henhaobio
Sign in

bioRxivBioinformaticsPreprint

PICRUSt2: An improved and customizable approach for metagenome inference

Gavin M. DouglasiD, Vincent J. Maffei, Jesse ZaneveldiD, Svetlana N. Yurgel, James R. Brown, Christopher M. TayloriD, Curtis HuttenhoweriD, Morgan G. I. LangilleiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

One major limitation of microbial community marker gene sequencing is that it does not provide direct information on the functional composition of sampled communities. Here, we present PICRUSt2 (https://github.com/picrust/picrust2), which expands the capabilities of the original PICRUSt method1 to predict the functional potential of a community based on marker gene sequencing profiles. This updated method and implementation includes several improvements over the previous algorithm: an expanded database of gene families and reference genomes, a new approach now compatible with any OTU-picking or denoising algorithm, and novel phenotype predictions. Upon evaluation, PICRUSt2 was more accurate than PICRUSt1 and other current approaches overall. PICRUSt2 is also now more flexible and allows the addition of custom reference databases. We highlight these improvements and also important caveats regarding the use of predicted metagenomes, which are related to the inherent challenges of analyzing metagenome data in general.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (20 Mar 2020) · CC BY 4.0

PICRUSt2 places 16S rRNA variants into a 20,000-gene reference tree, predicting gene families, pathways and phenotypes; gene-family predictions matched or beat four tools against shotgun metagenomes.

16S rRNA amplicon surveys lack direct functional readouts, shotgun metagenomics is impractical for host-contaminated, low-biomass or budget-limited samples, and PICRUSt1 needed closed-reference Greengenes operational taxonomic units (OTUs) and databases unchanged since 2013. PICRUSt2 accepts OTUs or amplicon sequence variants (ASVs) via phylogenetic placement, uses >20-fold more genomes, raises KEGG orthologs from 6,911 to 10,543, and adds custom databases and phenotypes.

Key results

  1. ShownPICRUSt2 KEGG ortholog (KO) predictions from 16S data tracked paired shotgun metagenome (MGS) KO profiles closely and matched or beat PICRUSt1, Piphillin, PanFP and Tax4Fun2 in all seven validation datasets; the authors note PICRUSt2 stood out most in non-human-associated datasets.

    Across seven datasets (6 to 137 samples each), mean Spearman rho spanned 0.79 in primate stool (sd = 0.028) to 0.88 in Cameroonian stool (sd = 0.019) over the 6,220 KOs all databases could output; PICRUSt2 was best or comparable to the best tool in every dataset (paired two-tailed Wilcoxon tests, P < 0.05). · Abstract; Abstract; Figure legends

  2. ShownThe default reference grows from 2,011 to 41,926 bacterial and archaeal genomes from Integrated Microbial Genomes (IMG), broadening taxonomic and gene-family coverage; placement-based inference keeps almost all input ASVs, whereas PICRUSt1's reference-OTU rule dropped many study sequences.

    41,926 genomes (IMG as of November 8, 2017; >20-fold over PICRUSt1), dereplicated to 20,000 16S rRNA clusters; species and genus diversity up 5.3-fold and 2.2-fold, phyla from 39 to 64; KOs from 6,911 to 10,543 (1.5-fold); PICRUSt1 excluded many sequences in four representative 16S datasets. · Abstract; Figure legends

  3. ShownMost of PICRUSt2's agreement with metagenomes stems from cross-taxon and dataset-wide gene content rather than sample-specific signal: it beat a mean-genome null everywhere but exceeded ASV-shuffled predictions by only a small, significant margin.

    PICRUSt2 tables were closer to MGS than a null genome (mean gene family abundance over all reference genomes) in every dataset, yet ASV-shuffled predictions still reached mean rho of 0.77 for primate stool (sd = 0.196) to 0.84 for blueberry rhizosphere (sd = 0.178), vs 0.79 to 0.88 for unshuffled PICRUSt2. · Abstract

  4. ShownIn KO differential abundance (DA) tests on four datasets, PICRUSt2 had the highest F1 among prediction tools (contrary to an earlier report favoring Piphillin), yet all tools had low precision, and reprocessed metagenomes themselves reached only marginally higher precision.

    PICRUSt2 F1 0.46-0.59 (mean=0.51; sd=0.06); precision 0.38-0.58 (mean=0.48; sd=0.08) vs Piphillin 0.06-0.66 (mean=0.45; sd=0.27) and ASV-shuffled PICRUSt2 0.22-0.42 (mean=0.30; sd=0.09); alternative MGS pipeline (direct KEGG alignment instead of HUMAnN2) precision 0.57-0.67 (mean=0.62; sd=0.04); Wilcoxon tests, FDR < 0.05. · Abstract; Figure legends

  5. ShownPICRUSt2 also outputs MetaCyc pathway abundances from structured mappings of enzyme classification (EC) numbers, plus presence/absence calls for 41 IMG-linked phenotypes; both beat null baselines, though pathway differential abundance signal was largely dataset-wide.

