Bio Applied Assembly Binning

Metagenomic Assembly, Binning, and MAGs

When to Use

• Assembling shotgun metagenomic reads (post-QC/decontamination) into contigs.
• Recovering metagenome-assembled genomes (MAGs) from a microbial community sample.
• Choosing/comparing binners (MetaBAT2, CONCOCT, MaxBin2) or dereplicating with DAS_Tool.
• Grading bin quality (completeness/contamination) against the MIMAG standard before deposition or downstream analysis.
• Deciding between single-sample vs multi-sample (co-abundance) binning strategy.

Version Compatibility

MEGAHIT ≥ 1.2.9, metaSPAdes (SPAdes) ≥ 3.15, MetaBAT2 ≥ 2.15, CONCOCT ≥ 1.1, MaxBin2 ≥ 2.2.7, DAS_Tool ≥ 1.1.6, CheckM ≥ 1.2 (or CheckM2 ≥ 1.0), bowtie2 ≥ 2.5, samtools ≥ 1.17, Python ≥ 3.10 with pandas/numpy for downstream QC.

Prerequisites

• Reads already quality-trimmed and host/contaminant-screened (see bio-read-qc-contamination-screening).
• Tools on PATH: megahit, bowtie2, samtools, jgi_summarize_bam_contig_depths (ships with MetaBAT2), metabat2, das_tool, checkm.
• pandas, numpy for parsing depth tables and CheckM output in Python.

Workflow Overview

reads → MEGAHIT assembly → filter ≥500 bp contigs
      → map reads back → jgi_summarize_bam_contig_depths
      → MetaBAT2/CONCOCT/MaxBin2 binning
      → DAS_Tool dereplication
      → CheckM quality assessment
      → filter HQ MAGs (≥90% complete, <5% contamination)

Goal: turn decontaminated paired-end reads into a set of genome-resolved MAGs. Approach: assemble once with MEGAHIT (fast, low RAM), recruit coverage signal by mapping reads back, bin with multiple tools and dereplicate with DAS_Tool, then gate bins on CheckM's MIMAG tiers.

# 1. Assembly
megahit \
    -1 decontam_1.fastq.gz \
    -2 decontam_2.fastq.gz \
    -o megahit_assembly/ \
    --min-contig-len 500 \
    --k-list 21,29,39,59,79,99,119,141 \
    -t 16 \
    -m 0.5              # cap at 50% of available RAM
seqkit stats megahit_assembly/final.contigs.fa

# 2. Coverage estimation (map reads back to own assembly)
bowtie2-build megahit_assembly/final.contigs.fa contigs_index
bowtie2 -x contigs_index -1 decontam_1.fastq.gz -2 decontam_2.fastq.gz \
    -p 16 --no-unal | samtools sort -@ 8 -o contigs_mapped.bam
samtools index contigs_mapped.bam

# For multi-sample binning: repeat mapping for each sample's BAM,
# then pass ALL bams to jgi_summarize_bam_contig_depths in one call.
jgi_summarize_bam_contig_depths \
    --outputDepth depths.txt \
    contigs_mapped.bam
# columns: contigName, contigLen, totalAvgDepth, sample1.bam, sample1.bam-var, ...

# 3. Binning (run multiple binners, then dereplicate)
metabat2 \
    -i megahit_assembly/final.contigs.fa \
    -a depths.txt \
    -o bins/bin \
    --minContig 1500 \
    --minClsSize 100000 \
    -t 8 --saveCls

das_tool \
    -i metabat2_bins,concoct_bins,maxbin2_bins \
    -l MetaBAT2,CONCOCT,MaxBin2 \
    -c megahit_assembly/final.contigs.fa \
    -o dastool_output/ \
    -t 8 --write_bins

# 4. Quality assessment
checkm lineage_wf dastool_output/_DASTool_bins/ checkm_out/ -x fa -t 16 --pplacer_threads 4
checkm qa checkm_out/lineage.ms checkm_out/ -o 2 > checkm_summary.txt

Binning signals used by MetaBAT2/CONCOCT:

1. Tetranucleotide frequency (TNF) — 256-dim composition vector, genome-specific.
2. Coverage depth across samples — organism-specific abundance (multi-sample = differential coverage = much cleaner bins).

