Machine-learning MAG binning with SemiBin2 and Snakemake (soil metagenomes)

Created in November 17, 2025

2025

Today’s post is about one of my favorite combinations right now:
machine-learning–guided binning with SemiBin2 + a fully automated Snakemake workflow for recovering MAGs from highly complex soil metagenomes.

This is part of my broader goal of machine-learning–guided multi-omics analysis.

🧬 Why soil metagenomes are painful (and exciting)

I’m working with soil metagenomes, which are notoriously challenging:

  • 1 g of soil can harbor millions of bacterial, archaeal, and fungal cells
  • Community richness and evenness are extremely high
  • Assemblies become highly fragmented with low N50
  • Short contigs make binning much harder

After exploring different strategies, I moved to a co-assembly approach:

  • Co-assembly with MEGAHIT of ~30 samples
  • Producing a single final.contigs.fa

(I will post separate details on MEGAHIT parameters.)

🧱 Classical binners: what failed and what worked

I initially tested the usual multi-binner approaches:

❌ CONCOCT

Likely failed because:

  • Assembly extremely large
  • Could not consistently segment into 10 kb fragments
  • Coverage patterns too noisy

❌ MaxBin2

Struggled due to:

  • Too many contigs
  • Sparse coverage across samples

⚠️ MetaBAT2 (worked partially)

  • Produced ~500 bins
  • Good, but clearly under-represented soil diversity
  • Bins were fragmented and incomplete

To do better, I needed something that:

  1. Scales to large, fragmented assemblies
  2. Leverages machine learning rather than fixed heuristics

🤖 Enter SemiBin2: machine-learning–guided binning

SemiBin2 integrates:

  1. Sequence composition (k-mers)
  2. Coverage profiles from multiple samples
  3. Deep-learning-based embeddings
  4. Density-based clustering for final bins

🔗 Repo: https://github.com/BigDataBiology/SemiBin

For complex and diverse soil data, this ML approach provides a huge improvement.

I automated everything through Snakemake so the full workflow runs with one command.

🔁 My SemiBin2 + Snakemake workflow (high-level)

Inputs

  • Co-assembly: final.contigs.fa
  • Paired-end reads: sample_R1_nodup.fq, sample_R2_nodup.fq

Workflow steps

  1. Build Bowtie2 index
  2. Map each sample → sorted BAM
  3. Create bam_list.txt
  4. SemiBin2 generate features
  5. SemiBin2 train_self (machine-learning step)
  6. SemiBin2 bin (produces bins)
  7. Unzip and convert .fa.gz → .fa
  8. MetaWRAP refinement
  9. dRep dereplication

All managed by Snakemake.

🚀 Running the workflow

Once the Snakefile is created:

snakemake -j 32

Snakemake will:

  • Detect required jobs
  • Build the DAG
  • Run everything in parallel
  • Restart only failed steps
  • Ensure reproducibility

📈 Results from this workflow

⭐ SemiBin2 initial output

6,563 bins generated from the co-assembly.

⭐ After refinement + dereplication

Using:

  • MetaWRAP bin refinement
  • dRep dereplication
  • ≥ 70% completeness
  • ≤ 5% contamination

I recovered:

🔥 ~1,000 high-quality MAGs

A major improvement over the ~500 bins from MetaBAT2 alone.

These high-quality MAGs will be used for:

  • Taxonomy
  • Functional annotation
  • CAZyme profiling
  • Energy metabolism markers
  • MG–MTX expression integration
  • Functional redundancy modeling (future post)

🔚 Wrap-up

This combined SemiBin2 + Snakemake workflow gave me:

  • A reproducible, automated MAG recovery pipeline
  • Better binning in extremely complex soil metagenomes
  • A strong foundation for multi-omics integration
  • ML-powered binning without manual tuning

💾 Full workflow on GitHub

I’ll maintain and update the full Snakemake pipeline here:

👉 https://github.com/jojyjohn28/semibin2-soil-mag-workflow