    Pathway predictions (575 pathways shared with HUMAnN2) beat the null on all metrics (paired Wilcoxon P < 0.05); pathway DA F1 0.23-0.62 (mean=0.41; sd=0.17) vs 0.22-0.60 (mean=0.34; sd=0.18) for ASV-shuffled predictions. Genome hold-out phenotypes: mean F1 84.8% (sd 9.01%), precision 86.5% (sd 6.21%), recall 83.5% (sd 11.4%), above null (P < 0.05). · Abstract; Figure legends

Methods, data and reproducibility

ASVs are placed into a reference tree of 20,000 IMG 16S rRNA genes (HMMER, EPA-ng, GAPPA), and castor hidden state prediction infers per-ASV gene family copy numbers, which are multiplied by 16S copy-corrected abundances to give metagenomes and, via EC mappings, MetaCyc pathways. Outputs were benchmarked against paired shotgun metagenomes.

  • Cameroonian stool samples with paired 16S rRNA amplicon and shotgun metagenomic profiles (validation set) · human (stool microbiome) · n = 57
  • Indian stool samples with paired 16S rRNA amplicon and shotgun metagenomic profiles · human (stool microbiome) · n = 91
  • Human Microbiome Project (HMP) samples spanning multiple body sites, with paired 16S and shotgun data · human (Human Microbiome Project) · n = 137
  • Non-human primate stool samples with paired 16S and shotgun data · non-human primates (species not stated) · n = 77
  • Mammalian stool samples with paired 16S and shotgun data · mammals (species not stated) · n = 8
  • Ocean water samples with paired 16S and shotgun data · n = 6
  • Bulk soil and blueberry rhizosphere samples with paired 16S and shotgun data · n = 22
  • Default PICRUSt2 reference: IMG bacterial and archaeal genomes (as of November 8, 2017) annotated with KO, EC, COG, Pfam and TIGRFAM families plus 41 binary phenotypes · bacteria and archaea · n = 41,926 genomes; 20,000 dereplicated 16S rRNA clusters
Methods & tools
PICRUSt2 (phylogenetic placement plus hidden state prediction), PICRUSt1 (baseline), Piphillin (baseline), PanFP (baseline), Tax4Fun2 (baseline), HMMER (ASV placement), EPA-ng (optimal placement in reference phylogeny), GAPPA (outputs tree with placed ASVs), castor R package (hidden state prediction; replaces ape), HUMAnN2 (shotgun metagenome functional profiling), Integrated Microbial Genomes (IMG) database, KEGG orthologs, EC, COG, Pfam, TIGRFAM gene families; MetaCyc pathways
Software & environment
Analysis and database-construction code in Python and R; hidden state prediction via the castor R package; default genomes from IMG as of November 8, 2017; no software versions or containers are given.
Reproducibility gaps
Sequencing data accessions are deferred to the supplementary information and absent from the main text; no hardware, runtime, or package version details are reported.
Limitations (6)
  • Predictions are skewed toward existing reference genomes, so rare functions specific to particular environments tend to be missed; expanding genome collections and user-supplied custom genomes only partly mitigate this. · stated by the authors
  • Amplicon-based inference cannot resolve strain-specific functions within a species, because taxa are separable only as far as their amplified marker gene sequences differ. · stated by the authors
  • Spearman agreement may overstate accuracy because gene families co-occur within genomes; gains over a mean-genome null mostly reflect dataset-level genome content, and sample-level predictions only slightly beat ASV-shuffled ones. · stated by the authors
  • Differential abundance findings replicate poorly: every tool showed low KO precision, pathway F1 varied widely across datasets and tests, and even an alternative metagenome-processing workflow gave only marginally better precision. · stated by the authors
  • Phenotype accuracy comes from leave-one-genome-out tests within IMG, where many strains share identical 16S genes and close relatives may remain in the reference; accuracy for novel or environmental taxa may be lower and was not tested against measured phenotypes. · noted in this summary
  • Correlation benchmarks used only the 6,220 KOs every tool could output, so accuracy for newly added KOs is not separately reported, and the ocean (n=6) and mammalian stool (n=8) datasets offer little statistical power. · noted in this summary

Read this if you infer community function from 16S rRNA amplicon or ASV data and need evidence on how far PICRUSt2 KO, pathway and phenotype predictions can stand in for shotgun metagenomics.

Citations
812

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSun '19Wright '25Matchado '24Ward '17Cuesta‐Zulu '19Tovo '20
75% semantic matchBioinformatics
Shan Sun, Roshonda B. Jones, Anthony A. Fodor·bioRxiv·May 30, 2019· 8 citations