Advancing Protein Function Annotation with Equivariant Graph Networks

Protein function research helps in understanding complex biological processes within cells. However, the intricate nature of protein structures and functions, along with the rapid growth of protein sequence data, presents a pressing challenge to develop efficient computational methods for protein annotation. In this study, we propose ENGINE, a multi-channel deep learning framework designed for robust protein function annotation. ENGINE integrates an equivariant graph convolutional network model to capture geometric features from protein 3D structurals, leverages ESM-C to encode evolutionary and sequence-derived information, and combines an innovative 3D sequence representation that unifies spatial and sequential signals. We demonstrate that ENGINE consistently surpasses current state-of-the-art methods across diverse protein function prediction benchmarks, demonstrating robust generalization and high predictive accuracy. Beyond performance, ENGINE provides interpretable insights into key sequence features and substructures, enabling the identification of functionally critical residues within proteins. This facilitates a deeper mechanistic understanding of protein function annotation outcomes and supports hypothesis generation for downstream biological studies. By offering reliable predictions with biological interpretability, ENGINE contributes to advancing research into cellular processes and disease mechanisms.

Table of Contents

• Advancing Protein Function Annotation with Equivariant Graph Networks
• Table of Contents
• Data
• Model
• Installation
  • Environment
  • Dependency
    • Bioinformatics
    • Machine Learning & GNN
    • Data & Visualization
    • Utilities
    • Development & System Libraries
  • Feature extraction and tool description
    • Foldseek
    • DSSP
    • ESM-C
  • Create Environment with Conda
• Usage
• GUI (Graphical user interface)
  • Introduction
  • Running Environment
  • Running Methods
  • Use ENGINE in this platform
    • Input Data
    • Select Parameters
    • Predict and save results
    • Results display
      • Plot
      • Results Table
• License

Data

The original input dataset is in FASTA format, containing amino acid sequences of target proteins. To obtain corresponding 3D structures in PDB format, we use ESMFold for structure prediction.

Model

Due to GitHub's file size limitations, the .pt file of the ENGINE model can be downloaded from ENGINE model.

Installation

Environment

Anaconda
python 3.8

Dependency

Bioinformatics

biopython==1.83
bio==1.6.2
mygene==3.2.2
biothings-client==0.3.1
obonet==1.1.0
gprofiler-official==1.0.0
propcache==0.2.0

Machine Learning & GNN

torch==1.11.0+cu113
torch-cluster==1.6.0
torch_geometric==2.5.3
torch-scatter==2.0.9
torch-sparse==0.6.14
torch-spline-conv==1.2.1
egnn-pytorch==0.2.8
scikit-learn==1.3.2
joblib==1.4.2
networkx==3.1
opencv-python==4.11.0.86

Data & Visualization

numpy==1.24.4
pandas==2.0.3
scipy==1.10.1
matplotlib==3.7.5
seaborn==0.13.2

Utilities

pyyaml==6.0.2
click==8.1.8
requests==2.32.3
aiohttp==3.10.11
ttach==0.0.3
grad-cam==1.5.4

Development & System Libraries

attrs==24.3.0
backcall==0.2.0
pyzmq==25.1.2
qdarkstyle==3.2.3
qt-material==2.14
qtpy==2.4.3

Feature extraction and tool description

Foldseek

Extracting 3Di Tokens with Foldseek : github.com/steineggerlab/foldseek. Change to the corresponding Foldseek path in the infer_main.py and GUI/engine.py.

foldseek_executable = 'foldseek/bin/foldseek' # Foldseek installation path

DSSP

Extraction of secondary structure features using DSSP. Change to the corresponding DSSP path in the infer_main.py and GUI/engine.py.

dssp_executable = '/usr/bin/dssp' # DSSP installation path

ESM-C

We use the ESM-C 6B model provided by ESM (https://github.com/evolutionaryscale/esm) for sequence embedding extraction, which is currently only supported by the Forge API. Go to https://forge.evolutionaryscale.ai/ and register an account to get the API Token.
The infer_main.py and GUI/engine.py files in the project uses the Features/esm-c path to read features by default. Please change the path to the actual location according to the result you downloaded. Example:

