DeepMM: Identify and correct Metagenome Misassemblies with deep learning

Abstract

Accurate metagenomic assemblies are essential for constructing reliable metagenome-assembled genomes (MAGs). However, the complexity of microbial genomes continues to pose challenges for accurate assembly. Current reference-free assembly evaluation tools primarily rely on hand-crafted features and suffer from poor generalization across different metagenomic data. To address these limitations, we propose DeepMM, a novel deep learning-based visual model de-signed for the identification and correction of metagenomic misassemblies. DeepMM transforms alignments between assemblies and reads into a multi-channel image for misassembly feature learning and applies contrastive learning to bring different views of misassemblies closer. Fur-thermore, DeepMM offers a fine-tuning process to match different sequencer data. Our results show that DeepMM outperforms state-of-the-art methods in identifying misassemblies, achieving the highest AUPRC score in five CAMI datasets. DeepMM provides accurate correction of misassemblies, significantly improving downstream binning results, increasing the number of near-complete MAGs from 905 to 1006 in a large real metagenomic sequencing dataset derived from a diarrhea-predominant Irritable Bowel Syndrome (IBS-D) cohort.