ResMiCo

ResMiCo identifies misassembled contigs in metagenome assemblies using a residual neural network to provide reference-free quality assessment of metagenome-assembled genomes.

Key Features:

• Residual neural network: Implements a residual neural network architecture tailored to detect misassembled contigs.
• Reference-free prediction: Operates without relying on reference genomes to identify misassemblies in assembled contigs.
• Robustness to taxonomic novelty: Demonstrates the ability to generalize to diverse and taxonomically novel genomic data and assembly methods.
• Benchmarking and accuracy: Shows substantial performance improvements over prior methods, which reported AUPRC not exceeding 0.57.
• Large-scale training dataset: Trained on a dataset of unprecedented size and complexity to support rigorous evaluation and model generalization.
• Assembly parameter optimization: Enables optimization of metagenome assembly hyperparameters with a focus on reducing misassemblies as well as improving contiguity.

Scientific Applications:

• Quality control: Applied to metagenomic studies to identify misassembled contigs and improve the fidelity of metagenome-assembled genomes.
• Methodology optimization: Used to inform adjustment of assembly parameters and workflows to reduce misassembly rates.
• Misassembly rate estimation: Applied to real-world datasets to estimate misassembly prevalence, reporting an average of approximately 7% misassembled contigs per metagenome in evaluations.

Methodology:

ResMiCo uses a residual neural network trained on a large, complex training dataset and evaluated via benchmarking metrics including AUPRC.