ROCker models for reliable detection and typing of short-read sequences carrying mcr , erm , mph , and lnu antibiotic resistance genes

Quantitative monitoring of emerging antimicrobial resistance genes (ARGs) using short-read sequences remains challenging due to the high frequency of amino acid functional domains and motifs shared with related but functionally distinct (non-target) proteins. To facilitate ARG monitoring efforts using unassembled short-reads, we present novel ROCker models for mcr , mph , erm , and lnu ARG families as well as models for variants of special public health concern within these families including mcr-1 , mphA , ermB , lnuF, lnuB, and lnuG genes. For this, we curated target gene sequence sets for model training and built these models using the recently updated ROCker V2 pipeline (Gerhardt et al., in review). To validate our models, we simulated reads from the whole genome of ARG-carrying isolates spanning a range of common read lengths and used them to challenge the filtering efficacy of ROCker vs. common static filtering approaches such as similarity searches using BLASTx with various e-value thresholds or hidden Markov models. ROCker models consistently showed F1 scores up to 10x higher (31% higher on average) and lower false-positive (by 30%, on average) and false-negative (by 16%, on average) rates based on 250 bp-long reads compared to alternative methods. The ROCker models and all related reference material and data are freely available through http://enve-omics.ce.gatech.edu/rocker/models, further expanding the available model collection developed previously for other genes. Their application to short-read metagenomes, metatranscriptomes, and PCR amplicon data should facilitate more accurate classification and quantification of unassembled short-read sequences for these ARG families and specific genes.