This is a pre-print that just came out from several UC Davis colleagues. I actually think the title is a bit misleading because these recommendations are a lot broader than just human fecal metagenomes. Definitely worth a read!
Background Shotgun metagenomes are often assembled prior to annotation of genes which biases the functional capacity of a community towards its most abundant members. For an unbiased assessment of community function, short reads need to be mapped directly to a gene or protein database. The ability to detect genes in short read sequences is dependent on pre- and post-sequencing decisions. The objective of the current study was to determine how library size selection, read length and format, protein database, e-value threshold, and sequencing depth impact gene-centric analysis of human fecal microbiomes when using DIAMOND, an alignment tool that is up to 20,000 times faster than BLASTX.
Results Using metagenomes simulated from a database of experimentally verified protein sequences, we find that read length, e-value threshold, and the choice of protein database dramatically impact detection of a known target, with best performance achieved with longer reads, stricter e-value thresholds, and a custom database. Using publicly available metagenomes, we evaluated library size selection, paired end read strategy, and sequencing depth. Longer read lengths were acheivable by merging paired ends when the sequencing library was size-selected to enable overlaps. When paired ends could not be merged, a congruent strategy in which both ends are independently mapped was acceptable. Sequencing depths of 5 million merged reads minimized the error of abundance estimates of specific target genes, including an antimicrobial resistance gene.
Conclusions Shotgun metagenomes of DNA extracted from human fecal samples sequenced using the Illumina platform should be size-selected to enable merging of paired end reads and should be sequenced in the PE150 format with a minimum sequencing depth of 5 million merge-able reads to enable detection of specific target genes. Expecting the merged reads to be 180-250bp in length, the appropriate e-value threshold for DIAMOND would then need to be more strict than the default. Accurate and interpretable results for specific hypotheses will be best obtained using small databases customized for the research question.