PheMaLe bacteria

Phenotype Machine Learning bacteria - A toolkit for bacterial phenotype prediction

Bacterial phenotypes prediction using high level features and multiple machine learning algorithms

Metagenomics and sequencing techniques have greatly improved in these last five years and, as a consequence, the amount of data from microbial communities is astronomic. An import part of the microbial community are phages, which have their own ecological roles in the environment. Besides that, they have also been given a possible human relevant (clinical) role as terminators of multidrug resistant bacterial infections. A lot of basic research still need to be done in the Phage therapy field, and part of this research involves gathering knowledge from new phages present in the environment as well as about their relationship with clinical relevant bacterial pathogens.

Having this scenario in mind, we have developed vHULK. A user-friendly tool for prediction of phage hosts given their complete or partial genome in FASTA format. Our tool outputs an ensemble prediction at the genus or species level based on scores of four different neural network models. Each model was trained with more than 4,000 genomes whose phage-host relationship was known. v.HULK also outputs a mesure of entropy for each final prediction, which we have demonstrated to be correlated with prediction's accuracy. The user might understand this value as additional information of how certain v.HULK is about a particular prediction. We also suspect that phages with higher entropy values may have a broad host-range. But that hypothesis is to be tested later. Accuracy results in test datasets were >99% for predictions at the genus level and >98% at the species level. vHULK currently supports predictions for 52 different prokaryotic host species and 61 different genera.

Target species

Acinetobacter_baumannii, Aeromonas_salmonicida, Arthrobacter_globiformis, Bacillus_cereus, Bacillus_megaterium, Bacillus_pumilus, Bacillus_subtilis, Bacillus_thuringiensis, Brucella_abortus, Campylobacter_jejuni, Caulobacter_crescentus, Cellulophaga_baltica, Citrobacter_freundii, Clostridium_difficile, Clostridium_perfringens, Cronobacter_sakazakii, Enterococcus_faecalis, Erwinia_amylovora, Escherichia_coli, Gordonia_terrae, Helicobacter_pylori, Klebsiella_pneumoniae, Lactococcus_lactis, Listeria_monocytogenes, Microbacterium_foliorum, Moraxella_catarrhalis, Mycobacterium_smegmatis, Paenibacillus_larvae, Pectobacterium_atrosepticum, Propionibacterium_acnes, Proteus_mirabilis, Pseudomonas_aeruginosa, Pseudomonas_fluorescens, Pseudomonas_syringae, Ralstonia_solanacearum, Rhodococcus_erythropolis, Salmonella_enterica, Salmonella_typhimurium, Shigella_flexneri, Shigella_sonnei, Staphylococcus_aureus, Staphylococcus_epidermidis, Streptococcus_pneumoniae, Streptococcus_thermophilus, Streptomyces_griseus, Streptomyces_lividans, Streptomyces_venezuelae, Sulfolobus_islandicus, Vibrio_anguillarum, Vibrio_cholerae, Vibrio_parahaemolyticus, Yersinia_enterocolitica

Target genera

Achromobacter, Acidianus, Acinetobacter, Aeromonas, Arthrobacter, Bacillus, Brevibacillus, Brucella, Burkholderia, Campylobacter, Caulobacter, Cellulophaga, Citrobacter, Clostridium, Corynebacterium, Cronobacter, Dickeya, Enterobacter, Enterococcus, Erwinia, Escherichia, Flavobacterium, Gordonia, Haloarcula, Halorubrum, Helicobacter, Klebsiella, Lactobacillus, Lactococcus, Leuconostoc, Listeria, Mannheimia, Microbacterium, Moraxella, Mycobacterium, Paenibacillus, Pectobacterium, Prochlorococcus, Propionibacterium, Proteus, Pseudoalteromonas, Pseudomonas, Ralstonia, Rhizobium, Rhodococcus, Ruegeria, Salmonella, Serratia, Shigella, Sinorhizobium, Staphylococcus, Stenotrophomonas, Streptococcus, Streptomyces, Sulfolobus, Synechococcus, Thermus, Vibrio, Xanthomonas, Xylella, Yersinia

Scripts

Main script:

• vHULK.py - Deep learning host prediction from phage genomes

Dependencies

All scripts from this project were coded in Python 3.7. So, first of all, make sure you have it installed and updated.

vHULK's main script (vHULK.py) requires Prokka and HMMER tools as dependencies.

• Prokka - Rapid Prokaryotic genome annotation.
• HMMER Tools - Biosequence analysis using profile hidden Markov models

These python libraries are required:

• Numpy - Fast calculations with numerical data
• Pandas - Data management
• Scipy - Efficient cientific computing with arrays
• Biopython - Handling biological sequences
• Tensorflow > 2.8 - Google's Neural Networks library

To install these dependencies, there are usually two ways: pip or conda.
We strongly recommend creating a specific conda environment containing the installed libraries and tools.

