name: aamb description: > AAMB / AVAMB — adversarial autoencoder metagenomic binning and its ensemble with VAMB for improved near-complete MAG recovery. Use when: (1) Running AAMB as a DL single binner, (2) AVAMB ensemble (VAMB+AAMB) for extra NC genomes, (3) Diversifying bins before DAS Tool/BASALT. Still QC with checkm2; prefer basalt ★ for multi-assembly catalogues. license: MIT metadata: version: "1.0.1" metagenomics.category: "analysis-tools" metagenomics.tags: "binning, AAMB, AVAMB, VAMB, deep-learning" metagenomics.upstream: "https://github.com/RasmussenLab/vamb" metagenomics.stage: "binning" metagenomics.maturity: "reviewed-legacy" metagenomics.upstream-version: "AVAMB workflow 4.1.3; current VAMB snapshot b146fb6" metagenomics.last-reviewed: "2026-10-18" metagenomics.command-status: "legacy workflow CLI reviewed; not executed here"

AAMB / AVAMB

Upstream: RasmussenLab/vamb (AAMB/AVAMB paths documented in the VAMB repository)

Citation

Piera Líndez, P. et al. Adversarial and variational autoencoders improve metagenomic binning. Commun. Biol. 6, 1073 (2023). https://doi.org/10.1038/s42003-023-05452-3

See also docs/references.md.

Analytical thinking

AAMB uses adversarial autoencoders on coabundance + TNF; AVAMB ensembles AAMB with VAMB and recovered substantially more near-complete genomes in the paper vs VAMB alone. Treat as a diverse DL binner for ensembles — not a replacement for basalt multi-assembly refinement.

Related VAMB-family tools: vamb, taxvamb. Current VAMB documentation calls AVAMB obsolete and recommends TaxVAMB when reliable contig taxonomy and sufficient compute are available; use AVAMB mainly for reproducing its published workflow.

Inputs and run

The archived AVAMB Snakemake workflow requires config.json, an ordered one-path-per-line contigs.txt, and a tab-separated samples2data.tsv containing sample ID, R1 and R2. Pin VAMB v4.1.3 for this workflow rather than mixing it with current-main internals.

set -euo pipefail
root=${VAMB_413:?set VAMB_413 to a VAMB v4.1.3 checkout}
[[ -s config.json && -s contigs.txt && -s samples2data.tsv ]] || exit 2
python -m json.tool config.json >/dev/null
snakemake --cores 20 --use-conda --configfile config.json \
  --snakefile "$root/workflow_avamb/avamb.snake.conda.smk" --dry-run
snakemake --cores 20 --use-conda --configfile config.json \
  --snakefile "$root/workflow_avamb/avamb.snake.conda.smk" \
  >avamb.stdout.log 2>avamb.stderr.log

Review outdir, quality thresholds and avamb_params before execution; use a new output when changing them.

Validation, recovery, and provenance

Require completed Snakemake rules, Final_clusters.tsv, Final_bins/quality_report.tsv, matching final FASTAs and VAE/AAE component assignments. Reconcile every final contig, ensure unique retained membership, review CheckM2 gates and compare VAMB/AAMB contributions under identical criteria. Reassess with GUNC and current CheckM2 rather than assuming workflow-era QC is final.

Resume only from unchanged Snakemake inputs/config/environments; otherwise archive and rerun. Record workflow/VAMB/CheckM2 commits, database, config, commands, ordered sample/read/assembly checksums, mapper settings, seeds/device, logs, component/final memberships and candidate checksums. AAMB/AVAMB clusters remain candidate genomes.

Related skills

vamb · taxvamb · semibin2 · dastool · basalt · tool-selection