MIMAG Quality Tiers (Bowers et al. 2017, Nature Biotechnology)

CheckM estimates completeness/contamination from ~100 lineage-specific single-copy marker genes expected exactly once in a complete genome; the marker set is chosen from each bin's phylogenetic placement.

Goal: parse assembly/CheckM output and classify bins programmatically instead of eyeballing text reports. Approach: compute N50 from contig lengths, then apply the MIMAG completeness/contamination thresholds to a CheckM qa table.

import numpy as np
import pandas as pd


def compute_n50(contig_lengths: np.ndarray) -> int:
    """Return the N50 of an assembly given an array of contig lengths (bp).

    N50 = length of the contig at which the cumulative sum of all contig
    lengths, sorted longest-first, first reaches 50% of the total assembly.
    """
    lengths_sorted = np.sort(contig_lengths)[::-1]
    cumsum = np.cumsum(lengths_sorted)
    half_total = cumsum[-1] / 2
    n50_idx = np.searchsorted(cumsum, half_total)
    return int(lengths_sorted[n50_idx])


def mimag_quality_tier(completeness: float, contamination: float) -> str:
    """Classify a MAG into a MIMAG tier from CheckM completeness/contamination (%)."""
    if completeness >= 90 and contamination < 5:
        return "High Quality"
    elif completeness >= 50 and contamination < 10:
        return "Medium Quality"
    return "Low Quality"


# Parse `checkm qa -o 2` tab-separated output
checkm_df = pd.read_csv("checkm_summary.txt", sep="\t")
checkm_df = checkm_df.rename(columns=lambda c: c.strip())
checkm_df["Quality"] = checkm_df.apply(
    lambda row: mimag_quality_tier(row["Completeness"], row["Contamination"]), axis=1
)

print(checkm_df["Quality"].value_counts())
hq_bins = checkm_df[checkm_df["Quality"] == "High Quality"]
print(f"High-quality MAGs: {len(hq_bins)} / {len(checkm_df)}")

Binner Comparison

Pitfalls

• Discard short contigs before binning: contigs < 1500 bp carry insufficient TNF signal for reliable clustering — use --minContig 1500 in MetaBAT2.
• Multi-sample binning dramatically improves quality: map all samples to the same assembly and feed all BAMs to jgi_summarize_bam_contig_depths in one call whenever you have ≥2 related samples — differential coverage is the strongest binning signal.
• Strain heterogeneity ≠ contamination: high CheckM contamination with high strain heterogeneity (≥90% AA identity between duplicated marker genes) indicates strain mixing within one population, not foreign contamination.
• Run all three binners, then DAS_Tool: MetaBAT2 alone leaves quality on the table; DAS_Tool-dereplicated bins from MetaBAT2 + CONCOCT + MaxBin2 consistently outperform any single binner.
• MEGAHIT vs metaSPAdes trade-off: MEGAHIT is RAM-efficient and fast; metaSPAdes is more accurate for low-coverage organisms but needs 100–400 GB RAM for complex communities.
• N50 expectations: a typical gut metagenome (5 Gbp data) assembles to ~100k–500k contigs with N50 of 1–10 kb — don't expect chromosome-scale contigs from short-read metagenomics.
• CheckM marker sets can mislead novel taxa: if a bin's closest reference lineage is distant, completeness/contamination estimates from generic (domain-level) marker sets are unreliable — use CheckM2 (ML-based, marker-set-free) for highly novel MAGs.

See Also

• bio-applied-genome-assembly
• bio-applied-taxonomic-profiling
• metagenomics-shotgun
• bio-applied-microbial-diversity