Abstract 2192: Transcriptomic classification of renal cancer: a machine learning approach

Abstract

Abstract Introduction: Accurate diagnosis is essential for cancer treatment. Sometimes, it is difficult to confirm histological diagnosis due to inadequate tissue sample or preparation. Using molecular markers for classification would be ideal in such cases. For renal cell carcinoma(RCC), we developed a molecular classification based on transcriptomics to address these issues and aid current diagnostic workflow. Method: We obtained transcriptomic and phenotype data of renal cell carcinoma from The cancer genome atlas (TCGA) data repository. We developed an unsupervised algorithm, Density-based UMAP, based on two clustering methods, UMAP (Uniform manifold approximation and projection) and DBScan (Density-based spatial clustering of applications with noise). We iterated this algorithm 1000 times with random 1000 genes each time and classified each sample in each iteration. We used plurality voting of at least 70% of the iterations for consensus groups. The algorithm was able to identify all the major histological subtypes. We ran a differential expression analysis for each group. Finally, we developed a classification gene signature and a supervised algorithm to classify subtypes with fewer genes and implement this signature in an RCC metanalysis. Results: Density-based UMAP Algorithm classified the samples with 91.4% concordance with WHO 2016 classification. Among the discrepant cases, 4.4 % cases (46) were not diagnosed the same as their histological classification. 4.2 % (44) cases were shown a mixed expression profile. These cases are of further interest as they might respond better with a different treatment. Based on molecular characteristics they share with major subtypes, they are named ClCh, Clear 2, and Clear 3. ClCh and Clear 2 have shown to be indolent and, Clear 3 has shown to be very aggressive. We manually curated 328 genes for supervised learning from differential expression analysis between the groups and able to identify the groups in a metanalysis. Conclusion: Transcriptomic profiling is fast, robust and, simple. It can help in diagnosing histologically discrepant cases. This classifier would be a good edition in renal cancer diagnostic workflow and can complement current diagnostic guidelines for renal cancer. Citation Format: Khaled Bin Satter, Paul MH Tran, Sharad B. Purohit, Jin-Xiong She. Transcriptomic classification of renal cancer: a machine learning approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 2192.