Optimizing methods and dodging pitfalls in microbiome research

Research on the human microbiome has yielded numerous insights into health and disease, but also has resulted in a wealth of experimental artifacts. Here, we present suggestions for optimizing experimental design and avoiding known pitfalls, organized in the typical order in which studies are carried out. We first review best practices in experimental design and introduce common confounders such as age, diet, antibiotic use, pet ownership, longitudinal instability, and microbial sharing during cohousing in animal studies. Typically, samples will need to be stored, so we provide data on best practices for several sample types. We then discuss design and analysis of positive and negative controls, which should always be run with experimental samples. We introduce a convenient set of non-biological DNA sequences that can be useful as positive controls for high-volume analysis. Careful analysis of negative and positive controls is particularly important in studies of samples with low microbial biomass, where contamination can comprise most or all of a sample. Lastly, we summarize approaches to enhancing experimental robustness by careful control of multiple comparisons and to comparing discovery and validation cohorts. We hope the experimental tactics summarized here will help researchers in this exciting field advance their studies efficiently while avoiding errors. Electronic supplementary material The online version of this article (doi:10.1186/s40168-017-0267-5) contains supplementary material, which is available to authorized users.

Bioinformatics processing DNA sequence data was analyzed using QIIME version 1.9.1 2 , followed by additional analysis in the R Language for Statistical Computing 3 . Read pairs were assembled to form a complete sequence for the V1-V2 variable region of the 16S rRNA gene, using a minimum overlap of 35bp and a maximum difference of 15%. The resultant sequences were quality filtered, using a minimum quality threshold of Q20. The sequences were clustered into operational taxonomic units (OTUs) using UCLUST 4 , according to the de novo OTU clustering workflow in QIIME. Taxonomic assignments were performed against the Greengenes database 5 , version 13_8, using the default method in QIIME 1.9.1. Representative sequences for each OTU were aligned with PyNAST 6 , and a phylogenetic tree was estimated with FastTree 7 . Weighted and unweighted UniFrac distances were computed between each pair of samples 8 .

Sequencing results
We collected 2.3 million total reads, with a median value of 59,000 reads per sample. Oral swab samples consisted of typical oral taxa, such as Streptococcus, Veillonella, Rothia, Prevotella, and Haemophilus (Fig. A1). Positive control samples from pond water contained a number Proteobacteria, Firmicutes, and Chloroflexi groups, which were rarely observed in the oral swab samples. Our blank extraction sample consisted predominantly of Streptophyta, which has been previously identified as a contaminant of extraction kits 9 . The community composition of positive and negative control samples was significantly different from that of oral swabs, when compared using unweighted UniFrac distance (Fig. A2, PERMANOVA test, R 2 = 0.26, P = 0.002). Results were similar with weighted UniFrac distance (R 2 = 0.51, P = 0.001).

Analysis of storage methods
The storage method used had no detectable effect on the within-sample diversity (richness, Fig. A3), when tested with a repeated-measures ANOVA (P = 0.7) or a linear mixed-effects model (P = 0.6). Conversely, the subject ID was statistically significant when included as a random effect (P < 0.001). We compared the community composition of oral samples using weighted and unweighted UniFrac distance. Based on unweighted UniFrac distance, samples from the same subject appeared more similar than samples from different subjects (Fig. A4, PERMANOVA test, R 2 = 0.47, P = 0.001). Storage method did not have a statistically significant effect on unweighted UniFrac distance (Fig. A5, P = 0.2). Our results were similar using weighted UniFrac distance: we observed a significant effect of subject ID (Fig. A6, R 2 = 0.38, P = 0.04) but not of storage method (Fig. A7, P = 0.9).

Figure A4
: Principal coordinates analysis of unweighted UniFrac distance between oral swab samples, colored by subject. Figure A5: Principal coordinates analysis of unweighted UniFrac distance between oral swab samples, colored by storage method.  To see if the method of storage was associated with alterations in specific bacterial taxa, we identified the top 12 most abundant bacterial genera in oral swab samples (Fig. A8).
We tested the effect of storage condition using a linear mixed model on log-transformed proportion (Fig. A9), with subject ID included as a random effect. We observed no statistically significant effect for any storage method on any of the top 12 taxa. The minimum p-value observed was 0.06, before correction for multiple comparisons.