PANDA-3D: protein function prediction based on AlphaFold models

This repository includes code and a pre-trained model of PANDA-3D for protein function prediction.

Protein function prediction

PANDA-3D can predict protein function using AlphaFold-predicted structures. PANDA-3D is offered as a stand-alone tool and a web server.

Stand-alone PANDA-3D

1. Download the trained model to panda3code/

wget http://dna.cs.miami.edu/PANDA-3D/download_files/trained.model panda3code/

2. Download a predicted structure to example/ from the AlphaFold Protein Structure Database. The protein "INTS1" with UniProtID "A0A6J7ZWZ8" has already been downloaded.
3. Run PANDA-3D

>> python run_PANDA-3D.py example/

Dependencies
PANDA-3D is built under Python 3.10.
The conda environment is shared via "conda_environment.yml".

Web-server PANDA-3D

Submit predicted structures at http://dna.cs.miami.edu/PANDA-3D/

Output explaination:

The output is saved in example/prediction.txt. The file format is as follows:

• The first two lines contain model information.
• Subsequent lines follow this format:
  • The first column is the name of the PDB structure file.
  • The second column is the GO term ID.
  • The third column is the confidence score predicted by PANDA-3D.
• The last line contains only "END."

MODEL 1
KEYWORDS sequence embedding, geometric vector perceptron, transformer.
AF-A0A6J7ZWZ8-F1-model_v4       GO:0003006      0.93
AF-A0A6J7ZWZ8-F1-model_v4       GO:0010605      0.93
AF-A0A6J7ZWZ8-F1-model_v4       GO:0070727      0.92
...
AF-A0A6J7ZWZ8-F1-model_v4       GO:0016798      0.09
END

Training, validation, and testing datasets

Datasets are available at http://dna.cs.miami.edu/PANDA-3D/download_files/alphafoldDB_pdb_1115/

Proteins are randomly split into training (80%), validation (10%), and testing (10%). The UniProt IDs for training, validation, and testing are saved into train_0130_df.pkl, valid_0130_df.pkl, and test_0130_iden_pd3_dpfri.pkl, respectively.

The protein features, including UniProt ID, true GO annotations, 3D coordinates, GO term label, pLDDT scores, amino acid sequence, ESM features, and sequence length, are saved into a pickle file in the format of a Python dictionary. These features are further divided into subfolders based on the first two letters of the protein ID. For example, the features of protein A0A0A1C3I2 can be found in A0/A0A0A1C3I2.pkl.

Citation

@article{zhao2023PANDA3D,
  title={PANDA-3D: protein function prediction based on
AlphaFold models},
  author={Zhao, Chenguang and Liu, Tong and Wang, Zheng},
  publisher={Under Review}
}