Antimicrobial resistance surveillance in the genomic age

Abstract

The loss of effective antimicrobials is reducing our ability to protect the global population from infectious disease. However, the field of antibiotic drug discovery and the public health monitoring of antimicrobial resistance (AMR) is beginning to exploit the power of genome and metagenome sequencing. The creation of novel AMR bioinformatics tools and databases and their continued development will advance our understanding of the molecular mechanisms and threat severity of antibiotic resistance, while simultaneously improving our ability to accurately predict and screen for antibiotic resistance genes within environmental, agricultural, and clinical settings. To do so, efforts must be focused toward exploiting the advancements of genome sequencing and information technology. Currently, AMR bioinformatics software and databases reflect different scopes and functions, each with its own strengths and weaknesses. A review of the available tools reveals common approaches and reference data but also reveals gaps in our curated data, models, algorithms, and data-sharing tools that must be addressed to conquer the limitations and areas of unmet need within the AMR research field before DNA sequencing can be fully exploited for AMR surveillance and improved clinical outcomes.

DeepARG: A deep learning approach for predicting antibiotic resistance genes from metagenomic data

“…Existing bioinformatics tools focus on detecting known ARG sequences in genomic or metagenomic sequence libraries and thus are biased towards specific ARGs (McArthur and Tsang, 2016). For instance, ResFinder (Moran et al, 2016) and SEAR (Rowe et al, 2015) predict specifically plasmid-borne ARGs, and Mykrobe predictor (Bradley et al, 2015) is dedicated to 12 types of antimicrobials, while PATRIC (Davis et al, 2016) is limited to identifying carbapenem, methicillin, and beta lactam ARGs.…”

“…Although the best hit approach has a low false positive rate, that is, few non-ARGs are predicted as ARGs (Forsberg et al, 2014), the false negative rate can be very high and a large number of actual ARGs end up being predicted as non-ARGs McArthur and Tsang, 2016). Figure 1 shows the distribution of manually curated potential ARGs from the Universal Protein Resource (UNIPROT) database against the Comprehensive Antibiotic Resistance Database (CARD) and the Antibiotic Resistance Genes Database (ARDB).…”

A preliminary investigation of major bacterial antibiotic resistance genes present in Australian fecal samples using shotgun metagenomics and targeted sequencing

“…The declining costs of metagenomic sequencing technologies has led to increased molecular investigation of ARGs and has enabled a shift from phenotype to genotype-based approaches to investigate AMR 21,22 . The ultimate goal of ARG sequencing and related bioinformatic analyses is the accurate detection of the resistome and prediction of the antibiogram from genomic and metagenomic data 23 . Several major databases provide collections of ARGs covering broad categories of AMR mechanisms including the Comprehensive Antibiotic Resistance Database (CARD) 24 , Antibiotic Resistance Gene-ANNOTation (ARG-ANNOT) 25 , the Bacterial Antimicrobial Resistance Reference Gene Database 26 hosted by the National Center for Biotechnology Information (NCBI) as part of the National Database of Antibiotic Resistant Organisms (NDARO), and ResFinder 27 (see 23 for a more exhaustive list of available AMR bioinformatics software, databases, and data-sharing resources).…”

Capreomycin resistance prediction in two species ofMycobacteriumusing a stacked ensemble method

“…These methods work well for finding known and highly conserved AR sequences and have low probability of producing false positives, that is, predicting non‐AR genes as AR genes (Forsberg et al ). However, they may produce high false‐negative rates when they fail to identify AR genes that have lower sequence identity with known AR genes (Xavier et al ; Yang et al ; McArthur and Tsang ). Machine learning can be a useful alternative computational framework for identifying AR phenotypes accurately using features, that is, characteristics of known AR protein sequences, found in the genomic data regardless of sequence similarity (Niehaus et al ; Santerre et al ).…”