Metagenome assembly

Version Compatibility

Reference examples tested with: Flye 2.9+, SPAdes 3.15+ (metaSPAdes), MEGAHIT 1.2+, hifiasm-meta 0.3+, metaMDBG 1.0+, MetaBAT2 2.15+, MaxBin 2.2.7+, CONCOCT 1.1+, SemiBin 2.0+, VAMB 4.1+, DAS_Tool 1.1.6+, CheckM2 1.0+, GUNC 1.0+, GTDB-Tk 2.4+, inStrain 1.7+, minimap2 2.26+, samtools 1.19+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags
• Python: pip show <package> then help(module.function) to check signatures

GTDB-Tk results track the reference-package RELEASE (e.g. R214 vs R220); the DB release MUST match the GTDB-Tk binary or classification silently fails. CheckM2 and GUNC each download their own DIAMOND DB. SemiBin2's pretrained --environment models are versioned. If code throws an error, introspect the installed tool and adapt rather than retrying.

Metagenome Assembly

"Assemble genomes from my metagenome" -> Co-assemble a community at uneven, strain-mixed coverage, then bin the contigs into a set of consensus population genomes (MAGs) and QC each against MIMAG. The deliverable is MAGs, not a single assembly.

• CLI: flye --meta --nano-hq reads.fq (ONT), spades.py --meta -1 R1.fq -2 R2.fq or megahit -1 R1.fq -2 R2.fq (Illumina), hifiasm_meta/metaMDBG (HiFi); then binners -> DAS_Tool -> checkm2 predict + gunc run + gtdbtk classify_wf

The Single Most Important Modern Insight -- A Metagenome Is Not a Genome; the Assembler Cannot Assume Uniform Coverage

Every isolate assembler is built on the premise that the true sequence sits at roughly one depth, so a coverage drop or spike signals a repeat or an error. In a community that premise is false by construction: an abundant species at 500x and a rare one at 3x are both real. Running plain SPAdes/Unicycler or single-genome Flye on a community treats the abundance spread and strain bubbles as errors to "fix" and produces garbage. Use --meta modes. Three consequences cascade:

1. A MAG is a population consensus, not an organism's genome. Co-occurring strains differing by <1% ANI become bubbles the assembler collapses into one consensus path -- a sequence that may match no actual cell in the sample. A 99%-complete circular MAG is still the consensus of the dominant strain; minority-strain accessory genome is averaged away. Treat every per-strain, per-allele, or pangenome claim from a single consensus MAG as suspect until read-level microdiversity (inStrain) or strain-aware assembly confirms it.
2. The deliverable is a community of MAGs, not one assembly -- a community has no N50. N50 is dominated by whichever few abundant genomes assembled well and says nothing about the community; a "better N50" assembly can have recovered fewer genomes. Report MAG count split by MIMAG tier (HQ/medium/low) and community fraction binned. Bigger total assembly size is not better -- it can mean more chimeras and strain-fragmentation.
3. Modern practice is multi-binner -> consolidate -> CheckM2 + GUNC -> GTDB-Tk. Never trust one binner; run several (each weights composition vs coverage differently and recovers a partially-different genome set), reconcile with DAS_Tool, then QC every bin with CheckM2 AND GUNC (completeness lies about chimeras) before classifying against the MIMAG 90%/5% bar. HiFi/long reads are the single biggest quality jump: they span the conserved rRNA operons and strain bubbles short reads shred, yielding complete, circular, genuinely-HQ MAGs.

