depht 1.1.1

Discovery and Extraction of Phages Tool

Detection and Extraction of Phages Tool (DEPhT)

DEPhT is a new tool for identifying prophages in bacteria, and was developed with a particular interest in being able to rapidly scan hundreds to thousands of genomes and accurately extract complete (likely active) prophages from them.

A detailed manuscript has been submitted to Nucleic Acids Research, but in brief DEPhT works by using genome architecture (rather than homology) to identify genomic regions likely to contain a prophage. Any regions with phage-like architecture (characterized as regions with high gene density and few transcription direction changes) are then further scrutinized using two passes of homology detection. The first pass identifies genes on putative prophages that are homologs of (species/clade/genus-level) conserved bacterial genes, and uses any such genes to disrupt the prophage prediction. The second pass (disabled in the 'fast' runmode) identifies genes on putative prophages that are homologs of conserved, functionally annotated phage genes. Finally, prophage regions that got through the previous filters are subjected to a BLASTN-based attL/attR detection scheme that gives DEPhT better boundary detection than any tool we are aware of.

Installation

DEPhT runs natively on MacOS and Linux operating systems, and in theory should work on Windows using WSL.

Conda install

DEPhT has several dependencies, and as a result by far the easiest way to install it is to use Anaconda or the lightweight Miniconda with this single command:

conda create -n depht -c laa89 -c bioconda -c conda-forge depht -y

It may take up to a couple of minutes to complete.

Manual install

For users that would prefer to manage their own dependencies, you'll need to install each of the following:

• BLAST 2.9 or higher
• HHsuite3
• MMseqs2
• Prodigal
• Aragorn
• Python 3.6 or higher
• DEPhT~=1.1.0
• Python dependencies:
  • biopython~=1.78
  • bitarray~=2.0.0
  • bokeh~=2.2.2
  • dna-features-viewer~=3.0.3
  • kaleido~=0.2.1
  • matplotlib~=3.4.1
  • numpy~=1.20.2
  • pandas~=1.2.4
  • pretty-html-table~=0.9.10
  • plotly~=5.1.0
  • scipy~=1.7.0
  • phamclust~=0.1.0 (for model-training)
• ClustalO (for model-training)

Setup

DEPhT requires at least one genus-specific model to be installed before it will be able to run. At present, there are a few models available in our repository at the Open Science Framework. New models can be trained (instructions below), though this process is currently not very streamlined. We have a script nearly finished that should make the process much simpler, which should be available by late December 2021 or early January 2022.

Once a model has been downloaded (the easiest way is through a web browser), it needs to be decompressed and moved into a directory for DEPhT. For example, if you downloaded the Mycobacterium model:

if ! [[ -d ~/.depht/models ]]; then
    mkdir -p ~/.depht/models
fi

unzip ~/Downloads/Mycobacterium.zip -d ~/.depht/models/

Models trained using depht_train will be put in this directory by default. We are generally amenable to aiding in the construction of new models - the easiest way to accomplish this is by emailing either [e-mail] or [e-mail]. Note that some genera are better suited than others for DEPhT model creation.

Running DEPhT

Basics

After installation and setup, check that DEPhT can be run on the command line. NOTE: If you installed using conda, you'll need to activate your environment first (e.g. conda activate depht). Typing depht at the commandline should display something similar to the following (number of CPUs and models available will vary):

usage: depht [-h] [--model] [-c] [-n] [-m {fast,normal,strict}] [-s] [-d] [-v] [-t] [-p] [-l]
             infile [infile ...] outdir

DEPhT scans bacterial genomes looking for prophages. Regions identified as prophage 
candidates are further scrutinized, and attachment sites identified as accurately as 
possible before prophage extraction and generating the final report.

