Supplementary Figure 3: HUMAnN2 tiered search performance on human metagenomes.

We applied HUMAnN2′s tiered search to profile 397 first-visit HMP metagenomes on Harvard University’s Odyssey Research Computing Cluster (8 CPU cores per job). Sample counts per body site were as follows: 54 for anterior nares, 65 for buccal mucosa, 68 for supragingival plaque, 73 for tongue dorsum, 76 for stool, and 34 for posterior fornix. (a) At most body sites, ~ 60% of reads were explained by detected pangenomes, with (b) an additional ~ 20% explained by downstream translated search (~80% total). Pangenome search performance (c) consistently exceeded translated search performance (d) by 1–2 orders of magnitude. From smallest to largest, box plot elements in panels a–d represent the lower inner fence, first quartile, median, third quartile, and upper inner fence. Horizontal red lines indicate the median value over all samples. (e) Total runtime is largely dictated by the number of reads passed to translated search, and (for HMP samples with < 100 million reads) was approximately linear in the number of input reads (~1 h/5 million input reads). (f) Peak memory use was sublinear in the number of input reads and very predictable. The cluster of outliers in f results from large samples that were requeued during their runs: these samples resumed later in the HUMAnN2 workflow and hence display smaller peak memory use