Assembler Taxonomy

Tool	Citation	Mechanism / role	When
metaFlye	Kolmogorov 2020 Nat Methods	repeat-graph long-read meta-assembler (--meta mandatory)	ONT/CLR community de novo; polish after
metaSPAdes	Nurk 2017 Genome Res	strain-aware multi-k de Bruijn (spades.py --meta)	Illumina, contiguity priority; ONE paired library only
MEGAHIT	Li 2015 Bioinformatics	succinct de Bruijn, low-memory	huge/complex Illumina co-assemblies, soil; lower contiguity
hifiasm-meta	Feng 2022 Nat Methods	strain-resolved HiFi string graph	PacBio HiFi communities; keeps strains apart
metaMDBG	Benoit 2024 Nat Biotechnol	minimizer de Bruijn for HiFi	HiFi; often ~2x the HQ circular MAGs, low RAM
OPERA-MS	Bertrand 2019 Nat Biotechnol	short-read meta scaffolded by long reads	hybrid short + long

Binner Taxonomy

Binner	Citation	Signal	When
MetaBAT2	Kang 2019 PeerJ	tetranucleotide freq (TNF) + coverage, parameter-free	fast default workhorse (-m 1500)
MaxBin2	Wu 2016 Bioinformatics	EM over TNF + marker genes + coverage	single/few samples
CONCOCT	Alneberg 2014 Nat Methods	GMM on composition+coverage of cut-up contigs	many samples; 4-step pipeline, not one command
SemiBin2	Pan 2023 Bioinformatics	self-supervised contrastive deep learning	current SOTA; short + long; pretrained env models
VAMB	Nissen 2021 Nat Biotechnol	variational autoencoder of coabundance + k-mer	multi-sample; separates close strains
DAS_Tool	Sieber 2018 Nat Microbiol	dereplicate-aggregate-score across binners (consolidation)	ALWAYS run; non-redundant set beats any single binner

Decision Tree by Scenario

Scenario	Recommended	Why
Illumina, complex/huge/soil, low RAM	MEGAHIT --presets meta-sensitive	succinct dBG fits in memory; multi-library
Illumina, contiguity priority, tractable size	metaSPAdes (spades.py --meta)	strain-aware repeat resolution; merge libraries first (one paired lib only)
ONT-only community	metaFlye --meta --nano-hq -> polish	repeat-graph meta mode; ONT needs polishing -> assembly-polishing
PacBio HiFi community	hifiasm-meta or metaMDBG	complete circular strain-resolved MAGs; fixes rRNA collapse
Hybrid short + long	OPERA-MS	short-read meta scaffolded with long reads
Recover MAGs (any assembly)	>=2-3 binners -> DAS_Tool -> CheckM2 + GUNC -> GTDB-Tk	ensemble beats one binner; chimera + taxonomy gates
Only ONE sample	composition-only binning, expect weak bins	differential coverage needs multiple samples; add samples, not tuning
Multiple samples available	map ALL samples to each assembly for binning depth	differential-coverage is the strongest binning signal
Strain-level question	-> inStrain on reads mapped to MAGs	consensus MAGs blur strains; needs read-level microdiversity
Read-based taxonomy / rare biosphere	-> metagenomics/kraken-classification	assembly is blind below the abundance-detection limit
Reads not QC'd / host-contaminated	-> long-read-sequencing/long-read-qc	remove host reads vs a T2T reference before assembly
MAG contamination forensics	-> contamination-detection	detailed CheckM2/GUNC interpretation

metaFlye (ONT / Long Reads)

flye --meta --nano-hq ont.fastq.gz --out-dir flye_out -t 32
#   --meta        uneven-coverage metagenome mode (REQUIRED for communities)
#   read-type flag (mutually exclusive): --nano-hq (Guppy5+/Q20) | --nano-raw (older) |
#       --pacbio-hifi | --pacbio-raw (CLR)
# outputs: assembly.fasta, assembly_graph.gfa, assembly_info.txt (circularity flag in col 'circ.')

