kiharalab/domain-pfp

Domain-PFP is a self-supervised method to predict protein functions from the domains

Domain-PFP

Domain-PFP is a self-supervised method to learn functional representations of protein domains that can be used for protein function prediction.

License: GPL v3. (If you are interested in a different license, for example, for commercial use, please contact us.)

Contact: Daisuke Kihara ([email protected])

For technical problems or questions, please reach to Nabil Ibtehaz ([email protected]).

Citation:

Ibtehaz, N., Kagaya, Y. & Kihara, D. Domain-PFP allows protein function prediction using function-aware domain embedding representations. Commun Biol 6, 1103 (2023). https://doi.org/10.1038/s42003-023-05476-9

Online Platform (run easily and freely on Google Colab)

https://bit.ly/domain-pfp-colab

Introduction

Domains are functional and structural units of proteins that govern various biological functions performed by the proteins. Therefore, characterization of domains in a protein can serve as a proper functional representation of proteins. Here, we employ a self-supervised protocol to derive functionally consistent representations for domains, through learning domain-Gene Ontology (GO) co-occurrences and associations. Domain embeddings constructed with the self-supervised protocol learned functional associations, which turned out effective to perform in actual function prediction tasks. An extensive evaluation shows that the protein representation using the domain embeddings are superior to that of large-scale protein language models in GO prediction tasks. Moreover, the new function prediction method, Domain-PFP, significantly outperformed the state-of-the-art function predictors. Notably, Domain-PFP achieved increase of area under precision-recall curve by 2.43%, 14.58% and 9.57% over the state-of-the-art method for molecular function (MF), biological process (BP) and cellular components (CC), respectively. Moreover, Domain-PFP demonstrated competitive performance in CAFA3 evaluation, by achieving overall the best performance among top teams that participated in the assessment.

Overall Protocol

Overview of Domain-PFP.

1. The network architecture used for self-supervised learning of domain embeddings.
2. The overall pipeline of learning the functionally aware domain embeddings.
3. The steps of computing the embeddings of a protein and inferring the functions.

Pre-required software

Python 3.9 : https://www.python.org/downloads/

Installation

1. Install git

2. Clone the repository in your computer

git clone https://github.com/kiharalab/Domain-PFP && cd Domain-PFP

3. Build dependencies.

You have two options to install dependency on your computer:

3.1 Install with pip and python.

3.1.1install pip.

3.1.2 Install dependency in command line.

pip3 install -r requirements.txt --user

If you encounter any errors, you can install each library one by one:

!pip3 install numpy==1.23.5
!pip3 install tqdm==4.64.1
!pip3 install scipy==1.9.3
!pip3 install matplotlib==3.6.2
!pip3 install matplotlib-inline==0.1.6
!pip3 install pandas==1.5.2
!pip3 install seaborn==0.12.1
!pip3 install torch==1.13.0
!pip3 install tabulate==0.9.0
!pip3 install scikit-learn==1.2.0
!pip3 install click==8.0.3

Installing the dependencies only require a few minutes on a standard desktop computer.

3.2 Install with anaconda

3.2.1 install conda.

3.2.2 Install dependency in command line

conda create -n domainpfp python=3.9
conda activate domainpfp
pip3 install -r requirements.txt 

Each time when you want to run this code, simply activate the environment by

conda activate domainpfp
conda deactivate    (If you want to exit) 

Prepare Data

Please download and unzip the data.zip and saved_models.zip files. Optinally, you may download our blast and ppi database (blast_ppi_database.zip) if you wish to use blast or ppi in your prediction.

https://kiharalab.org/domainpfp/

wget https://kiharalab.org/domainpfp/data.zip
unzip data.zip
wget https://kiharalab.org/domainpfp/saved_models.zip
unzip saved_models.zip
wget https://kiharalab.org/domainpfp/blast_ppi_database.zip
unzip blast_ppi_database.zip

Source Codes

Our implementation of Domain-PFP is provided in the DomainPFP directory.

Experiments and Reproducibility

All the codes to run the experiments presented in the paper, are provided in the /experiments directory.

