SemiBin 2.1.0

Metagenomic binning with siamese neural networks

SemiBin: Metagenomic Binning Using Siamese Neural Networks for short and long reads

SemiBin is a command tool for metagenomic binning with deep learning, handles both short and long reads.

CONTACT US: Please use GitHub issues for bug reports and the SemiBin users mailing-list for more open-ended discussions or questions.

If you use this software in a publication please cite:

Pan, S.; Zhu, C.; Zhao, XM.; Coelho, LP. A deep siamese neural network improves metagenome-assembled genomes in microbiome datasets across different environments. Nat Commun 13, 2326 (2022). https://doi.org/10.1038/s41467-022-29843-y

The self-supervised approach and the algorithms used for long-read datasets (as well as their benchmarking) are described in

Pan, S.; Zhao, XM; Coelho, LP. SemiBin2: self-supervised contrastive learning leads to better MAGs for short- and long-read sequencing. Bioinformatics Volume 39, Issue Supplement_1, June 2023, Pages i21–i29; https://doi.org/10.1093/bioinformatics/btad209

Basic usage of SemiBin

A tutorial of running SemiBin from scrath can be found here SemiBin tutorial.

Installation:

conda create -n SemiBin
conda activate SemiBin
conda install -c conda-forge -c bioconda semibin

This will install both the SemiBin2 command as well (for backwards compatibility), the old SemiBin command. For new projects, it is recommended that you exclusively use SemiBin2: both commands do the same thing, but SemiBin2 has a slightly nicer interface.

The inputs to the SemiBin are contigs (assembled from the reads) and BAM files (reads mapping to the contigs). In the docs you can see how to generate the inputs starting with a metagenome.

Running with single-sample binning (for example: human gut samples):

SemiBin2 single_easy_bin -i contig.fa -b S1.sorted.bam -o output --environment human_gut

(if you are using contigs from long-reads, add the --sequencing-type=long_read argument).

Running with multi-sample binning:

SemiBin2 multi_easy_bin -i contig_whole.fa -b *.sorted.bam -o output

The output includes the bins in the output_bins directory (including the bin.*.fa and recluster.*.fa).

Please find more options and details below and read the docs.

Advanced Installation

SemiBin runs (and is continuously tested) on Python 3.7-3.12

Bioconda

The simplest mode is shown above. However, if you want to use SemiBin with GPU (which is faster if you have one available), you need to install PyTorch with GPU support:

conda create -n SemiBin
conda activate SemiBin
conda install -c conda-forge -c bioconda semibin
conda install -c pytorch -c nvidia pytorch pytorch-cuda=11.8

MacOS note: you can only install the CPU version of PyTorch in MacOS with conda and you need to install from source to take advantage of a GPU (see #72). For more information on how to install PyTorch, see their documentation.

Source

You will need the following dependencies:

• Bedtools, Hmmer
• Samtools
• HMMER

The easiest way to install the dependencies is with conda:

conda install -c bioconda bedtools hmmer samtools

Once the dependencies are installed, you can install SemiBin by running:

python setup.py install

Optional extra dependencies for running SemiBin1:

• MMseqs2
• Prodigal
• FragGeneScan

Examples of binning

SemiBin runs on single-sample, co-assembly and multi-sample binning. Here we show the simple modes as an example. For the details and examples of every SemiBin subcommand, please read the docs.

Binning assemblies from long reads

Since version 1.4, SemiBin proposes new algorithm (ensemble based DBSCAN algorithm) for binning assemblies from long reads. To use it, you can used the subcommands bin_long or pass the option --sequencing-type=long_read to the single_easy_bin or multi_easy_bin subcommands.

Self-supervised mode

Since version 1.3, SemiBin supports completely self-supervised learning, which bypasses the need to annotate contigs with MMSeqs2. In benchmarks, self-supervised learning is both faster (4x faster; using only 11% of RAM at peak) and generates 8.3-21.5% more high-quality bins compared to the version tested in the manuscript To use it, pass the option --training-mode=self to the single_easy_bin or multi_easy_bin subcommands.

Easy single/co-assembly binning mode

Single sample and co-assembly are handled the same way by SemiBin.

