VAMB

VAMB applies deep variational autoencoders to integrate contig sequence composition (k-mer distributions) and co-abundance profiles from BAM files for metagenomic binning and microbial genome reconstruction.

Key Features:

• Integration of Sequence Information: Combines sequence composition from a contig catalogue with co-abundance data sourced from BAM files.
• Deep Variational Autoencoders: Encodes k-mer distributions and sequence co-abundance jointly using deep variational autoencoders to learn representative features without prior dataset-specific knowledge.
• Clustering in Latent Space: Clusters the latent representations produced by the autoencoder to generate metagenomic bins.

Scientific Applications:

• Enhanced Genome Reconstruction: Reconstructs substantially more near-complete (NC) genomes, reporting 29–98% more NC genomes on simulated datasets and a 45% increase on real data versus prior binners.
• Strain Separation: Separates closely related strains up to an average nucleotide identity (ANI) of 99.5%, exemplified by differentiation of sample-specific Bacteroides vulgatus and Bacteroides dorei in 1,000 human gut microbiome samples.
• Geographical Distribution Analysis: Enables analysis of species geographical distribution patterns using 2,606 NC bins from the human gut microbiome dataset.

Methodology:

VAMB encodes k-mer distributions and sequence co-abundance derived from a contig catalogue and BAM files using deep variational autoencoders, then clusters the resulting latent representations for binning.