DeepMAsED: evaluating the quality of metagenomic assemblies

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Metagenome assembly advances generate numerous assemblies, but identifying misassemblies is difficult without reference genomes.
• Existing reference-free misassembly detection methods are outdated, rely on strong assumptions, and lack validation on large datasets.

Purpose of the Study:

• To develop a novel deep learning approach, DeepMAsED, for accurate misassembly identification in metagenomic contigs.
• To create an in silico pipeline for generating realistic, large-scale metagenome assemblies for model training and testing.

Main Methods:

• Implemented a deep learning model (DeepMAsED) for reference-free misassembly detection.
• Developed a computational pipeline for generating simulated metagenome assemblies.

Main Results:

• DeepMAsED significantly outperforms existing methods on large and complex metagenome assemblies.
• The model estimates a 1% misassembly rate in recent large-scale metagenome assembly publications.
• DeepMAsED accurately identifies misassemblies across diverse bacteria and archaea.

Conclusions:

• DeepMAsED provides a flexible and accurate solution for metagenome misassembly detection without requiring reference genomes.
• The tool and its associated dataset generation pipeline are straightforward to use and retrain.
• DeepMAsED is applicable to a wide array of metagenome assembly projects, enhancing data quality.