You will need the following inputs:

1. A contig file (contig.fa in the example below)
2. BAM file(s) from mapping short reads to the contigs, sorted (mapped_reads.sorted.bam in the example below)

The single_easy_bin command can be used to produce results in a single step.

For example:

SemiBin2 \
    single_easy_bin \
    --input-fasta contig.fa \
    --input-bam mapped_reads.sorted.bam \
    --environment human_gut \
    --output output

Alternatively, you can train a new model for that sample, by not passing in the --environment flag:

SemiBin2 \
    single_easy_bin \
    --input-fasta contig.fa \
    --input-bam mapped_reads.sorted.bam \
    --output output

The following environments are supported:

• human_gut
• dog_gut
• ocean
• soil
• cat_gut
• human_oral
• mouse_gut
• pig_gut
• built_environment
• wastewater
• chicken_caecum (Contributed by Florian Plaza Oñate)
• global

The global environment can be used if none of the others is appropriate. Note that training a new model can take a lot of time and disk space. Some patience will be required. If you have a lot of samples from the same environment, you can also train a new model from them and reuse it.

Easy multi-samples binning mode

The multi_easy_bin command can be used in multi-samples binning mode:

You will need the following inputs:

1. A combined contig file
2. BAM files from mapping

For every contig, format of the name is <sample_name>:<contig_name>, where : is the default separator (it can be changed with the --separator argument). NOTE: Make sure the sample names are unique and the separator does not introduce confusion when splitting. For example:

>S1:Contig_1
AGATAATAAAGATAATAATA
>S1:Contig_2
CGAATTTATCTCAAGAACAAGAAAA
>S1:Contig_3
AAAAAGAGAAAATTCAGAATTAGCCAATAAAATA
>S2:Contig_1
AATGATATAATACTTAATA
>S2:Contig_2
AAAATATTAAAGAAATAATGAAAGAAA
>S3:Contig_1
ATAAAGACGATAAAATAATAAAAGCCAAATCCGACAAAGAAAGAACGG
>S3:Contig_2
AATATTTTAGAGAAAGACATAAACAATAAGAAAAGTATT
>S3:Contig_3
CAAATACGAATGATTCTTTATTAGATTATCTTAATAAGAATATC

You can use this to get the combined contig:

SemiBin2 concatenate_fasta -i contig*.fa -o output

If either the sample or the contig names use the default separator (:), you will need to change it with the --separator,-s argument.

After mapping samples (individually) to the combined FASTA file, you can get the results with one line of code:

SemiBin2 multi_easy_bin -i concatenated.fa -b *.sorted.bam -o output

Running with abundance information from strobealign-aemb

Strobealign-aemb is a fast abundance estimation method for metagenomic binning. As strobealign-aemb can not provide the mapping information for every position of the contig, so we can not run SemiBin2 with strobealign-aemb in binning modes where samples used smaller 5 and need to split the contigs to generate the must-link constratints.

1. split the fasta files

python script/generate_split.py -c contig.fa -o output

2. map reads using strobealign-aemb to generate the abundance information

strobealign --aemb output/split.fa read1_1.fq read1_2.fq -R 6 > sample1.txt
strobealign --aemb output/split.fa read2_1.fq read2_2.fq -R 6 > sample2.txt
strobealign --aemb output/split.fa read3_1.fq read3_2.fq -R 6 > sample3.txt
strobealign --aemb output/split.fa read4_1.fq read4_2.fq -R 6 > sample4.txt
strobealign --aemb output/split.fa read5_1.fq read5_2.fq -R 6 > sample5.txt

3. Running SemiBin2 (like running SemiBin with BAM files)

SemiBin2 generate_sequence_features_single -i contig.fa -a *.txt -o output
SemiBin2 generate_sequence_features_multi -i contig.fa -a *.txt -s : -o output
SemiBin2 single_easy_bin -i contig.fa -a *.txt -o output
SemiBin2 multi_easy_bin i contig.fa -a *.txt -s : -o output

Output

The output folder will contain:

1. Features computed from the data and used for training and clustering
2. Saved semi-supervised deep learning model
3. Output bins
4. Table with basic information about each bin
5. Some intermediate files

By default, reconstructed bins are in output_recluster_bins directory.

For more details about the output, read the docs.