COMEBin

COMEBin bins metagenomic contigs into metagenome-assembled genomes (MAGs). It learns a representation of each contig from its sequence composition and read coverage with contrastive multi-view learning, and clusters the contigs with the Leiden algorithm.

COMEBin

License

Free to use and open source under GNU GPLv3.

Available

Roihu-CPU: 1.1.0, via the bio-apps module.
Roihu-GPU: 1.1.0 (GPU-accelerated training), via the bio-apps module.

Usage

COMEBin is part of the bio-apps collection on Roihu. Load the bio-apps module tree and then the COMEBin module:

module load bio-apps/v202603
module load comebin/1.1.0

COMEBin needs the contigs as an uncompressed FASTA file, and a directory of sorted BAM files from mapping the reads of each sample to the contigs. The gen_cov_file.sh script included with COMEBin maps paired-end reads with BWA and writes the BAM files to the directory given with -b:

gen_cov_file.sh -a contigs.fa -o cov -b bams -t 16 \
  -f _1.fastq.gz -r _2.fastq.gz reads/*.fastq.gz

-f and -r give the endings of the forward and reverse read files (the defaults are _1.fastq and _2.fastq). Give all options before the read files.

run_comebin.sh then runs the three steps of COMEBin: data augmentation, which computes the contig features from the contigs and BAM files, training of the model, and clustering. Only the training step can use a GPU. Run the mapping and all the COMEBin steps in batch jobs, not on the login node. See the example batch script below.

Running on Roihu-CPU

Run all three steps on Roihu-CPU with -d cpu, which trains the model on the CPU. The default, cuda, fails on CPU nodes:

run_comebin.sh -a contigs.fa -p bams -o comebin_out -d cpu -t $SLURM_CPUS_PER_TASK

The bins are written to comebin_out/comebin_res/comebin_res_bins, and the bin of each contig to comebin_out/comebin_res/comebin_res.tsv. If a run stops, run the same command again: COMEBin skips the data augmentation and training steps that have already finished.

To combine the bins of COMEBin with those of other binning tools, use MAGScoT.

Training on a GPU

Training is faster on a GH200 GPU, but the other steps only use CPUs. Run only the training as a GPU batch job, and the other steps as batch jobs on Roihu-CPU. You submit the GPU job from the Roihu-GPU login node (roihu-gpu.csc.fi), but note that the login node itself has no GPUs.

On Roihu-CPU, make the BAM files and run the data augmentation step in a batch job (see the example batch script):

python $COMEBIN_INSTROOT/bin/COMEBin/main.py generate_aug_data --contig_file contigs.fa \
  --bam_file_path bams --out_augdata_path comebin_out/data_augmentation --num_threads $SLURM_CPUS_PER_TASK

Log in to roihu-gpu.csc.fi, go to the same directory as in step 1, and submit the example GPU batch script, which trains the model.
On Roihu-CPU, run run_comebin.sh in a batch job with the same options as in Running on Roihu-CPU. It finds the finished data augmentation and training, and runs only the clustering.

Example batch script

This runs all three steps on Roihu-CPU. For GPU training, replace the run_comebin.sh line with the generate_aug_data command from step 1 above.

#!/bin/bash
#SBATCH --job-name=comebin
#SBATCH --account=<project>
#SBATCH --output=output_%j.txt
#SBATCH --error=errors_%j.txt
#SBATCH --partition=small
#SBATCH --time=24:00:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --mem=64G

module load bio-apps/v202603
module load comebin/1.1.0

gen_cov_file.sh -a contigs.fa -o cov -b bams -t $SLURM_CPUS_PER_TASK \
  -f _1.fastq.gz -r _2.fastq.gz reads/*.fastq.gz
run_comebin.sh -a contigs.fa -p bams -o comebin_out -d cpu -t $SLURM_CPUS_PER_TASK

Replace <project> with your CSC project (for example project_2001234).

Example GPU batch script

This runs only the training step, on the data augmentation results from Roihu-CPU. Submit it from roihu-gpu.csc.fi:

#!/bin/bash
#SBATCH --job-name=comebin-train
#SBATCH --account=<project>
#SBATCH --output=output_%j.txt
#SBATCH --error=errors_%j.txt
#SBATCH --partition=gpumedium
#SBATCH --time=08:00:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:gh200:1

module load bio-apps/v202603
module load comebin/1.1.0

python $COMEBIN_INSTROOT/bin/COMEBin/main.py train --data comebin_out/data_augmentation \
  --output_path comebin_out/comebin_res --earlystop --addvars --vars_sqrt \
  --num_threads $SLURM_CPUS_PER_TASK

See creating a batch job script for Roihu for more information about running batch jobs, and the GPU partitions for the available GPU resources.

Support

CSC Service Desk

More information

COMEBin GitHub repository