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 with conda:

bash 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):

bash 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:

bash 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 3.7-3.13

pixi

The current recommended way to install SemiBin with GPU-support is to use pixi. Pixi will use the packages from conda-forge and bioconda to install SemiBin and its dependencies. See the docs for more details, but the basic idea is to create a pixi.toml file with the following content:

```toml [project] authors = ["Luis Pedro Coelho luis@luispedro.org"] channels = ["conda-forge", "bioconda"] name = "semibin_install" platforms = ["linux-64"] version = "0.1.0"

[tasks]

[dependencies] semibin = ">=2.2.0,<3" pytorch-gpu = "*"

[system-requirements] cuda = "12.0" ```

This will install SemiBin with GPU support, but it does require a CUDA-compatible GPU. Alternatively, you can install SemiBin in CPU-only mode by removing the pytorch-gpu and cuda lines.

Source

You will need the following dependencies:

• Bedtools, Hmmer
• Samtools
• HMMER

The easiest way to install the dependencies is with conda:

bash conda install -c bioconda bedtools hmmer samtools

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

bash pip install .

Optional extra dependencies:

• MMseqs2
• Prodigal

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.

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:

bash 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:

bash 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:Contig1 AGATAATAAAGATAATAATA S1:Contig2 CGAATTTATCTCAAGAACAAGAAAA S1:Contig3 AAAAAGAGAAAATTCAGAATTAGCCAATAAAATA S2:Contig1 AATGATATAATACTTAATA S2:Contig2 AAAATATTAAAGAAATAATGAAAGAAA S3:Contig1 ATAAAGACGATAAAATAATAAAAGCCAAATCCGACAAAGAAAGAACGG S3:Contig2 AATATTTTAGAGAAAGACATAAACAATAAGAAAAGTATT S3:Contig3 CAAATACGAATGATTCTTTATTAGATTATCTTAATAAGAATATC ```

You can use this to get the combined contig:

bash 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:

bash 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 to generate the must-link constraints bash python script/generate_split.py -c contig.fa -o output
2. map reads using strobealign-aemb to generate the abundance information bash 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) bash 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, bins are in output_bins directory.

For more details about the output, read the docs.