A new self-supervised metagenome binning tool based on semi-contrastive learning method using the framework of BYOL (Bootstrap Your Own Latent) model. It only requires positive augmentated pairs for contrastive learning and at least 2x faster than COMEBin. It integrates k-mer sequence embedding vectors from GenomeFace composition model and sampled read coverage profiles using multinomial sampling of contigs to learn embedding space. It applies Leiden community detection algorithm to cluster contigs in the embedding space and obtain metagenomic bins.
conda env create -n mcdevol_env --file=environment.yml
conda activate mcdevol_env
export PATH=${PATH}/$(pwd)/mcdevol
python mcdevol.py --help
McDevol requires glibc2.25. Currently, McDevol was tested only on Linux system. On cluster node, if the default command doesn't work, try creating conda environment using CONDA_OVERRIDE_GLIBC=2.17 CONDA_OVERRIDE_CUDA=10.2 conda env create --file=environment.yml.
usage: mcdevol.py [-h] (-a ABUNDANCE | -i INPUTDIR) -c CONTIGS [-l MINLENGTH] [-o OUTDIR] [-n NCORES] [--abundformat ABUNDFORMAT] [-v] [-f NFRAGMENTS] [--multi_split]
McDevol: An accurate metagenome binning of contigs based on decovolution of abundance and k-mer embedding
optional arguments:
-h, --help show this help message and exit
-a ABUNDANCE, --abundance ABUNDANCE
abundance file in TSV format separated by tabs
-i INPUTDIR, --inputdir INPUTDIR
directory that contains SAM files
-c CONTIGS, --contigs CONTIGS
contigs fasta (or zip)
-l MINLENGTH, --minlength MINLENGTH
minimum length of contigs to be considered for binning
-o OUTDIR, --outdir OUTDIR
output directory
-n NCORES, --ncores NCORES
Number of cores to use
--abundformat ABUNDFORMAT
Format of abundance ('std|metabat', default='std') std:[contigname, s1meancov, s2meancov,...,sNmeancov]; metabat:[contigName, contigLen, totalAvgDepth,
s1meancov, s1varcov,...,sNmeancov, sNvarcov]
-v, --version print version and exit
-f NFRAGMENTS, --nfragments NFRAGMENTS
number of augumented fragments to generate
--multi_split split clusters based on sample id, separator (default='C')
python mcdevol.py -i <samfilesdir> -c <contigseq.fasta> -o <outputdir> --abundformat metabat -n 24
-i and -a options are mutually exclusive.
For multi-sample binning run
python mcdevol.py -i <samfilesdir> -c <contigseq.fasta> -o <outputdir> --abundformat metabat -n 24 --multi-split
McDevol is most powerful when input contigs have high read coverage profiles.