Benchmark Results

The result files of CAFA3 and PROBE benchmarks, generated using the official evaluation tool, are provided in the /results directory.

Usage

Here we provide the following functionalities :

1. calculate domain-GO association probabilities

You can use DomainGO_prob to calculate the association probability of a domain and GO term, by providing the domain and GO term

python3 domaingo_prob.py:

  -domain              input InterPro domain
  -GO                  input GO term

Example

python3 domaingo_prob.py --domain IPR000003 --GO GO:0006355

This usually takes <2 minutes to run.

2. compute functionally aware protein embedding representation

You can use Domain-PFP to compute functionally aware embedding representation of a protein by providing the protein ID or path to a fasta file. You also need to provide the path to the savefile, where the embedding will be saved as a pickle file

python3 compute_embeddings.py:

  -protein              UniProt ID of protein
  -fasta                Or provide the fasta file path
  -savefile             Path to save the protein embeddings (as pickle file)
                        (default: emb.p)  

Example

python3 compute_embeddings.py --protein Q6NYN7 --savefile emb_Q6NYN7.p

This usually takes <5 minutes to run, depending on the availability of InterProScan server.

Note: If you wish to use this representation as feature for some functionally relevant downstream task.
Please consider applying proper normalization

3. predict protein functions using Domain-PFP

You can use Domain-PFP to predict the functions by either providing the protein ID or path to a fasta file.

python3 predict_functions.py:
  --protein               UniProt ID of protein
  --fasta                 Or provide the fasta file path
  --threshMFO             Threshold for MFO prediction (default: 0.36)
  --threshBPO             Threshold for BPO prediction (default: 0.31)
  --threshCCO             Threshold for CCO prediction (default: 0.36)
  --blast_flag            Optional flag to use DiamondBlast for function prediction
                          (DiamondBlast needs to be installed and assigned to path)
  --diamond_path          Path to Diamond Blast (by default the colab release path is provided)
                          (default='/content/Domain-PFP/diamond')
  --ppi_flag              Optional flag to use String PPI for function prediction
                          (Only works for Uniprot IDs or properly formatted fastas)
  --outfile               Path to the output csv file (optional)
  

Example

python3 predict_functions.py --protein Q6NYN7 --outfile sample_functions/Q6NYN7_functions.csv

python3 predict_functions.py --fasta sample_protein/Q6NYN7.fasta --outfile sample_functions/Q6NYN7_functions.csv

python3 predict_functions.py --fasta sample_protein/Q6NYN7.fasta --threshCCO 0.5 --outfile sample_functions/Q6NYN7_functions.csv

python3 predict_functions.py --fasta sample_protein/Q6NYN7.fasta --threshCCO 0.5 --outfile sample_functions/Q6NYN7_functions.csv --blast_flag --ppi_flag

This usually takes <5 minutes to run, depending on the availability of InterProScan server.

(Note: we recommend using our google colab release https://bit.ly/domain-pfp-colab to avoid issues with DiamondBlast installation)

Example

Input File

Protein sequence in fasta format. Our example input can be found in the sample_protein directory

Output File

Predcited functions for the protein in csv format. Our example output can be found in the sample_functions directory

domain-pfp's Issues

Multiple protein sequences

Dear Author.

Thank you for your contribution to protein function prediction. According to your paper, Domain-PFP performs well against many of the latest tools. I am considering using your tool to assign functions to multiple protein sequences (tens of thousands) in a fasta file. The highest confidence MF, BP, CC will be selected for each sequence. so I was wondering if you have developed this script so I don't have to repeat it again :-)

Best,
XJ

Request for terms used in CAFA3 dataset

Hi,

Thank you for contributing to the protein function prediction. I noticed that only the terms used in NetGO2.0 benchmark were provided (terms.pkl) while that file of the CAFA3 dataset was not given.

I tried to reproduce the number of labels used for each sub-ontology by various methods, such as only considering downstream GO term nodes. However, I cannot get the exact same number of labels used for each category presented in the supplementary part of the publication. Therefore, I am wondering if it is possible to provide the terms.pkl file used for CAFA3 dataset as well?

Best regards
WL