ONT contigs are contiguous but error-prone (indels in homopolymers); polish before downstream use (-> assembly-polishing; medaka needs the matching basecaller model). HiFi usually needs no polishing.

metaSPAdes / MEGAHIT (Illumina)

# metaSPAdes -- contiguity priority; exactly ONE paired library
spades.py --meta -1 R1.fastq.gz -2 R2.fastq.gz -o spades_out -t 32 -m 500
#   -m memory cap in GB (SPAdes aborts if exceeded); -k auto by default
# outputs: contigs.fasta, scaffolds.fasta

# MEGAHIT -- huge/low-RAM; accepts comma-separated multiple libraries
megahit -1 a1.fq.gz,b1.fq.gz -2 a2.fq.gz,b2.fq.gz -o megahit_out -t 32 \
        --presets meta-sensitive --min-contig-len 1000
#   --presets meta-sensitive | meta-large (huge complex); raise --min-contig-len to ~1000 for binning

metaSPAdes --meta supports exactly ONE paired-end library -- a real constraint people miss; concatenate libraries first or use MEGAHIT for many. metaSPAdes is heavier on RAM/time and chokes on soil-scale co-assembly; MEGAHIT assembled a 252 Gbp soil set on one node at the cost of somewhat more fragmentation.

HiFi (the transformative case)

hifiasm_meta -t 32 -o asm hifi.fastq.gz
awk '/^S/{print ">"$2"\n"$3}' asm.p_ctg.gfa > asm.p_ctg.fa   # GFA -> FASTA

metaMDBG asm --out-dir mdbg_out --in-hifi hifi.fastq.gz --threads 32   # often ~2x HQ circular MAGs

Coverage for Binning, then Bin

# Map reads back to the assembly -- per sample for differential coverage
minimap2 -ax map-ont -t 32 contigs.fa reads.fq.gz | samtools sort -@ 32 -o s1.sorted.bam -
samtools index s1.sorted.bam
jgi_summarize_bam_contig_depths --outputDepth depth.txt s1.sorted.bam s2.sorted.bam   # all samples

metabat2 -i contigs.fa -a depth.txt -o metabat/bin -m 1500 -t 32   # -m 1500 = min contig (do not go below ~1000)
run_MaxBin.pl -contig contigs.fa -abund abund1.txt -out maxbin/bin -thread 32 -min_contig_length 1000
SemiBin2 single_easy_bin -i contigs.fa -b s1.sorted.bam -o semibin_out   # add --environment human_gut for a pretrained model

CONCOCT is a 4-step pipeline (cut_up_fasta.py -> concoct_coverage_table.py -> concoct -> merge_cutup_clustering.py -> extract_fasta_bins.py), not one command. Differential-coverage binning needs MULTIPLE samples with abundance variation; with one sample binning collapses to weak composition-only signal -- more samples, not more tuning.

Consolidate (DAS_Tool), then QC

# Convert each binner's output to contig2bin tables, then aggregate-and-score
Fasta_to_Contig2Bin.sh -i metabat/ -e fa    > metabat.tsv
Fasta_to_Contig2Bin.sh -i maxbin/  -e fasta > maxbin.tsv               # MaxBin emits .fasta
gunzip -k semibin_out/output_bins/*.gz 2>/dev/null || true             # SemiBin2 bins are gzipped
Fasta_to_Contig2Bin.sh -i semibin_out/output_bins/ -e fa > semibin.tsv
DAS_Tool -i metabat.tsv,maxbin.tsv,semibin.tsv -l metabat,maxbin,semibin \
         -c contigs.fa -o dastool/DAS --write_bins -t 32
# DAS_Tool assumes the SAME contig set across binners -- feeding bins from different assemblies is a silent error

checkm2 predict --input dastool/DAS_DASTool_bins/ -x fa --output-directory checkm2_out -t 32
gunc run --input_dir dastool/DAS_DASTool_bins/ --file_suffix .fa --out_dir gunc_out --threads 32
gtdbtk classify_wf --genome_dir dastool/DAS_DASTool_bins/ -x fa --out_dir gtdbtk_out --cpus 32

Run CheckM2 AND GUNC: CheckM2 counts marker copy number (completeness/contamination), GUNC tests whether a genome's genes share one lineage (chimerism). A bin made of two half-genomes can score high completeness, low contamination, and still be a chimera -- GUNC is the orthogonal catch. Dereplicate MAGs across samples (dRep ~95% ANI species, ~99% strain) before reporting counts.

