Bioinformatics Lecture: Geyu Zhou

Bayesian Transfer Learning with Two Dimensional Dirichlet Process Random Effects Model for Genetic Prediction of Complex Traits

Speaker: Geyu Zhou

Time: September 25th , 12:30-1:30 pm EDT.

Location: this event is virtual (see Zoom link below) 

Zoom Link https://purdue-edu.zoom.us/j/97231496999

Transfer learning is a machine learning technique to use knowledge learned from a task to boost performance on a related task. In recent years, the development and success of transfer learning have largely been driven by deep learning and its application in image recognition and natural language processing. Here, we present an alternative Bayesian transfer learning framework based on the two-dimensional Dirichlet process random effects model to jointly estimate effects from two datasets by assuming the joint distribution to be zero, dataset 1 specific, dataset 2 specific, or correlated between two datasets. Compared with deep learning, the proposed framework has four advantages: 1) consistency with the lack of detected epistasis (non-linear gene by gene interaction) in current human genetics data, 2) allowing the use of privacy-preserved sufficient statistics, 3) easier computation for high dimensional datasets, 4) no need for parameter tuning. We demonstrate the effectiveness of our framework for predicting the value of human complex traits in three cases. First, SDPRX is developed to improve the genetic prediction in non-European populations by jointly utilizing genetic data from European and non-European populations. Second, SDPR_admix is developed to improve the genetic prediction in admixed populations by leveraging the shared and distinct genetic information across local ancestries. Third, PleioSDPR is developed to improve the genetic prediction for two correlated traits by characterizing trait properties. This is a joint work with my former advisor Hongyu Zhao and other former lab members. 

Bio: Geyu Zhou is an Assistant Professor of Biology and Statistics from Purdue University. He received his PhD degree in Computational Biology and Bioinformatics, followed by postdoc training in Biostatistics from Yale University. He currently works on genetic risk prediction, admixture mapping, and gene by environment analysis in admixed populations. Dr. Zhou’s research is supported by the AnalytixIN fellowship.  

Website: https://eldronzhou.github.io/

Host: Boran Gao (Stat/Bio)

Coordinator: Daisuke Kihara (Bio/CS), Majid Kazemian (Biochem/CS)

Past lectures are posted on the Youtube channel: https://www.youtube.com/channel/UCK4OqdS6xAxTCiVev843f6Q

(the channel includes only lectures approved by the speakers)