GPU accelerated Bayesian mixture models for FCM analysis

We describe the integration of Bayesian non-parametric mixture models, massively parallel computing on GPUs and software development in Python to provide an extensible toolkit for automated statistical analysis in high-dimensional flow cytometry (FCM). The use of standard Bayesian non-parametric Dirichlet process mixture models allows the flexible density estimation of the posterior distribution (MCMC) or modes (EM) of high-dimensional FCM data, and provides a coherent statistical framework for data analysis and interpretation (1,2,3). By exploiting the massively parallel nature of GPUs to achieve greater than 100 fold speed-ups over serial code, it is now realistic to perform large scale data analysis using these methods (4). To facilitate dissemination to the computational cytometry community, the statistical and computational foundations have been wrapped into fcm (5), a Python library that makes it simple to design and implement analysis and visualization pipelines for FCM (6). Used interactively, Python is ideal for rapid prototyping and algorithm development. In addition, Python has robust support for databases, networking, XML processing, GUIs and web development, facilitating the development of full stack applications (7). Together, statistical mixture models, GPU computing and Pytho