Analytical Assessment of Metagenomic Workflows for Pathogen Detection with NIST RM 8376 and Two Sample Matrices

Abstract

We assessed the analytical performance of metagenomic workflows using NIST Reference Material 8376 DNA from bacterial pathogens spiked into two simulated clinical samples: cerebral spinal fluid (CSF) and stool. Sequencing and taxonomic classification were used to generate signals for each sample and taxa of interest, and to estimate the limit of detection (LOD), the response function, and linear dynamic range. We found that the LODs for taxa spiked into CSF ranged from approximately (0.1 to 0.3) copy/L, with a linearity of 0.96 to 0.99. For stool, the LODs ranged from (10 to 221) copy/L, with a linearity of 0.99 to 1.01. Further, discriminating different E. coli strains proved to be workflow-dependent, as only one classifier:database combination of the three tested showed the ability to differentiate the two pathogenic and commensal strains. Surprisingly, when we compared the response functions of the same taxa in the two different sample types, we found those functions to be the same, despite large differences in LODs. This suggests that the "agnostic diagnostic" theory for metagenomics may apply to different target organisms and different sample types. Using RMs, we were able to generate quantitative analytical performance metrics for each workflow and sample set, enabling relatively rapid workflow screening before employing clinical samples. This makes these RMs a useful tool that will generate data needed to support translation of metagenomics into regulated use. ImportanceAssessing the analytical performance of metagenomic workflows, especially when developing clinical diagnostics, is foundational for ensuring that the measurements underlying a diagnosis are supported by rigorous characterization. To facilitate the translation of metagenomics into clinical practice, workflows must be tested using control samples designed to probe the analytical limitations (e.g. limit of detection). Spike-ins allow developers to generate fit-for-purpose control samples for initial workflow assessments and inform decisions about further development. However, clinical sample types include a wide range of compositions and concentrations, each presenting different detection challenges. In this work, we demonstrate how spike-ins elucidate workflow performance in two highly dissimilar sample types (stool and CSF); and we provide evidence that detection of individual organisms is unaffected by background sample composition, making detection sample agnostic within a workflow. These demonstrations and performance insights will facilitate translation of the technology to the clinic.