optional arguments:
  -h, --help            show this help message and exit
  --model {Mycobacterium}
                        which local model should be used [default: Mycobacterium]
  -c , --cpus           number of CPU cores to use [default: 4]
  -n, --no-draw         don't draw genome diagram for identified prophage(s)
  -m {fast,normal,sensitive}, --mode {fast,normal,sensitive}
                        select a runmode that favors speed or accuracy
  -s , --att_sens       sensitivity parameter for att site detection.
  -d, --dump-data       dump all support data to outdir
  -v, --verbose         print progress messages as the program runs
  -t , --tmp-dir        temporary directory to use for file I/O [default: ~/.depht/tmp]
  -p , --products       minimum number of phage homologs to report a prophage
  -l , --length         select a minimum length for prophages [default: 20000]

In order to run DEPhT, you will need to provide two arguments:

1. One or more genome sequences in either FASTA or Genbank flatfile format
2. A desired output directory

DEPhT will infer the input file type(s) when it parses the files, not using the file extensions. As far as we are aware, this makes DEPhT somewhat unusual among prophage-detection tools, as in a single run you can provide a set of files with multiple file formats. FASTA files will be treated as un-annotated and the sequences parsed from these input files will be auto-annotated prior to prophage detection. Genbank flatfiles will be treated as annotated genomes, and will therefore bypass the auto-annotation step and run ~20-30 seconds faster than their FASTA counterparts.

Run DEPhT on a single FASTA file like this (use your own file paths/extensions):

depht /path/to/my/sequence.fasta /path/to/my/output/directory

Run DEPhT on a directory of FASTA files like this:

depht /path/to/my/directory/*.fasta /path/to/my/output/directory

A large set of mixed FASTA (here using .fasta extension) and Genbank (here using .gbk extension) flatfiles can be run like this:

depht /path/to/my/directory/*.fasta /path/to/my/directory/*.gbk /path/to/my/output/directory

In theory, you're limited only by the number of files your Terminal will let you expand by using *.

For Mac users who are uncomfortable with entering paths at the commandline, modern versions of MacOS let you drag files from a Finder window into the Terminal and will automatically populate the path in the Terminal for you. Some Linux distributions may also support this kind of action.

In the event that a prophage region is discovered, or if the -d argument is specified, DEPhT will create a directory at the specified output directory for each of the input sequences. For those sequences that have predicted prophages, DEPhT will write an .html file with a visualization of the discovered prophage region(s). It will also output a FASTA (sequence) file and a Genbank (annotation) file for each extracted prophage sequence. See below for more details DEPhT's output files.

Progress updates during DEPhT's runtime can be toggled with -v.

depht /path/to/my/sequence.fasta /path/to/my/output/directory -v

The amount of resources (CPU cores) that DEPhT is allowed to utilize can be specified with -c. Note that some of DEPhT's dependencies utilize hyper-threading, so on most modern computers DEPhT will utilize 2 threads per specified CPU core.

depht /path/to/my/sequence.fasta /path/to/my/output/directory -c 6

Other Options

What follows is a description of DEPhT's optional arguments. These are described in isolation, but can be mixed and matched using different values to specifically tune the behavior of DEPhT to suit your needs. Default parameters were all set to optimize performance in Mycobacterium genomes.

Model Selection

DEPhT was originally designed for the precise and efficient discovery and extraction of Mycobacterium prophages, but can be adapted for other genera with the --model flag. See above for instructions to download models that we have already trained, and below for the list of currently available models.

If you have more than one model installed locally, you will need to tell DEPhT which model you'd like to use. Otherwise, it will choose one more-or-less at random, which may result in unexpectedly low-quality outputs.

depht /path/to/my/sequence.fasta /path/to/my/output/directory --model Pseudomonas

Runmode Selection

DEPhT has multiple runmodes, intended to serve as a dial tuning the trade-off between runtime and accuracy. The -m argument lets you select one of the available runmodes:

• fast: DEPhT discovers prophage regions as fast as possible using gene size and transcription direction changes. Regions are trimmed using the identified shell genome content of the selected genera, and an effort is made to identify attL/attR, but are likely not as accurate as in the other runmodes.
• normal: DEPhT discovers prophage regions as in fast mode, then tries to differentiate between active and defective prophages by identifying homologs of phage genes essential for viability.
• sensitive: DEPhT discovers prophage regions as in normal mode, and then tries to further differentiate between active and defective prophages by identifying homologs of phage genes with a consensus annotated function.

