Construction of habitat-specific training sets to achieve species-level assignment in 16S rRNA gene datasets

Background: The low cost of 16S rRNA gene sequencing facilitates population-scale molecular epidemiological studies. Existing computational algorithms can resolve 16S rRNA gene sequences into high-resolution amplicon sequence variants (ASVs), which represent consistent labels comparable across studies. Assigning these ASVs to species-level taxonomy strengthens the ecological and/or clinical relevance of 16S rRNA gene-based microbiota studies and further facilitates data comparison across studies.

Results: To achieve this, we developed a broadly applicable method for constructing high-resolution training sets based on the phylogenic relationships among microbes found in a habitat of interest. When used with the naïve Bayesian Ribosomal Database Project (RDP) Classifier, this training set achieved species/supraspecies-level taxonomic assignment of 16S rRNA gene-derived ASVs. The key steps for generating such a training set are (1) constructing an accurate and comprehensive phylogenetic-based, habitat-specific database; (2) compiling multiple 16S rRNA gene sequences to represent the natural sequence variability of each taxon in the database; (3) trimming the training set to match the sequenced regions, if necessary; and (4) placing species sharing closely related sequences into a training-set-specific supraspecies taxonomic level to preserve subgenus-level resolution. As proof of principle, we developed a V1-V3 region training set for the bacterial microbiota of the human aerodigestive tract using the full-length 16S rRNA gene reference sequences compiled in our expanded Human Oral Microbiome Database (eHOMD). We also overcame technical limitations to successfully use Illumina sequences for the 16S rRNA gene V1-V3 region, the most informative segment for classifying bacteria native to the human aerodigestive tract. Finally, we generated a full-length eHOMD 16S rRNA gene training set, which we used in conjunction with an independent PacBio single molecule, real-time (SMRT)-sequenced sinonasal dataset to validate the representation of species in our training set. This also established the effectiveness of a full-length training set for assigning taxonomy of long-read 16S rRNA gene datasets.

Conclusion: Here, we present a systematic approach for constructing a phylogeny-based, high-resolution, habitat-specific training set that permits species/supraspecies-level taxonomic assignment to short- and long-read 16S rRNA gene-derived ASVs. This advancement enhances the ecological and/or clinical relevance of 16S rRNA gene-based microbiota studies. Video Abstract.