Evaluation of computational methods for human microbiome analysis using simulated data.

Testing of eight pipelines against 144 simulated metagenomic datasets initially produced 1,104 discrete results. Pipelines using a marker gene strategy; MetaPhlAn2 and ConStrains, were overall less sensitive, than other pipelines; with the notable exception of Taxator-tk. This difference in sensitivity was largely made up in terms of runtime, significantly lower than more sensitive pipelines that rely on whole-genome alignments such as PathoScope2.0. However, pipelines that used strategies to speed-up alignment between genomic references and metagenomic reads, such as kmerization, were able to combine both high sensitivity and low run time, as is the case with Kraken and Centrifuge. Absent species genomes in the database mostly led to assignment of reads to the most closely related species available in all pipelines. Our results therefore suggest that taxonomic profilers that use kmerization have largely superseded those that use gene markers, coupling low run times with high sensitivity and specificity. Taxonomic profilers using more time-consuming read reassignment, such as PathoScope 2.0, provided the most sensitive profiles under common metagenomic sequencing scenarios. All the results described and discussed in this paper can be visualized using the dedicated R Shiny application (https://github.com/microgenomics/HumanMicrobiomeAnalysis). All of our datasets, pipelines and results are made available through the GitHub repository for future benchmarking.