PHIEmbed (Phage-Host Interaction Prediction with Protein Embeddings)

PHIEmbed is a phage-host interaction prediction tool that uses protein language models to represent the receptor-binding proteins of phages. It presents improvements over using handcrafted (manually feature-engineered) sequence properties and eliminates the need to manually extract and select features from phage sequences.

Paper: https://doi.org/10.1371/journal.pone.0289030

If you find our work useful, please consider citing:

@article{10.1371/journal.pone.0289030,
    doi = {10.1371/journal.pone.0289030},
    author = {Gonzales, Mark Edward M. AND Ureta, Jennifer C. AND Shrestha, Anish M. S.},
    journal = {PLOS ONE},
    publisher = {Public Library of Science},
    title = {Protein embeddings improve phage-host interaction prediction},
    year = {2023},
    month = {07},
    volume = {18},
    url = {https://doi.org/10.1371/journal.pone.0289030},
    pages = {1-22},
    number = {7}
}

You can also find PHIEmbed on bio.tools.

Table of Contents

• 📰 News
• ♾️ Run on Google Colab
• 🚀 Installation & Usage
• 📚 Description
• 🧪 Reproducing Our Results
• 💻 Authors

📰 News

• 01 Sep 2024 - We created a Google Colab notebook for running our tool. Instructions here.

• 24 Apr 2024 - We added scripts to simplify running and training our tool. Instructions here.

• 23 Feb 2024 - We presented our work at the eAsia AMR Workshop 2024 held virtually and in person in Tokyo, Japan, and attended by antimicrobial resistance (AMR) researchers from Thailand, USA, Australia, Japan, and the Philippines. Slides here.

• 01 Dec 2023 - Presenting this work, the lead author (Mark Edward M. Gonzales) won 2nd Prize at the 2023 Magsaysay Future Engineers/Technologists Award. This award is conferred by the National Academy of Science and Technology, the highest recognition and scientific advisory body of the Philippines, to recognize outstanding research outputs on engineering and technology at the collegiate level. Presentation here (29:35–39:51) and slides here.

• 07 Jul 2023 - Our paper was accepted for publication in PLOS ONE.

↑ Return to Table of Contents.

♾️ Run on Google Colab

You can readily run PHIEmbed on Google Colab, without the need to install anything on your own computer: http://phiembed.bioinfodlsu.com

🚀 Installation & Usage

Operating System: Windows, Linux, or macOS

Clone the repository:

git clone https://github.com/bioinfodlsu/phage-host-prediction
cd phage-host-prediction

Create a virtual environment with all the necessary dependencies installed via Conda (we recommend using Miniconda):

conda env create -f environment.yaml

Activate this environment by running:

conda activate PHIEmbed

Running PHIEmbed

python3 phiembed.py --input <input_fasta> --model <model_joblib> --output <results_dir>

• Replace <input_fasta> with the path to the FASTA file containing the receptor-binding protein sequences. A sample FASTA file is provided here.
• Replace <model_joblib> with the path to the trained model (recognized format: joblib or compressed joblib, framework: scikit-learn). Download our trained model from this link. No need to uncompress, but doing so will speed up loading the model albeit at the cost of additional storage requirements. Refer to this guide for the list of accepted compressed formats.
• Replace <results_dir> with the path to the directory to which the results of running PHIEmbed will be written. The results of running PHIEmbed on the sample FASTA file are provided here.

The results for each protein are written to a CSV file (without a header row). Each row contains two comma-separated values: a host genus and the corresponding prediction score (class probability). The rows are sorted in order of decreasing prediction score. Hence, the first row pertains to the top-ranked prediction.

Under the hood, this script first converts each sequence into a protein embedding using ProtT5 (the top-performing protein language model based on our experiments) and then passes the embedding to a random forest classifier trained on our entire dataset. If your machine has a GPU, it will automatically be used to accelerate the protein embedding generation step.

Note: Running this script for the first time may take a few extra minutes since it involves downloading a model (ProtT5, around 2 GB) from Hugging Face.

Training PHIEmbed

python3 train.py --input <training_dataset>

• Replace <training_dataset> with the path to the training dataset. A sample can be downloaded here.
• The number of threads to be used for training can be specified using --threads. By default, it is set to -1 (that is, all threads are to be used).

The training dataset should be formatted as a CSV file (without a header row) where each row corresponds to a training sample. The first column is for the protein IDs, the second column is for the host genera, and the next 1,024 columns are for the components of the ProtT5 embeddings.

This script will output a gzip-compressed, serialized version of the trained model with filename phiembed_trained.joblib.gz.

↑ Return to Table of Contents.

📚 Description

Motivation: With the growing interest in using phages to combat antimicrobial resistance, computational methods for predicting phage-host interactions have been explored to help shortlist candidate phages. Most existing models consider entire proteomes and rely on manual feature engineering, which poses difficulty in selecting the most informative sequence properties to serve as input to the model.

