Thursday, December 8, 2016

BISTRO - Hongjiu Zhang

4:00 PM

2036 Palmer Commons

BISTRO is restricted to U-M Bioinformatics Graduate Program students and faculty.

"An ultra-fast tumor heterogeneity inference method using density-based infinite mixture model"

Abstract

Tumor development, being a rapid evolution process, involves mixtures of different cell subpopulations carrying distinct sets of mutations. Such innate heterogeneity may render a particular treatment ineffective. We developed a subclone inference algorithm that quickly predicts the number, the compositions, and the phylogenetic relationship of the subclones from bulk sequencing data. The algorithm is tested in the DREAM SMC-Het Challenge and ranked 1st in the final benchmark test, outperforming many complicated models. The method evolves from well-established mixture models over allelic frequencies of somatic mutations, but uses kernel density estimation to simplify the subclone detection. This allows us to iteratively determine the full mixture model, which reduces runtime by 100-fold in comparison with Monte-Carlo simulation methods. We further simplified the phylogeny inference using a greedy algorithm. The whole inference process takes only a few seconds. In the post challenge phase, we tested the robustness of this software with tumors collected from collaborative hospitals. In this case, the real-world tumors are represented by much less mutations (~200 mutations, compared to up to thousands in the simulated data). While existing methods often report unrealistic number of clusters, our new method identified clusters well supported by mutation frequency distributions.