Evaluating Novel Metagenomic Classification Algorithms for Forensic Microbial Detection

As genome sequencing technologies become more rapid, robust, and affordable, there is increasing interest in its application to microbial forensics investigations. One goal of microbial forensics is to detect specific microorganisms with a high degree of certainty, which is more challenging than techniques that compare the abundance of different organisms across many environmental samples. There are no published algorithms for metagenomic classification that present their results with associated probabilities which would be forensically useful. In this study we tested the current state-of-the-art collection of metagenomic classifiers, including our own novel algorithm (“SIANN”) which incorporates near neighbors and non-parametric likelihood estimation, and evaluates the specificity and sensitivity of detection using real-world Illumina next-generation sequencing (NGS) data from bacterial pathogens. We find that the utility of binary classification varies widely with abundance and significance thresholds across all methods evaluated here. Moreover, the presence of near neighbor organisms can be highly confounding to the detection of human bacterial pathogens. We present these findings in order to establish thresholds for metagenomic classification with associated probabilities that can be used to make NGS data interpretable to the microbial forensic community.