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

Abstract

Growing concerns regarding increasing rates of antibiotic resistance call for global monitoring efforts. Monitoring of environmental media (e.g., wastewater, agricultural waste, food, and water) is of particular interest as these media can serve as sources of potential novel antibiotic resistance genes (ARGs), as hot spots for ARG exchange, and as pathways for the spread of ARGs and human exposure. Next-generation sequencebased monitoring has recently enabled direct access and profiling of the total metagenomic DNA pool, where ARGs are identified or predicted based on the "best hits" of homology searches against existing databases. Unfortunately, this approach tends to produce high rates of false negatives. To address such limitations, we propose here a deep leaning approach, taking into account a dissimilarity matrix created using all known categories of ARGs. Two models, deepARG-SS and deepARG-LS, were constructed for short read sequences and full gene length sequences, respectively.Performance evaluation of the deep learning models over 30 classes of antibiotics demonstrates that the deepARG models can predict ARGs with both high precision (>0.97) and recall (>0.90) for most of the antibiotic resistance categories. The models show advantage over the traditional best hit approach by having consistently much lower false negative rates and thus higher overall recall (>0.9). As more data become available for under-represented antibiotic resistance categories, the deepARG models' performance can be expected to be further enhanced due to the nature of the . CC-BY-NC 4.0 International license peer-reviewed) is the author/funder. It is made available under a The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/149328 doi: bioRxiv preprint first posted online Jun. 12, 2017; underlying neural networks. The deepARG models are available both in command line version and via a Web server at http://bench.cs.vt.edu/deeparg. Our newly developed ARG database, deepARG-DB, containing predicted ARGs with high confidence and high degree of manual curation, greatly expands the current ARG repository.DeepARG-DB can be downloaded freely to benefit community research and future development of antibiotic resistance-related resources.

Insights into the antibiotic resistance dissemination in a wastewater effluent microbiome: bacteria, viruses and vesicles matter

“…4B). In the case of ARGs sub-categories or groups, we found 412 out of a total of 2149 described groups (Arango-Argoty et al, 2018). The most abundant ARG category in the control and antibiotic-treated samples (50%) was multidrug efflux pumps (e.g.…”

“…The comparison between ARG and PRG genes normalized by metagenome size between control and treated samples for all the fractions was performed with one-way ANOVA, as indicated above. Unassembled metagenomic data were also analysed for ARG presence following the recently described machine learning algorithm developed for metagenomic data (Arango-Argoty et al, 2018). For this, cleaned filtered metagenomes were analysed and submitted to the comprehensive DeepARG platform that contains 14 933 reference ARG sequences.…”

Self-Attention based model forde-novoantibiotic resistant gene classification with enhanced reliability for Out of Distribution data detection

“…Dataset We used a modified version of the dataset curated in the DeepARG study [9]. Briefly, The dataset was created from the CARD [22], ARDB [23] and UNIPROT [24] databases with a combination of computational and manual curation.…”

“…This property can be useful in identifying metagenomic sample resistance for the purpose of providing a focused drug treatment. Traditional methods [6,7,8] to identify antibiotic-resistant genes usually take an alignment based best-hit approach which causes the methods to produce many false negatives [9]. Recently, a deep learning based approach was developed that used normalized bit scores as features that were acquired after aligning against known antibiotic resistant genes [9].…”

“…Traditional methods [6,7,8] to identify antibiotic-resistant genes usually take an alignment based best-hit approach which causes the methods to produce many false negatives [9]. Recently, a deep learning based approach was developed that used normalized bit scores as features that were acquired after aligning against known antibiotic resistant genes [9]. In contrast, our model only uses the raw protein sequence as its input, and as such doesn't need do any kind of alignment against a set database.…”