DeepARG

DeepARG predicts antibiotic resistance genes (ARGs) in metagenomic sequences using deep learning to improve detection and annotation of ARGs.

Key Features:

• Deep learning framework: Employs deep learning models trained on ARG sequence dissimilarities for classification.
• Dissimilarity matrix: Leverages a dissimilarity matrix constructed from all known categories of ARGs.
• DeepARG-SS: Implements a model optimized for short read sequences.
• DeepARG-LS: Implements a model optimized for full gene-length sequences.
• Performance metrics: Evaluated across 30 antibiotic resistance categories with reported precision >0.97 and recall >0.90.
• Reduced false negatives: Outperforms best-hit sequence similarity approaches by reducing false negatives and improving recall.
• Threshold flexibility: Predicts ARGs without relying on strict cutoffs to capture broader ARG diversity.
• DeepARG-DB: Provides a database of high-confidence predicted ARGs that were subjected to extensive manual inspection.

Scientific Applications:

• ARG detection and annotation: Identification and annotation of antibiotic resistance genes in metagenomic datasets.
• Environmental surveillance: Monitoring ARGs in environmental reservoirs such as wastewater, agricultural waste, food, and water.
• ARG diversity and dissemination studies: Expanding ARG catalogs and capturing under-represented ARG categories for studies of resistance spread.
• Method comparison: Benchmarking and comparison against traditional best-hit sequence similarity methods to assess false-negative rates and recall.

Methodology:

Uses deep learning models trained on a dissimilarity matrix of known ARG categories, comprising DeepARG-SS for short reads and DeepARG-LS for full-length genes; DeepARG-DB was constructed from high-confidence predicted ARGs with manual inspection, and models were evaluated across 30 ARG categories with precision >0.97 and recall >0.90.