Strain Resolution

# Consensus MAGs blur strains; recover read-level microdiversity
inStrain profile sample.sorted.bam mags.fa -o instrain_out -p 16 -g genes.fna
inStrain compare -i instrain_A instrain_B -o instrain_compare   # popANI: shared-strain detection across samples

Per-Method Failure Modes

Isolate assembler on a community

Trigger: plain spades.py/flye (no --meta) on community reads. Mechanism: uniform-coverage assumption deletes rare-taxon contigs and mis-resolves strain bubbles as errors. Symptom: few short contigs, missing abundant taxa. Fix: --meta mode for every meta-assembler.

Single-binner pipeline

Trigger: "we used SemiBin2 because it's SOTA," one binner. Mechanism: each binner recovers a partially-different genome set. Symptom: real MAGs left on the table; lower count than peers. Fix: run >=2-3 binners -> DAS_Tool consolidation.

Single-sample differential-coverage expectation

Trigger: MetaBAT2/CONCOCT on one sample, surprised bins are bad. Mechanism: one coverage value -> composition (TNF) only, which is weak (related genera share TNF). Symptom: poor, split bins. Fix: more samples with abundance variation, then map all back; do not retune.

Short-read MAG reported HQ on 90/5 alone

Trigger: calling a 95%-complete/2%-contam short-read MAG "high-quality." Mechanism: conserved + multi-copy rRNA operon tangles the short-read graph and stays unbinned. Symptom: MAG fails MIMAG HQ for missing 16S/23S/5S despite great completeness. Fix: check the FULL MIMAG HQ definition (rRNA + tRNA); use HiFi/long reads to span the operon.

High CheckM2 completeness read as quality

Trigger: trusting completeness/contamination alone. Mechanism: non-overlapping markers from two half-genomes score high completeness, low contamination. Symptom: "clean" MAG that is a chimera. Fix: always pair CheckM2 with GUNC -> contamination-detection.

Strain claims from a consensus MAG

Trigger: per-allele/per-strain interpretation of one MAG. Mechanism: the assembler collapsed strains to consensus. Symptom: strain findings that read-level data contradict. Fix: inStrain popANI or strain-aware (hifiasm-meta) assembly.

Reading MAG richness as community richness

Trigger: "200 MAGs cover 60% of reads" treated as the whole community. Mechanism: the rare biosphere never crosses the assembly detection limit. Symptom: species richness massively undercounted. Fix: complement with read-based profiling -> metagenomics/kraken-classification.

Quantitative Thresholds

Threshold	Source	Rationale
MIMAG high-quality: completeness >90%, contamination <5%, AND 16S/23S/5S rRNA + >=18 tRNAs	Bowers 2017 Nat Biotechnol	rRNA criterion is the silent killer; short-read MAGs fail it
MIMAG medium-quality: completeness >=50%, contamination <10%	Bowers 2017	most short-read MAGs land here
Min contig length for binning ~1000-1500 bp	MetaBAT2 default 1500	shorter contigs have unreliable TNF/coverage -> noise/chimeras
Differential-coverage binning needs N >= ~3-5 samples with abundance variation	binning practice	one coverage value cannot separate co-abundant genomes
Assembly detection limit ~3-5x coverage	de Bruijn requirement	below this the rare biosphere does not assemble at all
MAG dereplication 95% ANI (species), 99% (strain)	dRep/ANI convention	collapse redundant per-sample MAGs before counting
metaSPAdes input: exactly ONE paired library	--meta constraint	merge libraries or use MEGAHIT for many
N50 / contiguity	not a community metric	report MAG count + MIMAG tiers instead

Common Errors

Error / symptom	Cause	Solution
Few short contigs, missing abundant taxa	isolate mode on a community	add --meta
metaSPAdes rejects multiple libraries	--meta supports one paired library	concatenate, or use MEGAHIT
metaSPAdes aborts / out of memory	RAM cap hit on a complex co-assembly	raise -m, or switch to MEGAHIT
Poor bins from one sample	no differential-coverage signal	add samples; map all back for depth
"HQ MAG" lacks rRNA	rRNA-operon collapse in short reads	full MIMAG check; HiFi/long reads
Clean CheckM2 but suspect bin	chimera invisible to marker counts	run GUNC alongside CheckM2
GTDB-Tk classify_wf errors	DB release != binary version	match the GTDB reference-package release
Strain finding not reproducible	consensus MAG blurs strains	inStrain popANI / strain-aware assembly

