Cancer Research • 2024-04-07

Unraveling sarcomatoid features in clear cell renal cell carcinoma with RNA-seq

Authors

Kirill Kryukov, Suren Davitavyan, Danil Stupichev, Anna Sharun, Anna Love, Mikhail Kleimenov, Stepan Kuznetsov, Arina Tkachuk, Vladimir Kushnarev
  1. BostonGene, Corp., Waltham, MA, US

Abstract

Introduction: Sarcomatoid сlear cell renal cell carcinoma (ccRCC) represents a challenging and aggressive subset of kidney cancers with poor prognosis and resistance to conventional treatments. Recent studies have shown the benefit of checkpoint inhibitors (CPIs) for patients with sarcomatoid components; therefore, precise and objective evaluation of sarcomatoid differentiation is essential for CPI selection. Here, we developed a predictive model to identify sarcomatoid features in ccRCC.

Methods: The machine learning-based transcriptomic sccRCC model was developed on 512 samples from publicly available datasets: TCGA-KIRC and CPTAC-ccRCC. Samples were excluded with the following criteria: inconsistent with ccRCC, less than 10% sarcomatoid component, or <20% tumor neoplastic cellularity. Sarcomatoid features for each case were assessed by 3 pathologists. The meta-cohort was divided into training (n = 348) and validation (n = 164) cohorts.The model was tested on a publicly available cohort of ccRCC patients from Sato et al. (n = 100). Clinical validation was performed using an additional prospective cohort that received BostonGene’s Tumor PortraitTM test (n = 32) with RNA-seq and whole exome sequencing (WES). The JAVELIN Renal 101 cohort was used to uncover differences in progression-free survival (PFS) associated with a predicted sarcomatoid component in advanced or metastatic ccRCCs treated with a combination of CPIs and tyrosine kinase inhibitors (TKIs).

Results: H&E analysis of the TCGA, CPTAC, and BostonGene cohorts identified zones with sarcomatoid features in 129 out of 556 (23.2%) samples. The sarcomatoid predictive model demonstrated robust performance, with weighted F1-scores of 0.85 when applied to the test cohort (Sato et al.) and 0.70 for the prospective clinical cohort. The model accurately classified 9 out of 10 total sarcomatoid samples in the prospective clinical cohort, highlighting its potential for real-world clinical applications. When applied to TCGA validation samples, sarcomatoid patients identified by the transcriptomic model exhibited median overall survival (mOS) of 55 months, whereas sarcomatoid patients based on H&E assessment had a mOS of 67 months, indicative of the model’s superior performance when detecting sarcomatoid features for OS prediction. Moreover, the sarcomatoid ccRCCs predicted by transcriptomic model that were treated with a combination of CPIs and TKIs had statistically (p=0.0001) improved PFS compared to the TKI-treated group.

Conclusion: The developed transcriptomic model effectively differentiated sarcomatoid from non-sarcomatoid ccRCC samples with robust performance metrics. With additional investigation, the predictive value of the sarcomatoid ccRCC transcriptomic model can be leveraged to guide CPI and TKI selection.
Read full publication Download