REMAG

REMAG recovers high-quality eukaryotic metagenome-assembled genomes from mixed prokaryotic-eukaryotic metagenomic samples using non-contrastive self-supervised learning.

Key Features:

• Eukaryotic contig filtering: Uses a finetuned HyenaDNA genomic foundation model to efficiently filter eukaryotic contigs from mixed community data.
• Dual-encoder embedding learning: Employs a Siamese network trained with the non-contrastive Barlow Twins loss to learn a shared embedding space integrating contig composition and differential coverage.
• Greedy iterative binning: Extracts high-quality bins using greedy iterative Leiden clustering optimized with eukaryotic single-copy core gene constraints.
• Long-read data optimization: Specifically suited for long-read metagenomic data and scales effectively with increasing dataset size and sequencing depth.

Scientific Applications:

• Eukaryotic genome recovery from mixed communities: Recovers near-complete eukaryotic genomes from simulated and real mixed prokaryotic/eukaryotic metagenomic communities of varying sizes and origin.
• Protist and fungal community analysis: Enables exploration of protist and fungal populations in diverse ecosystems where these eukaryotes are ecologically relevant but genome recovery has lagged behind prokaryotes.

Methodology:

REMAG uses a finetuned HyenaDNA genomic foundation model for eukaryotic contig filtering, then employs a dual-encoder Siamese network trained with the non-contrastive Barlow Twins loss to integrate contig composition and differential coverage in a shared embedding space, and finally extracts bins using greedy iterative Leiden clustering with eukaryotic single-copy core gene constraints.