Method: In this paper, we framed phage-host interaction prediction as a multiclass classification problem that takes as input the embeddings of a phage's receptor-binding proteins, which are known to be the key machinery for host recognition, and predicts the host genus. We explored different protein language models to automatically encode these protein sequences into dense embeddings without the need for additional alignment or structural information.

Results: We show that the use of embeddings of receptor-binding proteins presents improvements over handcrafted genomic and protein sequence features. The highest performance was obtained using the transformer-based protein language model ProtT5, resulting in a 3% to 4% increase in weighted F1 and recall scores across different prediction confidence thresholds, compared to using selected handcrafted sequence features.

↑ Return to Table of Contents.

🧪 Reproducing Our Results

Project Structure

The experiments folder contains the files and scripts for reproducing our results. Note that additional (large) files have to be downloaded (or generated) following the instructions in the Jupyter notebooks.

Click here to show/hide the list of directories, Jupyter notebooks, and Python scripts, as well as the folder structure.

Directories

Directory	Description
inphared	Contains the list of phage-host pairs in TSV format. The GenBank and FASTA files with the genomic and protein sequences of the phages, the embeddings of the receptor-binding proteins, and the phage-host-features CSV files should also be saved in this folder
preprocessing	Contains text files related to the preprocessing of host information and the selection of annotated RBPs
rbp_prediction	Contains the JSON file of the trained XGBoost model proposed by Boeckaerts et al. (2022) for the computational prediction of receptor-binding proteins. Downloaded from this repository (under the MIT License)
temp	Contains intermediate output files during preprocessing and performance evaluation

↑ Return to Table of Contents.

Jupyter Notebooks

Each notebook provides detailed instructions related to the required and output files, including the download links and where to save them.

Notebook	Description	Required Files	Output Files
1. Sequence Preprocessing.ipynb	Preprocessing of host information and selection of annotated receptor-binding proteins	GenomesDB (Partial. Complete populating following the instructions in the notebook), GenBank file of phage genomes and/or proteomes	FASTA files of genomic and protein sequences
2. Exploratory Data Analysis.ipynb	Exploratory data analysis	Protein embeddings (Part 1 and Part 2), Phage-host-features CSV files	–
3. RBP Computational Prediction.ipynb	Computational prediction of receptor-binding proteins	Protein embeddings (Part 1 and Part 2)	Protein embeddings (Part 1 and Part 2)
3.1. RBP FASTA Generation.ipynb	Generation of the FASTA files containing the RBP protein sequences	Protein embeddings (Part 1 and Part 2)	FASTA files of genomic and protein sequences
4. Protein Embedding Generation.ipynb	Generation of protein embeddings	FASTA files of genomic and protein sequences	Protein embeddings (Part 1 and Part 2)
5. Data Consolidation.ipynb	Generation of phage-host-features CSV files	FASTA files of genomic and protein sequences, Protein embeddings (Part 1 and Part 2)	Phage-host-features CSV files
6. Classifier Building & Evaluation.ipynb	Construction of phage-host interaction model and performance evaluation	Phage-host-features CSV files	Trained models
6.1. Additional Model Evaluation (Specificity + PR Curve).ipynb	Addition of metrics for model evaluation	Phage-host-features CSV files	–
7. Visualization.ipynb	Plotting of t-SNE and UMAP projections	Phage-host-features CSV files	–

↑ Return to Table of Contents.

Python Scripts

Script	Description
ClassificationUtil.py	Contains the utility functions for the generation of the phage-host-features CSV files, construction of the phage-host interaction model, and performance evaluation
ConstantsUtil.py	Contains the constants used in the notebooks and scripts
EDAUtil.py	Contains the utility functions for exploratory data analysis
RBPPredictionUtil.py	Contains the utility functions for the computational prediction of receptor-binding proteins
SequenceParsingUtil.py	Contains the utility functions for preprocessing host information and selecting annotated receptor-binding proteins
boeckaerts.py	Contains the utility functions written by Boeckaerts et al. (2021) for running their phage-host interaction prediction tool (with which we benchmarked our model). Downloaded from this repository (under the MIT License)

↑ Return to Table of Contents.

Folder Structure

Once you have cloned this repository and finished downloading (or generating) all the additional required files following the instructions in the Jupyter notebooks, your folder structure should be similar to the one below:

• phage-host-prediction (root)
  • datasets
    • inphared
      • inphared
        • GenomesDB (Downoad partial. Complete populating following the instructions here)
          • AB002632
          • ...
  • experiments
    • inphared
      • data (Download)
        • rbp.csv
        • rbp_embeddings_esm.csv
        • ...
      • embeddings (Download Part 1 and Part 2)
        • esm
        • esm1b
        • ...
      • fasta (Download)
        • complete
        • hypothetical
        • nucleotide
        • rbp
      • 16Sep2022_data_excluding_refseq.tsv
      • 16Sep2022_phages_downloaded_from_genbank.gb (Download)
    • models (Download)
      • boeckaerts.joblib
      • esm.joblib
      • ...
    • preprocessing
    • rbp_prediction
    • temp
    • 1. Sequence Preprocessing.ipynb
    • ...
    • ClassificationUtil.py
    • ...

