To the Editor—One limitation of microbial community marker-gene sequencing is that it does not provide information about the functional composition of sampled communities. PICRUSt1 was developed in 2013 to predict the functional potential of a bacterial community on the basis of marker gene sequencing profiles, and now we present PICRUSt2 (https://github. com/picrust/picrust2), which improves on the original method. Specifically, PICRUSt2 contains an updated and larger database of gene families and reference genomes, provides interoperability with any operational taxonomic unit (OTU)-picking or denoising algorithm, and enables phenotype predictions. Benchmarking shows that PICRUSt2 is more accurate than PICRUSt and other competing methods overall. PICRUSt2 also allows the addition of custom reference databases. We highlight these improvements and also important caveats regarding the use of predicted metagenomes. The most common method for profiling bacterial communities is to sequence the conserved 16S rRNA gene. Functional profiles cannot be directly identified using 16S rRNA gene sequence data owing to strain variation, so several methods have been developed to predict microbial community functions from taxonomic profiles (amplicon sequences) alone1–5. Shotgun metagenomics sequencing (MGS), which sequences entire genomes rather than marker genes, can also be used to characterize the functions of a community, but does not work well if there is host contamination—for example, in a biopsy—or if there is very little community biomass. PICRUSt (hereafter “PICRUSt1”) was developed for prediction of functions from 16S marker sequences, and it is widely used but has some limitations. Standard PICRUSt1 workflows require input sequences to be OTUs generated from closed-reference OTU-picking against a compatible version of the Greengenes database6. Due to this restriction to reference OTUs, the default PICRUSt1 workflow is incompatible with sequence denoising methods, which produce amplicon sequence variants (ASVs) rather than OTUs. ASVs have finer resolution, allowing closely related organisms to be more readily distinguished. Furthermore, the bacterial reference databases used by PICRUSt1 have not been updated since 2013 and lack thousands of recently added gene families. We expected that optimizing genome prediction would improve accuracy of functional predictions. Therefore, the PICRUSt2 algorithm (Fig. 1a) includes steps that optimize genome prediction, including placing sequences into a reference phylogeny rather than relying on predictions limited to reference OTUs (Fig. 1b); basing predictions on a larger database of reference genomes and gene families (Fig. 1c); more stringently predicting pathway abundance (Supplementary Fig. 1); and enabling predictions of complex phenotypes and integration of custom databases. PICRUSt2 integrates existing open-source tools to predict genomes of environmentally sampled 16S rRNA gene sequences. ASVs are placed into a reference tree, which is used as the basis of functional predictions. This reference tree contains 20,000 full 16S rRNA genes from bacterial and archaeal genomes in the Integrated Microbial Genomes (IMG) database7. Phylogenetic placement in PICRUSt2 is based on the output of three