Outlier Detection for RNA-Seq Data via Improved Surprise Index

• Activity Number: 320
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Ching-Wei Chang* and Claire Boyle and Yu-Chung Wei and Nysia George
• Companies: National Center for Toxicological Research and Florida State University and NCTR and FDA/NCTR
• Keywords: surprise index ; RNA-seq ; outlier detection ; negative binomial distribution

Abstract

Recent work has highlighted the impact of outliers on differential expression analysis in RNA sequencing (RNA-seq) studies. Current methods of outlier detection for RNA-seq data utilize a test-based approach and the resulting test statistic to identify outliers. However, consideration of the empirical discrete data structure may provide more biological information. In this study, we developed a generalized modification of the surprise index (SI) to accommodate features of RNA-seq data. In lieu of applying the widely used arbitrary cutoff of 1000 to classify observations as surprising, the proposed index models cutoffs as a function of sample size and parameter estimates of the assumed discrete distribution for a given target. Simulated and real RNA-seq datasets are used to demonstrate an improved outlier detection rate using the modified SI over the universal SI cutoff and a popular competing methodology.