remag 0.1.2

Recovery of high-quality eukaryotic genomes from complex metagenomes

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

# Install from PyPI
pip install remag

# Run REMAG
remag -f contigs.fasta -c alignments.bam -o output_directory

Installation

From PyPI (recommended)

# Create conda environment (optional but recommended)
conda create -n remag python=3.9
conda activate remag

# Install from PyPI
pip install remag

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]"

GPU-accelerated installation

For GPU-accelerated clustering (requires NVIDIA GPU):

# Install with RAPIDS support
pip install "remag[gpu]"

Usage

Command line interface

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

remag -f contigs.fasta -c alignments.bam -o output_directory

Python module mode

python -m remag -f contigs.fasta -c alignments.bam -o output_directory

How REMAG Works

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

1. Bacterial Pre-filtering: By default, REMAG automatically filters out bacterial contigs using the integrated 4CAC 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. HDBSCAN Clustering: Density-based clustering on the learned contig embeddings to form bins
5. Quality Assessment: Uses miniprot to align bins against a database of eukaryotic core genes to detect contamination
6. Iterative Refinement: Automatically splits contaminated bins based on core gene duplications to improve bin quality

Key Features

• Automatic Bacterial Filtering: The 4CAC classifier automatically identifies and removes bacterial sequences before binning
• Multi-Sample Support: Can process coverage information from multiple samples (BAM/CRAM files) simultaneously
• 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

Options

  -f, --fasta PATH                Input FASTA file with contigs to bin. Can be gzipped.  [required]
  -c, --coverage PATH             Coverage files for calculation. Supports BAM, CRAM (indexed), and TSV formats. Auto-detects format by extension. Each file represents one sample. Supports space-separated paths and glob patterns (e.g., "*.bam", "*.cram", "*.tsv"). Use quotes around glob patterns.
  -o, --output PATH               Output directory for results.  [required]
  --epochs INTEGER RANGE          Training epochs for neural network.  [default: 400; 50<=x<=2000]
  --batch-size INTEGER RANGE      Batch size for training.  [default: 2048; 64<=x<=8192]
  --embedding-dim INTEGER RANGE   Embedding dimension for contrastive learning.  [default: 256; 64<=x<=512]
  --base-learning-rate FLOAT RANGE
                                  Base learning rate for optimizer.  [default: 0.008; 0.00001<=x<=0.1]
  --min-cluster-size INTEGER RANGE
                                  Minimum fragments per cluster.  [default: 2; 2<=x<=100]
  --min-samples INTEGER RANGE     Minimum samples for HDBSCAN core points.  [default: None; 1<=x<=100]
  --cluster-selection-epsilon FLOAT RANGE
                                  Epsilon for HDBSCAN cluster selection.  [default: 0.0; 0.0<=x<=1.0]
  --min-contig-length INTEGER RANGE
                                  Minimum contig length in bp.  [default: 1000; 500<=x<=10000]
  --max-positive-pairs INTEGER RANGE
                                  Maximum positive pairs for contrastive learning.  [default: 5000000; 100000<=x<=10000000]
  -t, --threads INTEGER RANGE     Number of CPU threads.  [default: 8; 1<=x<=64]
  --min-bin-size INTEGER RANGE    Minimum bin size in bp.  [default: 100000; 50000<=x<=10000000]
  -v, --verbose                   Enable verbose logging.
  --skip-bacterial-filter         Skip bacterial contig filtering (4CAC classifier + contrastive learning).
  --skip-refinement               Skip bin refinement.
  --skip-kmeans-filtering         Skip K-means filtering on embeddings.
  --max-refinement-rounds INTEGER RANGE
                                  Maximum refinement rounds.  [default: 2; 1<=x<=10]
  --num-augmentations INTEGER RANGE
                                  Number of random fragments per contig.  [default: 8; 1<=x<=32]
  --keep-intermediate             Keep intermediate files (training fragments, etc.).
  -h, --help                      Show this message and exit.

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
• remag.log: Detailed log file
• *_non_bacterial_filtered.fasta: Filtered FASTA file with bacterial contigs removed (when bacterial filtering is enabled)

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

• embeddings.csv: Contig embeddings from the neural network
• umap_embeddings.csv: UMAP projections for visualization
• umap_plot.pdf: UMAP visualization plot with cluster assignments
• siamese_model.pt: Trained Siamese neural network model
• params.json: Complete run parameters for reproducibility
• features.csv: Extracted k-mer and coverage features
• fragments.pkl: Fragment information used during training
• classification_results.csv: 4CAC bacterial classification results
• refinement_summary.json: Summary of the bin refinement process
• kmeans_filtering_stats.json: Statistics from k-means pre-filtering (if enabled)
• core_gene_duplication_results.json: Core gene duplication analysis from refinement
• temp_miniprot/: Temporary directory for miniprot alignments (removed unless --keep-intermediate)

Requirements

• Python 3.8+
• PyTorch (≥1.11.0)
• scikit-learn (≥1.0.0)
• XGBoost (≥1.6.0) - for 4CAC classifier
• HDBSCAN (≥0.8.28)
• UMAP (≥0.5.0)
• pandas (≥1.3.0)
• numpy (≥1.21.0)
• matplotlib (≥3.5.0)
• pysam (≥0.18.0)
• loguru (≥0.6.0)
• tqdm (≥4.62.0)
• rich-click (≥1.5.0)
• joblib (≥1.1.0)

The package includes a pre-trained 4CAC classifier model for bacterial contig filtering. The 4CAC classifier code and models are adapted from the Shamir-Lab/4CAC repository.

Acknowledgments

The integrated 4CAC classifier (xgbclass module) is adapted from the work by Shamir Lab:

• Repository: Shamir-Lab/4CAC
• Paper: Pu L, Shamir R. 4CAC: 4-class classifier of metagenome contigs using machine learning and assembly graphs. Nucleic Acids Res. 2024;52(19):e94–e94.

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.