semibin2

name	semibin2
description	Self-supervised contrastive learning binning for short- and long-read assemblies. Use this skill when working with semibin2: installing, choosing parameters, running commands, interpreting outputs, or troubleshooting. Upstream: https://github.com/BigDataBiology/SemiBin. For stage routing use tool-selection / metagenomics-workflow.
license	MIT
category	analysis-tools
tags	["binning","semibin","self-supervised"]
upstream	https://github.com/BigDataBiology/SemiBin
stage	binning

SemiBin2

Upstream: SemiBin2

Citation

Pan, S., Zhao, X.-M. & Coelho, L. P. SemiBin2: self-supervised contrastive learning leads to better MAGs for short- and long-read sequencing. Bioinformatics 39, i21–i29 (2023). https://doi.org/10.1093/bioinformatics/btad209

See also docs/references.md.

Analytical thinking

SemiBin1 used semi-supervised contrastive learning with contig annotation — costly and reference-biased. SemiBin2 switches to self-supervised embeddings from contigs alone, improving HQ MAG counts while cutting runtime (~25%) and peak memory (~11%) vs SemiBin1 on real short-read samples in the paper.

For long reads, SemiBin2 adds an ensemble-based DBSCAN clustering path (reported 13–26% more HQ genomes than the next-best long-read binner in that study). Compare with dedicated lorbin on long-read assemblies.

Still a single binner: for catalogues, feed into basalt ★ / metawrap / dastool. Prefer multi-sample coverage when related samples exist.

How to run

conda install -c bioconda semibin2  # if packaged; else follow upstream

# short-read / general easy mode
SemiBin2 single_easy_bin \
  -i contigs.fa \
  -b temp/map/*.bam \
  -o temp/bins/semibin2 \
  -t 16

# multi-sample / long-read modes: see SemiBin2 docs (environment + DBSCAN options)

Key parameters

Mode	When
single_easy_bin	One assembly + BAMs
multi-sample modes	Related samples with shared genomes
long-read / DBSCAN path	Long-read contigs — verify flags in upstream

Literature notes

• Prefer SemiBin2 over SemiBin1 unless reproducing an old Methods block.
• Self-supervision removes annotation bottleneck but does not remove need for CheckM2/GUNC.
• Ensemble with classical binners increases recovery diversity (DAS Tool / MetaWRAP / BASALT logic).

Decision notes

• Stage: binning
• Long-read alternative/complement: lorbin
• Ensemble: basalt ★ · metawrap · dastool

Related skills

metabat2 · vamb · comebin · lorbin · dastool · basalt · bowtie2 · tool-selection · metagenomics-workflow