PIME: a package for discovery of novel differences among microbial communities

Abstract

Massive sequencing of genetic markers, such as the 16S rRNA gene for prokaryotes, allows the comparative analysis of diversity and abundance of whole microbial communities. However, the data used for profiling microbial communities is usually low in signal and high in noise preventing the identification of real differences among treatments. PIME (Prevalence Interval for Microbiome Evaluation) fills this gap by removing those taxa that may be high in relative abundance in just a few samples but have a low prevalence overall. The reliability and robustness of PIME were compare against the existing methods and verified by a number of approaches using 16S rRNA independent datasets. To remove the noise, PIME filters microbial taxa not shared in a per treatment prevalence interval starting at 5% with increments of 5% at each filtering step. For each prevalence interval, hundreds of decision trees are calculated to predict the likelihood of detecting differences in treatments. The best prevalence-filtered dataset is user-selected by choosing the prevalence interval that keeps the majority of the 16S rRNA reads in the dataset and shows the lowest error rate. To obtain the likelihood of introducing bias while building prevalence-filtered datasets, an error detection step based in random permutations is also included. A reanalysis of previews published datasets with PIME uncovered previously missed microbial associations improving the ability to detect important organisms, which may be masked when only relative abundance is considered.