# Enter the token you applied.
token= '******' # API Token

# Modify feature loading location
esm3_6B_path = 'Features/esm-c'  # Please change the path to the file you actually downloaded

Users can use the following script to extract ESM-c features on a large scale.

python ESM3C_feature_extractor.py --input_file pdbs.txt

The content of pdbs.txt is as follows, each line is the path to the protein pdb file. The ESM-c features will be saved to GUI/Features/esm-c

6ASO-C.pdb
4RH6-A.pdb
4FEZ-A.pdb

Create Environment with Conda

First, create the environment. Download and install the anaconda platform. (Refer to https://www.anaconda.com/docs/getting-started/anaconda/install#linux-installer).

conda create -n ENGINE python=3.8

Then, activate the "ENGINE" environment and enter into the workspace.

conda activate ENGINE
pip install -r requirements.txt

Usage

The infer_temp folder allows you to batch store PDB files that need to be tested and enter MF, BP or CC categories at runtime. The model file corresponding to the corresponding path in infer_main.py is saved in here, just download it and save it to the model folder. For example:

python infer_main.py --data_list ./infer_temp --cate cc

The sample output file is infer_result.csv with the following sample content:

4RH6-A,cell outer membrane,GO:0009279,0.5082
4RH6-A,cell division site,GO:0032153,0.0011
4RH6-A,cytoplasmic side of plasma membrane,GO:0009898,0.0002
4RH6-A,nuclear chromatin,GO:0000790,0.0000
4RH6-A,virion membrane,GO:0055036,0.0049
4RH6-A,inner mitochondrial membrane protein complex,GO:0098800,0.0000
4RH6-A,ribosome,GO:0005840,0.0011
4RH6-A,thylakoid,GO:0009579,0.0015
4RH6-A,nuclear envelope,GO:0005635,0.0027
4RH6-A,spliceosomal snRNP complex,GO:0097525,0.0000
4RH6-A,NADH dehydrogenase complex,GO:0030964,0.0001
4RH6-A,nucleoid,GO:0009295,0.0001
4RH6-A,respirasome,GO:0070469,0.0035
4RH6-A,neuron to neuron synapse,GO:0098984,0.0000
4RH6-A,bacterial-type flagellum,GO:0009288,0.0004
4RH6-A,cell surface,GO:0009986,0.1694
4RH6-A,host cell membrane,GO:0033644,0.1396

GUI (Graphical user interface)

Introduction

The platform integrates protein functional annotation with ENGINE. It allows you to use PDB inputs to predict information about protein function and visualise score alignments.

Running Methods

• Step 1. Enter the "GUI" folder.

  cd GUI/

• Step 2. Using the conda environment ENGINE above

  conda activate ENGINE

• Step 3. Run ENGINE

  python ENGINE_GUI_platform.py

Use ENGINE in this platform

After successfully running the ENGINE_GUI_platform.py file, you can follow the steps below to use ENGINE on our platform.

Input Data

• Please upload a PDB file for one of the proteins you want to predict!

Select Parameters

• Set 'MF', 'BP', or 'CC' and choose one of the categories for GO terms prediction.
  

Predict and save results

• Click the 'Predict' button to perform the prediction operation.
• If you want to continue to predict other proteins, please click ‘Clean’ button first.

Results display

Plot

• In the 'Plot' on the right, we have provided a bar graph to show how the predicted scores for GO terms are arranged from highest to lowest.
• You can customise the number of ‘Top’ displayed.
  

Results Table

• On the right-hand side, in Prediction, we provide a table of prediction results showing the functional annotation information of the entered proteins, including the PDB ID of the underlying proteins as well as GO terms, in addition to the names corresponding to the GO terms, and the last column represents the predicted scores.
• Click ‘Save results’ button to save the prediction result as a csv file.
  

License

This source code is licensed under the MIT license.