↑ Return to Table of Contents.

Dependencies

Operating System: Windows, Linux, or macOS

Create a virtual environment with all the necessary dependencies installed via Conda (we recommend using Miniconda):

conda env create -f environment_experiments.yaml

Activate this environment by running:

conda activate PHIEmbed-experiments

Thanks to Dr. Paul K. Yu for sharing his environment configuration.

Click here to show/hide note on running the notebook for protein embedding generation.

The notebook 4. Protein Embedding Generation.ipynb has a dependency (bio_embeddings) that requires it to be run on Unix or a Unix-like operating system. If you are using Windows, consider using Windows Subsystem for Linux (WSL) or a virtual machine. We did not include bio_embeddings in environment_experiments.yaml to maintain cross-platform compatibility; you have to install it following the instructions here.

Moreover, generating protein embeddings should ideally be done on a machine with a GPU. The largest (and best-performing) protein language model that we used, ProtT5, consumes 5.9 GB of GPU memory. If your local machine does not have a GPU or if its GPU has insufficient memory, we recommend using a cloud GPU platform.

UPDATE (12 Jun 2023): In May 2023, Google Colab upgraded its Python runtime, resulting in compatibility issues with bio_embeddings. An alternative cloud GPU platform is Paperspace, which provides a PyTorch 1.12 runtime that is compatible with bio_embeddings.

Click here to show/hide the complete list of Python libraries and modules used in this project (excluding those that are part of the Python Standard Library).

Library/Module	Description	License
pyyaml	Supports standard YAML tags and provides Python-specific tags that allow to represent an arbitrary Python object	MIT License
jsonnet	Domain-specific language for JSON	Apache License 2.0
protobuf	Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data	BSD 3-Clause "New" or "Revised" License
regex	Provides additional functionality over the standard re module while maintaining backwards-compatibility	Apache License 2.0
nltk	Provides interfaces to corpora and lexical resources, along with a suite of text processing libraries for classification, tokenization, stemming, tagging, parsing, and semantic reasoning	Apache License 2.0
biopython	Provides tools for computational molecular biology	Biopython License Agreement, BSD 3-Clause License
ete3	Provides functions for automated manipulation, analysis, and visualization of phylogenetic trees	GNU General Public License v3.0
pandas	Provides functions for data analysis and manipulation	BSD 3-Clause "New" or "Revised" License
numpy	Provides a multidimensional array object, various derived objects, and an assortment of routines for fast operations on arrays	BSD 3-Clause "New" or "Revised" License
scipy	Provides efficient numerical routines, such as those for numerical integration, interpolation, optimization, linear algebra, and statistics	BSD 3-Clause "New" or "Revised" License
scikit-learn	Provides efficient tools for predictive data analysis	BSD 3-Clause "New" or "Revised" License
xgboost	Implements machine learning algorithms under the gradient boosting framework	Apache License 2.0
imbalanced-learn	Provides tools when dealing with classification with imbalanced classes	MIT License
joblib	Provides tools for lightweight pipelining in Python	BSD 3-Clause "New" or "Revised" License
cudatoolkit	Parallel computing platform and programming model for general computing on GPUs	NVIDIA Software License
bio_embeddings	Provides an interface for the use of language model-based biological sequence representations for transfer-learning	MIT License
torch	Optimized tensor library for deep learning using GPUs and CPUs	BSD 3-Clause "New" or "Revised" License
transformers	Provides pretrained models to perform tasks on different modalities such as text, vision, and audio	Apache License 2.0
sentencepiece	Unsupervised text tokenizer and detokenizer mainly for neural network-based text generation systems	Apache License 2.0
matplotlib	Provides functions for creating static, animated, and interactive visualizations	Matplotlib License (BSD-Compatible)
umap-learn	Implements uniform manifold approximation and projection, a dimensionality reduction technique	BSD 3-Clause "New" or "Revised" License

The descriptions are taken from their respective websites.

↑ Return to Table of Contents.

💻 Authors

• Mark Edward M. Gonzales
  gonzales.markedward@gmail.com

• Ms. Jennifer C. Ureta
  jennifer.ureta@gmail.com

• Dr. Anish M.S. Shrestha
  anish.shrestha@dlsu.edu.ph

This is a research project under the Bioinformatics Laboratory, Advanced Research Institute for Informatics, Computing and Networking, De La Salle University, Philippines.

This research was partly funded by the Department of Science and Technology – Philippine Council for Health Research and Development (DOST-PCHRD) under the e-Asia JRP 2021 Alternative therapeutics to tackle AMR pathogens (ATTACK-AMR) program. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.