 359 genes (TCGA359) on 48 RNA samples extracted from FFPE tissue from the SWOG S-1011 cohort (NCT 01224665). Whole transcriptome RNA-sequencing was performed to determine the “true” tumor subtype based on the TCGA single patient RNA-based classifier. We revised the codeset to target 399 genes (TCGA399) to optimize the assay’s fidelity to the RNA-seq subtype calls and ran the revised assay on these same 48 samples. We then processed the raw data using nCounter software and assigned the TCGA subtype using the published TCGA-classification script in R. To assess the agreement between the predicted subtypes from NanoString codeset and RNA-sequencing, classification error and accuracy were determined, resulting in a measure of concordance between the two profiling methods.

Results:

RNA sequencing and subtyping based on the TCGA patient classifier demonstrated the following “true” tumor classifications in the 48 samples: 21 (44%) basal/squamous, 3 (6%) luminal, 12 (25%) luminal infiltrated, 11 (23%) luminal papillary, and 1 (2%) neuronal. The subtype classification for the TCGA359 codeset included 19 (40%) basal/squamous, 4 (8%) luminal, 8 (17%) luminal infiltrated, 16 (33%) luminal papillary, and 1 (2%) neuronal.  The subtype classification for the TCGA399 codeset included 20 (43%) basal/squamous, 1 (2%) luminal, 8 (17%) luminal infiltrated, 16 (34%) luminal papillary, and 2 (4%) neuronal (Table 1). TCGA359 achieved 83% accuracy with a classification error of 17% when compared to the RNAseq subtype calls. TCGA399, with an additional 40 genes, showed an accuracy of 74%, and classification error of 26%. Luminal vs basal subtype prediction accuracy was 95.7% for TCGA 359 and 91.3% for TCGA399, respectively.

Conclusion:

Our custom NanoString codeset demonstrated accurate subtype calls based on the TCGA single patient classifier for patients with MIBC. TCGA399 did not improve accuracy of tumor subtype classification compared to TCGA359 despite the additional probes designed to increase precision of subtype calls. Nonetheless, this NanoString based classifier may provide faster throughput and reduced cost for subtyping and can be incorporated into clinical trials testing utility of molecular subtypes. Validation of this platform for tumor subtyping in MIBC in the full SWOG S-1011 cohort with comparison to RNA seq subtype calls is in progress.

Funding: NIH/NCI grants U10CA180888, U10CA180819, U24CA196175