name: comebin description: > COMEBin — contrastive multi-view representation learning for metagenomic contig binning (composition + coverage). Use when: (1) Running COMEBin as a modern single binner, (2) Comparing to SemiBin2/VAMB/MetaBAT2, (3) Co- assembly, single-sample, or multi-sample binning modes, (4) Feeding bins into ensemble refinement. Prefer ensemble (basalt ★ / metawrap / dastool) for catalogues. license: MIT metadata: version: "1.0.1" metagenomics.category: "analysis-tools" metagenomics.tags: "binning, COMEBin, contrastive-learning, MAG" metagenomics.upstream: "https://github.com/ziyewang/COMEBin" metagenomics.stage: "binning" metagenomics.maturity: "reviewed" metagenomics.upstream-version: "snapshot 1b476a4" metagenomics.last-reviewed: "2026-10-18" metagenomics.command-status: "CLI reviewed against upstream; not executed here"

COMEBin

Upstream: ziyewang/COMEBin

Citation

Wang, Z. et al. Effective binning of metagenomic contigs using contrastive multi-view representation learning. Nat. Commun. 15, 585 (2024). https://doi.org/10.1038/s41467-023-44290-z

See also docs/references.md.

Analytical thinking

COMEBin learns contig embeddings with contrastive multi-view learning: data augmentation creates multiple views per contig; a coverage module maps variable sample counts to fixed-dimensional coverage embeddings. Benchmarks (CAMI II and real data) show strong near-complete MAG recovery without SemiBin1-style taxonomic semi-supervision.

Use as a diverse single-binner in ensembles (different inductive bias from MetaBAT2 / MaxBin2 / VAMB). Prefer multi-coverage inputs when related samples exist (Mattock & Watson 2023 — single-coverage hides contamination).

Prerequisites

How to run

set -euo pipefail
[[ -s contigs.fa && -d temp/map ]] || exit 2
find temp/map -maxdepth 1 -name '*.bam' -size +0c | grep -q .
command -v run_comebin.sh >/dev/null
run_comebin.sh -a "$(realpath contigs.fa)" \
  -o "$(realpath -m temp/bins/comebin)" \
  -p "$(realpath temp/map)" -d cuda -t 32 \
  >comebin.stdout.log 2>comebin.stderr.log

Use -d cpu when no compatible CUDA device is available. BAMs must be coordinate-sorted, indexed, and aligned to the exact assembly; inspect run_comebin.sh -h in the pinned release before changing modes.

Decision tree

Single binner choice?
├─ Classical TNF+coverage → metabat2 / concoct / maxbin2
├─ VAE latent → vamb
├─ Self-supervised contrastive → semibin2
├─ Multi-view contrastive → comebin
└─ Catalogue / multi-assembly → basalt ★ (ensemble)

Methods caveats

• Contrastive gains are benchmark-specific; always QC with CheckM2/GUNC.
• Report mode (co-assembly / single / multi-sample) and coverage recipe.
• Combine with ensemble tools rather than publishing one binner’s raw dump.

Validation, recovery, and provenance

Require completed logs, candidate FASTAs/assignment tables, exact contig membership and no duplicate final assignments. Reconcile all BAM reference headers/lengths to the assembly, summarize binned/unbinned bases, and evaluate bins with CheckM2/GUNC under the same policy used for comparator binners. GPU completion alone is not a quality result.

Use a fresh output after changing assembly, BAM set, device or release. Record COMEBin/PyTorch/CUDA commit and versions, command, assembly/BAM checksums, ordered sample mapping, minimum contig policy, device/seed where exposed, assignments and candidate checksums. Bins remain candidates until independent QC/refinement.

Related skills

semibin2 · vamb · metabat2 · basalt · dastool · checkm2 · tool-selection · metagenomics-workflow