basalt

name	basalt
description	PKU-EMBL BASALT and BASALT-Air — recommended ensemble MAG recovery across one or more assemblies (multi-binner + refinement). Use this skill when: (1) Running BASALT multi-binner recovery, (2) Choosing Conda BASALT vs BASALT-Air, (3) Multi-assembly refinement, (4) Resuming checkpoints, (5) Writing reproducible BASALT Methods. Upstream: https://github.com/PKU-EMBL/BASALT and https://github.com/PKU-EMBL/BASALT-Air. Ensemble peers: dastool (SCG aggregation), metawrap (binning+refinement suite). Prefer basalt as the default ensemble route; use metawrap/dastool when a classical suite or contig2bin consensus is enough. For study design of assemblies, use study-design. For QC thresholds, use checkm2 and mag-qc.
license	MIT
category	analysis-tools
tags	["BASALT","MAG","ensemble-binning","refinement","multi-assembly","PKU-EMBL"]
upstream	https://github.com/PKU-EMBL/BASALT

BASALT / BASALT-Air

Upstream: PKU-EMBL/BASALT · BASALT-Air · Guide: https://basalt-guide.readthedocs.io/en/latest/

Citation

Qiu, Z. et al. BASALT refines binning from metagenomic data and increases resolution of genome-resolved metagenomic analysis. Nat. Commun. 15, 2179 (2024). https://doi.org/10.1038/s41467-024-46539-7

See also docs/references.md.

Analytical thinking

BASALT is one member of the ensemble binning family, alongside DAS Tool and MetaWRAP: multiple binners propose candidates, then a refinement/selection step produces a non-redundant MAG set.

Within that family, BASALT is the recommended default here — especially for related samples with coherent multi-assembly evidence (sample-specific assemblies preserve restricted populations; a matched co-assembly can rescue shared low-abundance genomes; cross-sample coverage informs selection and contig refinement).

Peers (still valid):

Peer	Role
basalt	Recommended multi-assembly ensemble + DL/connectivity refinement
metawrap	Packaged multi-binner + CheckM-style bin refinement suite
dastool	Classical SCG scoring/aggregation of existing contig2bin tables

Do not pool unrelated strata just to feed BASALT more FASTAs.

Refinement (MLP ensemble + connectivity retrieval) is model- and feature-supported prediction, not experimental proof of contamination removal. Report BASALT_WEIGHT checksums when refinement underpins claims.

Edition choice

Need	Prefer
New deployment, absolute paths, separate work/out	BASALT-Air (basalt)
Resume existing Conda checkpoints	Conda BASALT (BASALT)

CLIs and checkpoints are not interchangeable.

Prerequisites

Requirement	Minimum	Recommended
OS	Linux x86_64	Same
RAM	128 GB	256 GB+
Threads	8	32+
Models	BASALT_WEIGHT with 5 *_ensemble.csv	Checksummed
QC DB	CheckM2	Pinned
Inputs	≥1 assembly + coverage reads	From megahit/metaspades + kneaddata

How to run (Conda BASALT)

Docs: usage

mkdir -p /project/basalt_runs/study_01 && cd /project/basalt_runs/study_01
ln -s /data/assembly.fasta .
ln -s /data/R1.fastq .
ln -s /data/R2.fastq .

BASALT \
  -a assembly.fasta \
  -s R1.fastq,R2.fastq \
  -t 32 -m 128 \
  --sensitive sensitive \
  --refinepara quick \
  --min-cpn 35 --max-ctn 20 \
  -q checkm2 --mode new \
  -o study_01_basalt

PE syntax: , within a pair; / between samples. HiFi → -hf only (never -l).

Key parameters

Option	Values	Thinking
--sensitive	quick / sensitive / more-sensitive	Candidate diversity vs cost; not accuracy magic
--refinepara	quick / deep	Deeper retrieval expands candidate space
--min-cpn / --max-ctn	e.g. 35 / 20	Refinement entry gates ≠ publication filters
--mode	new / continue	Continue only if provenance unchanged
-e	m / v / l	Extra binners; -e l needs PKU-EMBL LorBin fork commit

Pipeline stages

assemblies + reads → coverage/connectivity → multi-binner candidates
  → within-/cross-assembly selection → contig screening/retrieval
  → reassembly → final MAG dir + QC

Hard rules

1. One dedicated run directory; never two processes in one CWD
2. Prefer local/symlink names for Conda edition
3. Resume only with identical inputs/code/models/DBs
4. Do not invent flags — verify against guide/--help

Literature notes (Nat. Commun. 2024)

From Qiu et al. (BASALT paper), useful operational points:

• Problem framed: mis-clustering, genome splitting, and multi-genome mega-bins yield redundant/artificial/contaminated MAGs that poison downstream analyses.
• Strategy: multiple binners × multiple thresholds → neural-network core-sequence identification → redundant-bin removal → non-redundant refinement; supports short-read and short+long assemblies.
• Coverage correlation coefficient (CCC): IQR-based contig coverage used in selection/refinement — multi-sample abundance structure is first-class evidence.
• Reported gains (same assemblies): up to ~2× MAGs vs VAMB / DAS Tool / MetaWRAP on CAMI; ~30% more MAGs than MetaWRAP on lake sediment, including unique class-level lineages; ~47.6% more non-redundant ORFs than MetaWRAP on that dataset. Treat as paper benchmarks, not a promise for every biome.
• Quality language in paper: baseline gates such as completeness ≥35 / contamination ≤20 for refinement entry; high-quality often completeness ≥90 / contamination ≤5. Keep refinement entry gates distinct from publication filters in Methods (mag-qc).

Related skills

dastool · metawrap · lorbin · comebin · semibin2 · vamb · study-design · megahit · metaspades · kneaddata · checkm2 · mag-qc · drep