DPGOK

A Deep Learning-Based Method for Protein Function Prediction by Fusing GO Knowledge and Protein Features

Usage

Here we provide instructions for two use cases: (1) Retraining our model on our or your data. (2) Testing data on trained models.

If you encounter any bugs or issues, feel free to contact us.

Dependencies

• The code was developed and tested using python 3.10.
• Clone the repository: https://github.com/CSUBioGroup/DPGOK.git
• Create virtual environment with Conda or python3-venv module.
• Install PyTorch, DGL and other requirements: conda create --name new_env --file requirements.txt

Prepare Data

Using our dataset

• Download the dataset.tar.gz
• Extract tar xvzf dataset.tar.gz

Generating the data

If you like to generate a new dataset follow these steps:

• Download Gene Ontology. You'll need go.obo file and save them into dataset folder.
• Download Uniprot-KB You'll need three uniprot data by time stamps.
• Run python split_data.py to split data into train, test, valid sets.
• Run python generate_graph_data.py to get graph data.
• Download esm2_t36_3B_UR50D You'll need esm2_t36_3B_UR50D.pt file and save them into esm2 folder.
• Run python extract_esm2.py to generate esm2 features.

Data Construction

All data required save in ./dataset

• [cc|mf|bp]_direct_parents_pairs.pkl and [cc|mf|bp]_ancestors_pairs.pkl #graph data
• [train|valid|test]_esm2.pkl #esm features
• select_min_count_1_[mf|bp|cc]_labels.csv #all go terms
• [train|valid|test]_data_separate.pkl #proteins information including proteinID, sequence, selected_terms, all_terms...
• [train|valid|test]_seq.fasta #protein sequence

Train/valid/test data of each ontology save in ./dataset/[mf|bp|cc]

• [train|valid|test]_data_separate_[mf|bp|cc]_gos.pkl #go names of each subset
• [train|valid|test]_data_separate_[mf|bp|cc]_labels.pkl #onehot labels of each subset
• [train|valid|test]_data_separate_[mf|bp|cc]_proteins.csv #proteinID of each subset
• [train|valid|test]_data_separate_[mf|bp|cc]_sequences.pkl #sequences of each subset

Training DuGPro

If you have prepared the data, you can train our model on your data as follows (Ensure that your configure is right):

cd ./src
python train.py --ont cc --lr 5e-5

arguments:
    --ont: the ontology(mf/bp/cc)
    --lr: learning rate(defult: {'mf':7e-5, 'cc':5e-5,'bp':5e-5})

Predicting

• You could use retrained model or our trained model to predict functions.

cd ./src
python predict.py --ont cc
arguments:
    --ont: the ontology(mf/bp/cc)

If you use our trained model, please first download saved_model and save them into saved_model folder.

Evaluation

• We provide Fmax, AUPR, IC_AUPR and DP_AUPR metrics to evaluate the performance.

cd ./utils
python metrics.py

Contact

Please feel free to contact us for any further questions.

• Qiurong Yang [email protected]
• Min Li [email protected]