REMAG

REcovery of eukaryotic genomes using contrastive learning. A specialized metagenomic binning tool designed for recovering high-quality eukaryotic genomes from mixed prokaryotic-eukaryotic samples.

Quick Start

Option 1: Using Conda (Recommended - handles all dependencies)

# Create environment and install everything
conda create -n remag -c bioconda -c conda-forge remag
conda activate remag

# Run REMAG (output directory optional - defaults to remag_output)
remag contigs.fasta -c alignments.bam

Option 2: Using Docker (No local installation needed)

docker run --rm -v $(pwd):/data danielzmbp/remag:latest \
  /data/contigs.fasta -c /data/alignments.bam -o /data/output

Option 3: Using pip

# Create environment first
conda create -n remag python=3.9
conda activate remag

# Install dependencies and REMAG
conda install -c bioconda miniprot
pip install remag

# Run REMAG
remag contigs.fasta -c alignments.bam

Installation

Recommended: Conda Installation

This is the easiest method as conda handles all dependencies automatically:

# Create a new environment with all dependencies
conda create -n remag -c bioconda -c conda-forge remag
conda activate remag

# Verify installation
remag --help

Note: miniprot is pulled in automatically as a dependency of the conda package; no separate installation is required when installing remag via conda.

Alternative: PyPI Installation

If you prefer pip, you'll need to install the external dependency separately:

# Step 1: Create and activate environment
conda create -n remag python=3.9
conda activate remag

# Step 2: Install external dependency
conda install -c bioconda miniprot

# Step 3: Install REMAG from PyPI
pip install remag

Advanced Conda Setup

For additional features:

# Basic installation
conda create -n remag -c bioconda -c conda-forge remag
conda activate remag

# Add optional plotting capabilities
conda install -c conda-forge matplotlib umap-learn

Using Docker

# Pull and run the latest version (output directory defaults to remag_output)
docker run --rm -v $(pwd):/data danielzmbp/remag:latest \
  /data/contigs.fasta -c /data/alignments.bam

# Or specify output directory
docker run --rm -v $(pwd):/data danielzmbp/remag:latest \
  /data/contigs.fasta -c /data/alignments.bam -o /data/output

# For interactive use
docker run -it --rm -v $(pwd):/data danielzmbp/remag:latest /bin/bash

Using Singularity

# Pull and run the latest version directly
singularity run docker://danielzmbp/remag:latest \
  contigs.fasta -c alignments.bam

# Build Singularity image from Docker Hub
singularity build remag_v0.3.4.sif docker://danielzmbp/remag:v0.3.4

# Or build latest version
singularity build remag_latest.sif docker://danielzmbp/remag:latest

# Run with Singularity
singularity run --bind $(pwd):/data remag_v0.3.4.sif \
  /data/contigs.fasta -c /data/alignments.bam

# Or use exec for direct command execution
singularity exec --bind $(pwd):/data remag_v0.3.4.sif \
  remag /data/contigs.fasta -c /data/alignments.bam -o /data/output

# For interactive shell
singularity shell --bind $(pwd):/data remag_v0.3.4.sif

# Build a local Singularity image file (optional)
singularity build remag.sif docker://danielzmbp/remag:latest
singularity run remag.sif contigs.fasta -c alignments.bam

From source

# Create and activate conda environment
conda create -n remag python=3.9
conda activate remag

# Clone and install
git clone https://github.com/danielzmbp/remag.git
cd remag
pip install .

Development installation

For contributors and developers:

# Install with development dependencies
pip install -e ".[dev]"

Optional Features Installation

For visualization capabilities:

# Install with plotting dependencies
pip install "remag[plotting]"

Usage

Command line interface

After installation, you can use REMAG via the command line:

# Basic usage (output defaults to remag_output in FASTA directory)
remag contigs.fasta -c alignments.bam

# With explicit output directory
remag contigs.fasta -c alignments.bam -o output_directory

# Multiple samples using glob patterns
remag contigs.fasta -c "samples/*.bam"

# Using explicit -f flag (both styles work)
remag -f contigs.fasta -c alignments.bam

# Keep intermediate files with -k shorthand
remag contigs.fasta -c alignments.bam -k

Python module mode

python -m remag contigs.fasta -c alignments.bam

Getting help

# Quick reference (basic options)
remag -h

# Full documentation (all advanced options)
remag --help

How REMAG Works

REMAG uses a sophisticated multi-stage pipeline specifically designed for eukaryotic genome recovery:

1. Eukaryotic Filtering: By default, REMAG automatically filters for eukaryotic contigs using the integrated HyenaDNA LLM-based classifier (can be disabled with --skip-bacterial-filter)
2. Feature Extraction: Combines k-mer composition (4-mers) with coverage profiles across multiple samples. Large contigs are split into overlapping fragments for augmentation during training
3. Contrastive Learning: Trains a Siamese neural network using the Barlow Twins self-supervised loss function. This creates embeddings where fragments from the same contig are close together
4. Adaptive Resolution: Automatically determines optimal Leiden clustering resolution by testing multiple resolutions and selecting the one that maximizes individual bin completeness
5. Clustering: Graph-based Leiden clustering on the learned contig embeddings to form bins
6. Quality Assessment: Uses miniprot to align bins against a database of eukaryotic core genes to detect contamination
7. Iterative Refinement: Automatically splits contaminated bins based on core gene duplications, then tests lower resolutions to find the most conservative solution

