Relative transcript abundance has proven to be a valuable tool for understanding the function of genes in biological systems. For the differential analysis of transcript abundance using RNA sequencing data, the negative binomial model is by far the most frequently adopted. However, common methods that are based on a negative binomial model are not robust to extreme outliers, which were found to be abundant in public datasets. So far, no rigorous and probabilistic methods for detection of outliers have been developed for RNA sequencing data, leaving the identification mostly to visual inspection.
Recent advances in Bayesian computation allow large-scale comparison of observed data against its theoretical distribution given in a statistical model. Here researchers from the Walter and Eliza Hall Institute, propose ppcseq, a key quality-control tool for identifying transcripts that include outlier data points in differential expression analysis, which do not follow a negative binomial distribution. Applying ppcseq to analyse several publicly available datasets using popular tools, the researchers show that from 3 to 10 percent of differentially abundant transcripts across algorithms and datasets had statistics inflated by the presence of outliers.
Flow chart of the two-step strategy for outlier detection, including discovery and test steps
Because a model that includes outliers is ill-posed by definition, a first discovery step allows the flagging of potential outliers with relaxed criteria, while a second test step allows the evaluation of those potential outliers against a model fitted without them. The workflow includes a preliminary independent estimation of differential gene transcriptional abundance with methods such as edgeR or DESeq2. Genes which outliers will be selected from the significance rank. The first step of the outlier identification includes the fitting of the user-defined linear model on the user gene-selection. Then, the theoretical data distribution is generated from the join posterior, and genes are flagged as potential outliers with a default false positive rate threshold of 5%. Of those, only detrimental outliers (see ‘Materials and Methods’ section) are flagged. The test step includes the removal of possible detrimental outliers from the data, and the fit of the same model, compensating for data truncation. Then, the theoretical data distribution is generated from the join posterior and potential outliers are checked against, with a better calibrated false positive rate (0.01 by default).
Availability – The code used to conduct the analyses is available at github.com/stemangiola/ppcseq.