It is possible to detect bacterial species in shotgun metagenome datasets through the presence of only a few sequence reads. However, false positive results can arise, as was the case in the initial findings of a recent New York City subway metagenome project. False positives are especially likely when two closely related are present in the same sample. Bacillus anthracis, the etiologic agent of anthrax, is a high-consequence pathogen that shares >99% average nucleotide identity with Bacillus cereus group (BCerG) genomes. Our goal was to create an analysis tool that used k-mers to detect B. anthracis, incorporating information about the coverage of BCerG in the metagenome sample.

Using public complete genome sequence datasets, we identified a set of 31-mer signatures that differentiated B. anthracis from other members of the B. cereus group (BCerG), and another set which differentiated BCerG genomes (including B. anthracis) from other Bacillus strains. We also created a set of 31-mers for detecting the lethal factor gene, the key genetic diagnostic of the presence of anthrax-causing bacteria. We created synthetic sequence datasets based on existing genomes to test the accuracy of a k-mer based detection model.

We found 239,503 B. anthracis-specific 31-mers (the Ba31 set), 10,183 BCerG 31-mers (the BCerG31 set), and 2,617 lethal factor k-mers (the lef31 set). We showed that false positive B. anthracis k-mers—which arise from random sequencing errors—are observable at high genome coverages of B. cereus. We also showed that there is a “gray zone” below 0.184× coverage of the B. anthracis genome sequence, in which we cannot expect with high probability to identify lethal factor k-mers. We created a linear regression model to differentiate the presence of B. anthracis-like chromosomes from sequencing errors given the BCerG background coverage. We showed that while shotgun datasets from the New York City subway metagenome project had no matches to lef31 k-mers and hence were negative for B. anthracis, some samples showed evidence of strains very closely related to the pathogen.

This work shows how extensive libraries of complete genomes can be used to create organism-specific signatures to help interpret metagenomes. We contrast “specialist” approaches to metagenome analysis such as this work to “generalist” software that seeks to classify all organisms present in the sample and note the more general utility of a k-mer filter approach when taxonomic boundaries lack clarity or high levels of precision are required.