Key Features

• Automatic Eukaryotic Filtering: The HyenaDNA classifier uses a pre-trained genomic foundation model to identify and retain eukaryotic sequences
• Multi-Sample Support: Can process coverage information from multiple samples (BAM/CRAM files) simultaneously
• Adaptive Resolution: Automatically determines optimal clustering resolution based on bin completeness and contamination
• Barlow Twins Loss: Uses a self-supervised contrastive learning approach that doesn't require negative pairs
• Fragment Augmentation: Large contigs are split into multiple overlapping fragments during training to improve representation learning
• Conservative Refinement: After successful bin refinement, tests lower resolutions to find the most consolidated solution that maintains quality

Options

Use remag -h for quick reference or remag --help for full documentation.

Essential Options

  FASTA_ARG                       Input FASTA file (positional argument). Can also use -f/--fasta
  -f, --fasta PATH                Input FASTA file with contigs to bin. Can be gzipped.
  -c, --coverage PATH             Coverage files for calculation. Supports BAM, CRAM (indexed), and TSV formats.
                                  Auto-detects format by extension. Supports space-separated paths and glob patterns
                                  (e.g., "*.bam", "*.cram", "*.tsv"). Use quotes around glob patterns.
  -o, --output PATH               Output directory for results. [default: remag_output in FASTA directory]
  -t, --threads INTEGER           Number of CPU cores to use for parallel processing.  [default: 8]
  -v, --verbose                   Enable verbose logging.
  -k, --keep-intermediate         Keep intermediate files (embeddings, features, model, etc.).
  -h, --help                      Show quick reference or full help.

Advanced Options

For complete list of advanced options (neural network parameters, clustering settings, refinement options, etc.), run:

remag --help

Output

REMAG produces several output files:

Core output files (always created):

• bins/: Directory containing FASTA files for each bin
• bins.csv: Final contig-to-bin assignments
• embeddings.csv: Contig embeddings from the neural network
• remag.log: Detailed log file
• *_eukaryotic_filtered.fasta: Filtered FASTA file with only eukaryotic contigs retained (when eukaryotic filtering is enabled)

Additional files (with -k / --keep-intermediate option):

• siamese_model.pt: Trained Siamese neural network model
• kmer_embeddings.csv: K-mer encoder embeddings (before fusion)
• coverage_embeddings.csv: Coverage encoder embeddings (before fusion)
• params.json: Complete run parameters for reproducibility
• features.csv: Extracted k-mer and coverage features
• fragments.pkl: Fragment information used during training
• hyenadna_classification_results.csv: HyenaDNA eukaryotic classification results
• organism_estimation_gene_counts.json: Gene counts used for adaptive resolution determination
• refinement_summary.json: Summary of the bin refinement process
• gene_contig_mappings.json: Cached gene-to-contig mappings for faster refinement
• core_gene_duplication_results.json: Core gene duplication analysis from refinement
• temp_miniprot/: Temporary directory for miniprot alignments (removed unless --keep-intermediate)

Visualization (optional, requires plotting dependencies):

To generate UMAP visualization plots:

# Install plotting dependencies if not already installed
pip install remag[plotting]

# Generate UMAP visualization from embeddings
python scripts/plot_features.py --features output_directory/embeddings.csv --clusters output_directory/bins.csv --output output_directory

This creates:

• umap_coordinates.csv: UMAP projections for visualization
• umap_plot.pdf: UMAP visualization plot with cluster assignments

Requirements

Core dependencies (always installed):

• Python 3.9+
• PyTorch (≥1.11.0)
• einops (≥0.6.0) - for HyenaDNA model operations
• scikit-learn (≥1.0.0)
• leidenalg (≥0.9.0) - for graph-based clustering
• igraph (≥0.10.0) - for graph construction in Leiden clustering
• pandas (≥1.3.0)
• numpy (≥1.21.0)
• pysam (≥0.18.0)
• loguru (≥0.6.0)
• tqdm (≥4.62.0)
• rich-click (≥1.5.0)

External dependencies (must be installed separately):

• miniprot - Required for core gene analysis and quality assessment
  • Install with: conda install -c bioconda miniprot

Optional dependencies:

• For visualization: matplotlib (≥3.5.0), umap-learn (≥0.5.0)
  • Install with: pip install remag[plotting]

The package includes a pre-trained HyenaDNA classifier model for eukaryotic contig filtering. The HyenaDNA model is a genomic foundation model based on the Hyena operator architecture.

Acknowledgments

The integrated HyenaDNA classifier uses a pre-trained genomic foundation model:

• Repository: HazyResearch/hyena-dna
• Paper: Nguyen E, Poli M, Faizi M, et al. HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution. NeurIPS 2023.

License

MIT License - see LICENSE file for details.

Citation

If you use REMAG in your research, please cite:

@software{gomez_perez_2025_remag,
  author       = {Gómez-Pérez, Daniel},
  title        = {REMAG: Recovering high-quality Eukaryotic genomes from complex metagenomes},
  year         = 2025,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.16443991},
  url          = {https://doi.org/10.5281/zenodo.16443991}
}

Note: The DOI 10.5281/zenodo.16443991 represents all versions and will always resolve to the latest release. A manuscript describing REMAG is in preparation.