Portal de Periódicos da CAPES

Abstract 2277: Machine learning to identify prostate cancer mutations for screening cell-free DNA (cfDNA)

2018; American Association for Cancer Research; Volume: 78; Issue: 13_Supplement Linguagem: Inglês

10.1158/1538-7445.am2018-2277

1538-7445

Clinton L. Cario, Lancelote Leong, Man-Tzu Wang, John S. Witte,

Molecular Biology Techniques and Applications

Abstract Cell free DNA (cfDNA), a form of ‘liquid biopsy', has recently emerged as a promising technology to screen, diagnose, and monitor many types of disease. However, in the context of cancer, one unfortunate complication of cfDNA's use as a biomarker is the weak signal-to-noise ratio that arises primarily due to DNA contamination from healthy tissue DNA. To circumvent this, many researchers have focused on a limited set of frequently observed coding variants within a single gene or a small panel of hand selected genes, trading breadth for depth to improve signal. This approach, while sensible, can lead to detection issues in highly heterogeneous diseases— such as prostate cancer— in which variants may not be universally present at high proportion across all patients. Furthermore, many of these panels ignore regions that are non-coding but functional, and possibly even crucial in tumorigenesis. This project proposes a different, agnostic, and machine-learning based approach for panel generation. Using public whole genome sequence (WGS) datasets containing single nucleotide polymorphisms (SNPs) and copy variant information from the International Cancer Genome Consortium (ICGC) as training data and a large number of annotation features collected from a dozen biological databases, we've developed an open source push-button tool called orchid to build cancer variant prediction models. After training a Support Vector Machine (SVM) on known prostate cancer mutations in patients with few mutations (putatively driver enriched) and those in patients with many mutations (putatively passenger enriched), the model was used to suggest a set of mutations most “prostate cancer like”. Mutations furthest from the classifying hyperplane are then selected to form the SVM panel. Compared to manually curated panels, the SVM panel demonstrates superior in silico patient detection in both untrained ICGC data and in tumor/normal sequence data from a pilot study of 13 prostate cancer patients with multiple heterogeneous tumor foci. Moving forward, the SVM generated panel will be used to screen cfDNA from patients in the pilot study and 100 additional UCSF prostate cancer patients. This research has clinical implications in creating diagnostic, prognostic, and predictive tools for prostate cancer. In particular, the SVM panel may be used to screen cfDNA for cancer mutations and to assess tumor heterogeneity and residual disease. Additionally, the published software tool orchid can be used for other tumor mutation classification tasks, such as determining tissue-of-origin from cfDNA fragments using only the locations of mutations in the fragments. Citation Format: Clinton L. Cario, Lancelote Leong, Man-Tzu Wang, John Witte. Machine learning to identify prostate cancer mutations for screening cell-free DNA (cfDNA) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 2277.