Integrating central nervous system metagenomics and host response for diagnosis of tuberculosis meningitis and its mimics

Tuberculous meningitis is difficult to differentiate from meningitis caused by other pathogens. Here, the authors combine metagenomics-based pathogen detection in cerebrospinal fluid with a host gene expression-based machine learning classifier for diagnosis. The epidemiology of infectious causes of meningitis in sub-Saharan Africa is not well understood, and a common cause of meningitis in this region, Mycobacterium tuberculosis (TB), is notoriously hard to diagnose. Here we show that integrating cerebrospinal fluid (CSF) metagenomic next-generation sequencing (mNGS) with a host gene expression-based machine learning classifier (MLC) enhances diagnostic accuracy for TB meningitis (TBM) and its mimics. 368 HIV-infected Ugandan adults with subacute meningitis were prospectively enrolled. Total RNA and DNA CSF mNGS libraries were sequenced to identify meningitis pathogens. In parallel, a CSF host transcriptomic MLC to distinguish between TBM and other infections was trained and then evaluated in a blinded fashion on an independent dataset. mNGS identifies an array of infectious TBM mimics (and co-infections), including emerging, treatable, and vaccine-preventable pathogens including Wesselsbron virus, Toxoplasma gondii, Streptococcus pneumoniae, Nocardia brasiliensis, measles virus and cytomegalovirus. By leveraging the specificity of mNGS and the sensitivity of an MLC created from CSF host transcriptomes, the combined assay has high sensitivity (88.9%) and specificity (86.7%) for the detection of TBM and its many mimics. Furthermore, we achieve comparable combined assay performance at sequencing depths more amenable to performing diagnostic mNGS in low resource settings.

Automated summary

In this study, the authors combined metagenomics-based pathogen detection with a machine learning classifier based on host gene expression to enhance the diagnosis of tuberculous meningitis (TBM). By integrating these approaches, they achieved high sensitivity and specificity for detecting TBM and its mimics, even in low resource settings.