DEPhT will run in normal mode by default (e.g. if -m is not given), but if one is interested in getting an estimate of the number of prophages as quickly as possible, they may run DEPhT like this:

depht /path/to/my/sequence.fasta /path/to/my/output/directory -m fast

Alternatively, if one wants only the most likely prophages, with as many detailed functional annotations as possible, they might run:

depht /path/to/my/sequence.fasta /path/to/my/output/directory -m sensitive

Product Threshold

In normal and sensitive runmodes, DEPhT attempts to differentiate between active and defective prophages based on the number of identified prophage homologs in a region. This number of phage products can be raised or lowered by using the -p argument. In normal mode, the default value is 5; in sensitive mode, it is 10. If one feels that the default value is too high and would rather use 2 for example, this can be done by running:

depht /path/to/my/sequence.fasta /path/to/my/output/directory -p 2

Attachment Site Tuning

DEPhT employs a multi-feature scoring algorithm and a library of reference sequences to determine the best possible attachment site core (or if there is no appropriate sequence). The runtime of this component is heavily influenced by the runtime of the BLASTN algorithm, so runtime scales with the amount of sequence that is searched. However, the precision of extraction may benefit from searching a larger sequence space, particularly in genera where few high-quality reference sequences are available. The sequence space that is searched for an attachment site core can be controlled by using the -s flag, which acts as a multiplier against 5000 bp. By default, DEPhT uses -s 7, which corresponds to a search space of up to 7 x 5,000 = 35,000 bp at the left and right ends of each identified prophage. This can be raised to 50,000 bp by setting -s to 10, at the expense of some additional runtime:

depht /path/to/my/sequence.fasta /path/to/my/output/directory -s 10

Prophage Size

DEPhT mandates a minimum length for prophage regions reported for output quality assurance. This minimum length threshold can be lowered or raised with the -l flag, and is set at 20,000 base pairs by default - just over half the length of the shortest known Mycobacterium prophage. Reduce this threshold to 10,000 bases like this:

depht /path/to/my/sequence.fasta /path/to/my/output/directory -l 10000

Temporary Directory

DEPhT utilizes various software that require outputs and data intermediates that are written to files. These files are stored in a temporary directory, and removed once DEPhT finishes running. By default, DEPhT will use ~/.depht/tmp, but can use any other directory that your user account has read/write permissions in, by using the -t argument.

depht /path/to/my/sequence.fasta /path/to/my/output/directory -t /path/to/temporary/directory

Output

DEPhT's output consists of three main files:

1. An .html file with a visualization of the discovered prophage regions
2. A .csv spreadsheet with the primary data used to discern prophage regions - one file per contig
3. A .gbk Genbank flatfile with DEPhT's annotation of the inputted sequence - one file per contig

DEPhT's graphical .html output displays a cirular input genome map and linear phage region genome map with DnaFeaturesViewer as well as the coordinates of the regions discovered in a colored table with pretty-html-table.

In each of these genome maps and coordinate tables, prophage and/or protein-coding sequence features are colored green for forward-oriented features and colored red for reverse-oriented features. Above those prophage features in the circular genome map is annotated the prophage region name as given by DEPhT. Above those protein-coding features in the linear genome map(s) is annotated phage products as identified by DEPhT.

DEPhT's data .csv output contains data for each protein-coding feature in the inputted sequence file.

The columns in this output are the following:

• Gene ID: A protein-coding feature ID assigned by DEPhT
• Start: The start coordinate of a feature in the input sequence
• End: The end coodinate of a feature in the input sequence
• Prediction: The probability of a feature belonging to a prophage as analyzed by DEPhT
• Bacterial Homology: The identity of a feature as shell genome content as analyzed by DEPhT
• Phage Homology: The probability given by an alignment of a feature to a HMM profile of phage amino acid sequences

Training New Models

Models can be trained using the depht_train package, installed as part of DEPhT.

What follows will describe the workflow for training new models, as well as explain the thought process.

