TaxAss: Leveraging a Custom Freshwater Database Achieves Fine-Scale Taxonomic Resolution

Abstract

Taxonomy assignment of freshwater microbial communities is limited by the minimally curated phylogenies used for large taxonomy databases. Here we introduce TaxAss, a taxonomy assignment workflow that classifies 16S rRNA gene amplicon data using two taxonomy reference databases: a large comprehensive database and a small ecosystem-specific database rigorously curated by scientists within a field. We applied TaxAss to five different freshwater data sets using the comprehensive SILVA database and the freshwater-specific FreshTrain database. TaxAss increased the percentage of the data set classified compared to using only SILVA, especially at fine-resolution family to species taxon levels, while across the freshwater test data sets classifications increased by as much as 11 to 40% of total reads. A similar increase in classifications was not observed in a control mouse gut data set, which was not expected to contain freshwater bacteria. TaxAss also maintained taxonomic richness compared to using only the FreshTrain across all taxon levels from phylum to species. Without TaxAss, most organisms not represented in the FreshTrain were unclassified, but at fine taxon levels, incorrect classifications became significant. We validated TaxAss using simulated amplicon data derived from full-length clone libraries and found that 96 to 99% of test sequences were correctly classified at fine resolution. TaxAss splits a data set's sequences into two groups based on their percent identity to reference sequences in the ecosystem-specific database. Sequences with high similarity to sequences in the ecosystem-specific database are classified using that database, and the others are classified using the comprehensive database. TaxAss is free and open source and is available at https://www.github.com/McMahonLab/TaxAssIMPORTANCE Microbial communities drive ecosystem processes, but microbial community composition analyses using 16S rRNA gene amplicon data sets are limited by the lack of fine-resolution taxonomy classifications. Coarse taxonomic groupings at the phylum, class, and order levels lump ecologically distinct organisms together. To avoid this, many researchers define operational taxonomic units (OTUs) based on clustered sequences, sequence variants, or unique sequences. These fine-resolution groupings are more ecologically relevant, but OTU definitions are data set dependent and cannot be compared between data sets. Microbial ecologists studying freshwater have curated a small, ecosystem-specific taxonomy database to provide consistent and up-to-date terminology. We created TaxAss, a workflow that leverages this database to assign taxonomy. We found that TaxAss improves fine-resolution taxonomic classifications (family, genus, and species). Fine taxonomic groupings are more ecologically relevant, so they provide an alternative to OTU-based analyses that is consistent and comparable between data sets.

Keywords: 16S rRNA gene; amplicon sequencing; freshwater; limnology; microbial ecology; taxonomy; taxonomy assignment; taxonomy database.

FIG 1

TaxAss conceptual diagram. TaxAss separates OTUs into two groups that are classified separately and then recombined. OTUs similar to any ecosystem-specific reference sequences are classified using the ecosystem-specific database; otherwise, they are classified by the comprehensive database. BLAST is used to split the OTUs into groups (left arrows), and the Wang classifier is used to assign taxonomy (right arrows).

FIG 2

TaxAss validation with tags simulated from full-length Marathonas Reservoir clone libraries. Tags simulated by trimming full-length sequences to the V4 region were classified by TaxAss, and the resulting classifications were compared to “reference” classifications determined by manually aligning the full-length sequences to the FreshTrain. Correct classifications are in green, lost ecosystem-specific classifications are in yellow, and incorrect classifications are in red. (Left) Number of unique sequences in each classification category at fine-resolution taxon levels. (Right) Examples of classifications that fit into each classification category. Tabular results from this and additional amplicon region simulations are available in Table S1.

FIG 3

TaxAss performance compared to SILVA-only performance. (A and B) The left bars represent the SILVA-only classification, and the right bars represent the TaxAss classification that leveraged both SILVA and the FreshTrain. Within the right bars, red reads were classified by the FreshTrain using TaxAss and were unclassified using only SILVA; yellow reads were classified by the FreshTrain using TaxAss but received SILVA classifications using only SILVA, and gray reads remained classified by SILVA when using TaxAss. (A) In the Lake Mendota data set, TaxAss leveraged the FreshTrain and SILVA to achieve improved fine-resolution classifications. (B) TaxAss achieved improvements in a range of freshwater data sets despite the FreshTrain’s primary focus on temperate lake epilimnia. Few changes in classification were observed in the mouse gut control. Versions of this figure across all data sets and taxa levels can be found in Fig. S1.

FIG 4

TaxAss performance compared to FreshTrain-only performance. (A and B) Lake Mendota reads represented by blue bars were incorrectly classified as red bars in the FreshTrain-only classification. Rank order of the bars follows the TaxAss classification rank abundances. Only taxa with at least 0.5% relative abundance are included, and at the lineage level, the number of bars displayed is further truncated to 20. (A) TaxAss maintained phylum richness (blue bars) by classifying phyla using SILVA when they are not included in FreshTrain. (B) TaxAss prevented lineage-level inaccuracies from misclassifications and overclassifications (red bars over known taxa) and lineage-level underclassifications (red bar over “unclassified” category). Versions of this figure across all test ecosystems can be found in Fig. S2.

FIG 5

Percent identity where classifications are maximized. The percentages of reads classified when using different percent identity cutoffs to separate out ecosystem-specific OTUs are shown for each freshwater data set across taxon levels. Faint vertical lines highlight the 98% identity chosen for the analyses in this paper. OTUs are predominantly unclassified at fine resolution if they are placed in the wrong classification group, so this visualization is generated by TaxAss to help users choose a percent identity cutoff appropriate for their data set.