The $ signifies the command-line prompt

Install all requirements using these commands:

$ conda create -n vHULK -c conda-forge -c bioconda -c defaults prokka
$ conda install -n vHULK -c bioconda hmmer
$ conda install -n vHULK -c bioconda -c anaconda numpy pandas scipy biopython tensorflow=2.8.2
$ conda activate vHULK 

Note

Some people have been facing issues with newer versions of Prokka or TensorFlow. To avoid such problems, we froze a working version with TensorFlow 2.8.2 we know for sure it works with vHULK.

If all went well your command-line prompt will now be prefixed with the environment name you gave above.

(vHULK)$ python -V
Python 3.7.X

Installation

Getting vHULK ready to run is as simple as cloning this Github project or download and extract it to a directory inside your computer:

$ git clone https://github.com/LaboratorioBioinformatica/vHULK

Usage

The vHULK.py is an executable script. Its location can be included in your $PATH environmet variable if you so desire.

For example, assuming you have cloned this repo in /home/user/tools/vHULK, you can prepend this location to $PATH

$ export PATH="/home/user/tools/vHULK:$PATH"

This allows you to invoke vHULK.py from any location on your system.

Alternatively, all invocations must point to its full location e.g.

(vHULK)$ /full/path/to/vHULK.py -h

Here, it is assumed that it is available in your $PATH.

In the first use it will download the databases to a directory called database, which stores all necessary data. You can move the whole directory in a location of your preference and point to it with the -m option of vHULK.py when running vHULK. By default, vHULK searches for it in the same path where the executable is located.

To list all options available run

(vHULK)$ vHULK.py -h
usage: vHULK.py [-h] -i INPUT_DIR -o OUTPUT_DIR [-t THREADS] [-m MODELS_DIR]
                [-f FILES_DIR] [--all] [-v]

Predict phage draft genomes in metagenomic bins.

Required arguments:
  -i INPUT_DIR, --input-dir INPUT_DIR
                        Path to a folder containing metagenomic bins in .fa or
                        .fasta format (default: )
Optional arguments:
  -h, --help            show this help message and exit
  -o OUTPUT_DIR, --output-dir OUTPUT_DIR
                        Location to store results in. It is created if it
                        doesn't exist (default: /home/user/tools/vHULK/results)
  -t THREADS, --threads THREADS
                        Number of CPU threads to be used by Prokka and hmmscan
                        (default: 1)
  -m MODELS_DIR, --models-dir MODELS_DIR
                        Path to directory where all models are stored.
                        (default: /home/user/tools/vHULK/models)
  -f FILES_DIR, --files-dir FILES_DIR
                        Files directory provided with vHULK (default:
                        /home/user/tools/Projects/vHULK/files)
  --all                 Write predictions for all input bins/genomes, even if
                        they were skipped (size filtered or hmmscan failed)
                        (default: False)
  -v, --version         show program's version number and exit

Running the example datasets

We provide a folder with example datasets containing mocking bins of RefSeq viral and bacterial genomes.

To try these examples, run:

(vHULK)$ vHULK.py -i test_input -o test_output -t 4

It should take about 2 to 3 minutes to generate one prediction.

Input

vHULK is ready to accept whole or partial phage genomes. Keep in mind that vHULK's predictions are based in high level annotated features, i.e., features that depend on gene annotation. So, very small contigs with few or no entire genes will not provide features for vHULK to work with. In general, if you have a partial genome larger than 10 kbp you should be fine. Currently, vHULK skips all genomes or bins that are smaller than 5000bp at runtime.

Only FASTA nucleotide files are accepted, and vHULK will understand each individual file as one phage genome. Therefore, single-sequence FASTA files will be understood as whole or partial phage genomes and multiFASTA as a metagenomic BIN. Our tests indicate that prediction's accuracy is not affected in fragmented genomes (BINs).

Output

vHULK's main output file is 'prediction_[name of your file].csv' inside the output directory. If more than one fasta file is inside the input_folder, vHULK will understand and make predictions for each file.

Output example:

pred_genus,score_genus,entropy_genus,energy_genus,pred_species,score_species,entropy_species,energy_species
Agrobacterium,0.9385931,0.9803923149327602,1.4062266998288386,Aeromonas_hydrophila,0.9914063,715.14189738594,-7.647708205200186

Information in the output file.

• 1st to 4th columns: Genus host prediction and respective score, entropy and energy for the prediction
• 5th to 8th columns: Species host prediction and respective score, entropy and energy for the prediction

Entropy value and energy value can be used as a proxy of confidence for a particular prediction, i.e., values close to 0 correlate with higher accuracies and above 2 correlate with dubious accuracy.

The output directory also contains two more subdirectories holding the intermediate results of prokka and hmmscan. These are named accordingly.