Selection of Training Genomes

This is by far the highest hurdle for training new models. The better the training genomes are selected, the better the model will perform. We highly recommend only training against completely sequenced bacteria and manually annotated phages.

There's an important tradeoff you'll need to make when training models: volume of data versus quality of data. A relatively small dataset (~100 phages and 30-45 bacteria) can yield incredibly high-quality models if the genomes are chosen well and especially if the phage genomes are well-annotated. Assuming all the training data is high-quality, increasing the amount of training data will likely improve the quality of predictions made by DEPhT, with the caveat that larger models will necessarily increase the DEPhT runtime, which will be most noticeable in the fast runmode.

Ok so let's suppose you want to train a new model for Mycobacteria. A good start would be to head to PATRIC and navigate to the Mycobacteriaceae.

Retrieve Bacterial Genomes

In the taxonomy tree, the steps to get here are:

Terrabacteria group >> Actinobacteria >> Actinomycetia >> Corynebacteriales >> Mycobacteriaceae

The red box below shows where to click to get to the home page for the family or genus of interest.

From there, navigate to the "Genomes" tab to see all the available genomes in the chosen taxon. Click "Filters", and a good choice might be to select only those genomes where "Genome Status" is "Complete", and "Reference Genome" is either "Representative" or "Reference", and "Genome Quality" is "Good". Hit "Apply" to apply those filters. You can download FASTA files for these genomes by selecting all the genomes in the table, and clicking the "DWNLD" button.

Click "More Options", and in the popup dialog box, check the box next to "Genomic Sequences in FASTA (*.fna)" before pressing "Download".

Of course you are free to add any additional genomes you'd like to better populate the spectrum of diversity in the genus. In our case, we added several Mycobacterium abscessus strains to fill in the so-called Mycobacterium abscessus complex (MAC).

Check Bacteria for Prophages

Ideally, you'll run these genomes through PHASTER or some other prophage prediction tool to get the approximate coordinates of any complete prophages in these strains, and recording them in a CSV file that you'll pass to the training module. The coordinates don't have to be perfect, though the better they are the better the resultant model will perform. This step will reduce the probability that DEPhT treats a prophage found in multiple strains as "conserved bacterial genes", and also give the model an idea what integrated prophages are supposed to look, as opposed to only knowing what extracted phages/prophages look like.

Retrieve Phage Genomes

Lastly, you'll need to retrieve functionally annotated phages from Genbank or elsewhere. Like the bacteria, it's important that these phages represent the spectrum of diversity of phages infecting hosts in the genus. Ideally there will also be clusters of at least somewhat-related phages in this dataset.

Running the Training Pipeline

The training workflow is available as a single pipeline. The only required arguments are:

1. a name for the new model
2. path to a directory containing functionally annotated phage genomes for the genus of interest
3. path to a directory containing bacterial genomes for the genus of interest

Run the pipeline like this:

depht_train create_model model_name /path/to/annotated/phage/genomes /path/to/bacterial/genomes

If you're trying to create a new model with the same name as an existing one, depht_train will not overwrite the existing model by default, but it will then force you to pick a new name. If you'd like to overwrite the existing model, you can do so with the -f/--force argument:

depht_train create_model model_name /path/to/annotated/phage/genomes /path/to/bacterial/genomes -f

If one or more of your bacterial genomes has one or more known (or probable) prophage(s) in it, you can provide a CSV file formatted as above, using the --prophage-coords argument:

depht_train create_model model_name /path/to/annotated/phage/genomes /path/to/bacterial/genomes --prophage-coords /path/to/prophage_coords.csv

Training a model consists of several computationally expensive steps, and as such the amount of time it takes to train a model is highly variable, but generally influenced in these ways:

1. more genomes --> longer training time (and likely depht runtime)
2. more CPU cores --> shorter training time

Most new models will likely take somewhere between 15 minutes and an hour to train.

General Information

• Current version is 1.1.0
• Most recent stable version is 1.1.0
• We currently have models available for these bacterial genera:
  • Mycobacterium
  • Gordonia
  • Pseudomonas
• If you'd like to contact us about expanding this set, please email either [e-mail] or [e-mail]