maxbin-rs

A Rust reimplementation of MaxBin2 by Yu-Wei Wu, Blake A. Simmons, and Steven W. Singer — a metagenome binning tool that clusters assembled contigs into individual genomes using tetranucleotide frequency and abundance information.

Wu Y-W, Simmons BA, Singer SW. MaxBin 2.0: an automated binning algorithm to recover genomes from multiple metagenomic datasets. Bioinformatics. 2016;32(4):605–607. doi:10.1093/bioinformatics/btv638

This project reimplements their algorithm while addressing long-standing packaging and reliability issues. Not yet tested inside nf-core/mag or other pipeline managers — if you try it, please report issues.

If referring specifically to this reimplementation, please also cite:

Kroneman W. maxbin-rs: a Rust reimplementation of MaxBin2. 2026. https://github.com/werner291/maxbin-rs

Always cite the original MaxBin2 first — the algorithm is theirs.

Quick start

# Nix (builds everything from source, pinned deps)
nix run github:werner291/maxbin-rs -- \
  --contig contigs.fa.gz --reads reads.fastq.gz --out my_bins/

# Docker (for pipeline integration — pin to a version tag)
docker run --rm -v /path/to/data:/data \
  ghcr.io/werner291/maxbin-rs:v0.3.0 \
  --contig /data/contigs.fa.gz --reads /data/reads.fastq.gz --out /data/bins

Both bundle HMMER, Bowtie2, and FragGeneScan. A bioconda package is planned. You can also cargo build --release if you have these tools on PATH.

For a drop-in replacement that matches the original MaxBin2 CLI exactly, use a v0.1.x release (e.g. ghcr.io/werner291/maxbin-rs:v0.1.3).

Inputs

• --contig — assembled contigs (FASTA, optionally gzipped).
• --abund — per-contig read depth (tab-separated: contig name, depth). Pass once per sample: --abund sample1.txt --abund sample2.txt --abund sample3.txt
• --reads — raw reads (FASTQ, optionally gzipped). Abundance is computed via Bowtie2 mapping. Pass once per sample: --reads sample1.fq.gz --reads sample2.fq.gz

Use --abund or --reads, not both. Default minimum contig length is 1000 bp. Run maxbin-rs --help for all options.

Output

my_bins/
├── 001.fasta    # bin 1
├── 002.fasta    # bin 2
├── ...
├── summary      # per-bin completeness, genome size, GC
├── noclass      # contigs not assigned to any bin
├── marker       # marker gene counts per bin
└── log          # tool version and run metadata

EM subcommand

The EM algorithm is the core of MaxBin. If you already have abundance data and seed contigs (e.g. from your own marker gene pipeline), you can run it directly without any external tools:

maxbin-rs em --contig filtered.fa --abund depth.txt --seed seeds.txt --out result

The default mode (maxbin-rs pipeline) wraps this with gene calling (FragGeneScanRs), marker detection (HMMER), and read mapping (Bowtie2) to produce seeds and abundance automatically. You can also supply your own intermediates via --faa (protein FASTA) or --hmmout (HMMER output) to substitute individual stages. See maxbin-rs --help for all options.

Migrating from MaxBin2

# before
run_MaxBin.pl -contig contigs.fa -reads reads.fq -out results/my_sample
# produces: results/my_sample.001.fasta, results/my_sample.summary, ...

# after
maxbin-rs --contig contigs.fa --reads reads.fq --out results/my_sample/
# produces: results/my_sample/001.fasta, results/my_sample/summary, ...

What to update in your scripts:

• Binary: maxbin-rs instead of run_MaxBin.pl
• Flags use double dashes: --contig, --reads, --out
• --out is a directory, not a filename prefix
• Output files: 001.fasta instead of prefix.001.fasta
• Version string: maxbin-rs X.Y.Z instead of MaxBin 2.2.7
• Progress goes to stderr, not stdout
• Exit code 1 on error (original uses -1 / 255)
• Not yet produced: per-bin marker tarball, multi-sample .abundance file (both optional: true in nf-core/mag)

If you can't change your scripts at all, v0.1.x uses the original single-dash flags and prefix-based --out — it's a drop-in replacement for run_MaxBin.pl. Bug fixes and performance improvements are backported as long as they don't change output.

See VERSIONING.md for the full version policy.

Verification

The v0.1.x series was verified to produce identical output to MaxBin2 2.2.7. Later versions build on that as a baseline, adding bug fixes, performance improvements, and interface changes. The evidence is described in detail in PAPER.md. In short:

• Unit/property tests compare each Rust function against the original C++ via FFI on randomized inputs (cargo nextest run).
• End-to-end tests run the full pipeline (including recursive binning) on CAMI I High: all bins byte-identical.
• The EM core is approximately 5x faster than the original C++. Gene calling uses FragGeneScanRs (Rust library) instead of shelling out to FragGeneScan (C + Perl), roughly halving pipeline time on pre-computed abundance.

CAPES_S7, pre-computed abundance, AMD Ryzen 9 7940HS. When using raw reads instead of abundance, Bowtie2 read mapping dominates (~32 min single-threaded on CAPES_S7 vs ~45s for everything else). Regenerate: nix build .#bench-chart

These results have not been independently verified. If you do run your own comparison, we'd welcome the feedback.