Date of Award

Fall 2014

Document Type

Thesis

Degree Name

Master of Science in Bioinformatics - (M.S.)

Department

Computer Science

First Advisor

Usman W. Roshan

Second Advisor

Jason T. L. Wang

Third Advisor

Zhi Wei

Abstract

Quantitative phenotypes prediction from genotype data is significant for pathogenesis, crop yields, and immunity tests. The scientific community conducted many studies to find unobserved quantitative phenotype high predictive ability models. Early genome-wide association studies (GWAS) focused on genetic variants that are associated with disease or phenotype, however, these variants manly covers small portion of the whole genetic variance, and therefore, the effectiveness of predictions obtained using this information may possibly be circumscribed [ 1 ].

Instead, this study shows prediction ability from whole genome single nucleotide polymorphisms (SNPs) data of 1940 genotyped stoke mouse with - 12k SNPs, and 413 genotyped rice inbred lines with - 40k SNPs. The predictive accuracy measured as the Pearson coefficient correlation between predicted phenotype and actual phenotype values using cross validation (CV), and found a predictive ability for mouse phenotypes MCH, CD8 to be 0.64 and 0.72, respectively.

The study compares whole genome SNPs data prediction methods built using Support Vector Regression (SVR) and Pearson Correlation Coefficient (PCC) to perform SNPs selection and then predict unobserved phenotype using ridge regression and SVR. The investigation shows that ranking SNPs by SVR significantly increases predictive accuracy than ranking with PCC. In general, Ridge Regression perform slightly better prediction ability than predicting with SVR.

Share

COinS