References

• Nurk S, Meleshko D, Korobeynikov A, Pevzner PA. 2017. metaSPAdes: a new versatile metagenomic assembler. Genome Res 27:824-834.
• Li D, Liu CM, Luo R, Sadakane K, Lam TW. 2015. MEGAHIT: an ultra-fast single-node solution for large and complex metagenomics assembly via succinct de Bruijn graph. Bioinformatics 31:1674-1676.
• Kolmogorov M, et al. 2020. metaFlye: scalable long-read metagenome assembly using repeat graphs. Nat Methods 17:1103-1110.
• Feng X, Cheng H, Portik D, Li H. 2022. Metagenome assembly of high-fidelity long reads with hifiasm-meta. Nat Methods 19:671-674.
• Benoit G, et al. 2024. High-quality metagenome assembly from long accurate reads with metaMDBG. Nat Biotechnol 42:1378-1383.
• Bertrand D, et al. 2019. Hybrid metagenomic assembly enables high-resolution analysis of resistance determinants and mobile elements in human microbiomes (OPERA-MS). Nat Biotechnol 37:937-944.
• Kang DD, et al. 2019. MetaBAT 2: an adaptive binning algorithm for robust and efficient genome reconstruction from metagenome assemblies. PeerJ 7:e7359.
• Wu YW, Simmons BA, Singer SW. 2016. MaxBin 2.0: an automated binning algorithm to recover genomes from multiple metagenomic datasets. Bioinformatics 32:605-607.
• Alneberg J, et al. 2014. Binning metagenomic contigs by coverage and composition (CONCOCT). Nat Methods 11:1144-1146.
• Pan S, Zhao XM, Coelho LP. 2023. SemiBin2: self-supervised contrastive learning leads to better MAGs for short- and long-read sequencing. Bioinformatics 39:i21-i29.
• Nissen JN, et al. 2021. Improved metagenome binning and assembly using deep variational autoencoders (VAMB). Nat Biotechnol 39:555-560.
• Sieber CMK, et al. 2018. Recovery of genomes from metagenomes via a dereplication, aggregation and scoring strategy (DAS_Tool). Nat Microbiol 3:836-843.
• Chklovski A, et al. 2023. CheckM2: a rapid, scalable and accurate tool for assessing microbial genome quality using machine learning. Nat Methods 20:1203-1212.
• Orakov A, et al. 2021. GUNC: detection of chimerism and contamination in prokaryotic genomes. Genome Biol 22:178.
• Chaumeil PA, Mussig AJ, Hugenholtz P, Parks DH. 2020. GTDB-Tk: a toolkit to classify genomes with the Genome Taxonomy Database. Bioinformatics 36:1925-1927.
• Bowers RM, et al. 2017. Minimum information about a single amplified genome (MISAG) and a metagenome-assembled genome (MIMAG) of bacteria and archaea. Nat Biotechnol 35:725-731.
• Olm MR, et al. 2021. inStrain profiles population microdiversity from metagenomic data and sensitively detects shared microbial strains. Nat Biotechnol 39:727-736.

Related Skills

• contamination-detection - CheckM2/GUNC interpretation and chimerism forensics for MAGs
• assembly-qc - Isolate-assembly QC; the uniform-coverage assumption metagenomes abandon
• assembly-polishing - Polish ONT/CLR meta-contigs before binning (HiFi usually needs none)
• metagenomics/kraken-classification - Read-based taxonomy; recovers the rare biosphere assembly cannot
• metagenomics/abundance-estimation - Community abundance downstream of recovered MAGs
• metagenomics/functional-profiling - Functional potential complementary to genome recovery
• long-read-sequencing/long-read-qc - Read-level QC and host removal before assembly