A0917
Title: Bayesian clustering of prostate cancer patients with simultaneous feature selection of metabolites
Authors: Inga Huld Armann - Imperial College London (United Kingdom) [presenting]
Elizabeth Bancroft - The Royal Marsden and Institute of Cancer Research (United Kingdom)
Zsofia Kote-Jarai - Institute of Cancer Research (United Kingdom)
Ros Eeles - Institute of Cancer Research and The Royal Marsden (United Kingdom)
Ioanna Papatsouma - Imperial College London (United Kingdom)
Marina Evangelou - Imperial College London (United Kingdom)
Abstract: Prostate cancer (PrCa) is the most common cancer in men in the UK, with 55,000 cases and 12,000 deaths per year. Leveraging data obtained from accessible screening methods, such as blood tests, would enhance early diagnosis and improve patient care. Metabolomics extracted from biofluids such as blood offer a promising and easily accessible source of data and have been studied alongside other 'omics' datasets to advance cancer research. Metabolomics data from two case-control studies on PrCa, namely, PROFILE and IMPACT, were analyzed. Building on the success of previous analyses of metabolomics data for PrCa, clustering techniques are employed to identify potential cancer subtypes, particularly those linked to disease aggressiveness. Motivated by the importance of biomarker identification, Bayesian nonparametric clustering via Dirichlet process mixture models is proposed, with an extension allowing for simultaneous feature selection. The extension aims to identify features contributing to the clustering of observations. This is an important insight into potential biomarkers. The posterior is approximated through variational inference, offering a computationally efficient alternative to traditional Markov Chain Monte Carlo methods. The effectiveness of the proposed method is evaluated on both real and simulated data. Moreover, the method is applicable to other 'omics' datasets, including genomics.
