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

Area of Science:

• Environmental science
• Bioinformatics
• Genomics

Background:

• Rising antibiotic resistance necessitates enhanced global monitoring of environmental sources.
• Environmental media (wastewater, food, water) are key reservoirs for antibiotic resistance genes (ARGs) and pathways for human exposure.
• Current methods using next-generation sequencing and database lookups yield high false-negative rates for ARG identification.

Purpose of the Study:

• To develop and evaluate deep learning models for more accurate identification of antibiotic resistance genes (ARGs).
• To address limitations of existing bioinformatics approaches in detecting ARGs, particularly reducing false negatives.

Main Methods:

• A deep learning approach was developed, utilizing a dissimilarity matrix of known ARGs.
• Two models, DeepARG-SS and DeepARG-LS, were created for short and long sequence reads, respectively.
• The models were evaluated against 30 antibiotic resistance categories.

Main Results:

• DeepARG models achieved high precision (>0.97) and recall (>0.90) across 30 ARG categories.
• The models outperformed the traditional 'best hit' method by significantly reducing false negatives and increasing overall recall (>0.9).
• A new curated database, DeepARG-DB, was created, expanding known ARG repositories.

Conclusions:

• Deep learning models provide more accurate antimicrobial resistance annotation than current bioinformatics practices.
• DeepARG facilitates the identification of a wider diversity of ARGs without strict cutoffs.
• The DeepARG models and database are publicly available as a command-line tool and web service.