Comparison study of differential abundance testing methods using two large Parkinson disease gut microbiome datasets derived from 16S amplicon sequencing.

BACKGROUND: Testing for differential abundance of microbes in disease is a common practice in microbiome studies. Numerous differential abundance (DA) testing methods exist and range from traditional statistical tests to methods designed for microbiome data. Comparison studies of DA testing methods have been performed, but none performed on microbiome datasets collected for the study of real, complex disease. Due to this, DA testing was performed here using various DA methods in two large, uniformly collected gut microbiome datasets on Parkinson disease (PD), and their results compared.
RESULTS: Overall, 78-92% of taxa tested were detected as differentially abundant by at least one method, while 5-22% were called differentially abundant by the majority of methods (depending on dataset and filtering of taxonomic data prior to testing). Concordances between method results ranged from 1 to 100%. Average concordance for datasets 1 and 2 were 24% and 28% respectively, and 27% for replicated DA signatures. Concordances increased when removing rarer taxa before testing, increasing average concordances by 2-32%. Certain methods consistently resulted in higher concordances (e.g. ANCOM-BC, LEfSe), while others consistently resulted in lower (e.g. edgeR, fitZIG). Hierarchical clustering revealed three groups of DA signatures that were (1) replicated by the majority of methods on average and included taxa previously associated with PD, (2) replicated by a subset of methods and included taxa largely enriched in PD, and (3) replicated by few to one method(s).
CONCLUSIONS: Differential abundance tests yielded varied concordances, and amounts of detected DA signatures. Some methods were more concordant than others on both filtered and unfiltered data, therefore, if consistency with other study methodology is a key goal, one might choose among these methods. Even still, using one method on one dataset may find true associations, but may also detect false positives. To help lower false positives, one might analyze data with two or more DA methods to gauge concordance, and use a built-in replication dataset. This study will hopefully serve to complement previously reported DA method comparison studies by implementing and coalescing a large number of both previously and yet to be compared methods on two real gut microbiome datasets.