SeXei/AEGAN

a new model of catalytic sites prediction

A highly sensitive model based on graph neural networks for enzyme key catalytic residues prediction

A new enzyme catalytic sites prediction model use new designed Adaptive Edge-Gated Graph Attention Neural Network(AEGAN) to process both sequence and structure features of proteins at different levels was developed here.

Requirements

torch==1.10.0
tqdm==4.62.3
numpy==1.23.2
pytorch-lightning==1.7.2
biotite==0.34.1
pandas==1.3.4
scikit-learn==1.1.1
psiblast==2.12.0+

Step-by-step running:

Make Dataset

• First, install BlAST using:
  conda install -c bioconda blast
• Download protein database for alignment, using: wget ftp://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref90/uniref90.fasta.gz
• compile the protein database as follows:

  gzip -d uniref90.fasta.gz
  makeblastdb -in uniref90.fasta -parse_seqids -hash_index -dbtype prot
  

• and then, cd AEGAN/code ,you can create datasets, using
  DataProcess.py -s uni14230 -d your-database-path
  there are some parameters you can choose:
  • -s:specifies which dataset to create
    • uni14230 train dataset
    • uni3175 test dataset
    • EF_fold benchmark dataset
    • EF_family benchmark dataset
    • EF_superfamily benchmark dataset
    • HA_superfamily benchmark dataset
    • NN benchmark dataset
    • PC benchmark dataset
  • -t:(optional) specify where the dataset will be saved
  • -d:database path used for BLAST alignment

Train a model

• cd AEGAN/code, and use train.py to train a model: python train.py -sm model-saved-path
  there are some parameters you can choose:
  • --trainset:The path of the dataset of the training model (default:../database/trainset)
  • --batchsize:batchsize (default:200)
  • -lr:learning rate (default:1e-2)
  • -ly:layers of AEGAN module (default:24)
  • --epoch:Traing epochs (default:1000)
  • -sm:Path of saving trained model
  • --accelerator: Uesd device to train model (default:gpu)
    • cpu:use cpu to train model
    • gpu:use gpu to train model

Test model

• cd AEGAN/code, and use validate.py to test model's performance on datasets:
  python validate.py --model trained-model-path
• if you haven't trained a model yet, you can use the model we've trained to test it:
  python validate.py --model trained_full
• and you can also get the results of ablation experiment using the trained model: python validate.py --model EP --lack PSSM
  there are also some parameters you can choose:
  • --model:specifies the path of the model to test. it can be a model file or folder contains model files.
    • trained_full:trained full features model
    • EP:trained model which features exclude PSSM
    • EAT:trained model which features exclude Atchley
    • EAM:trained model which features exclude AtomModel
  • --testset:Test dataset (default:uni3175)
    • uni3175
    • EF_fold
    • EF_family
    • EF_superfamily
    • HA_superfamily
    • NN
    • PC
  • --lack:specify which feature is missing
    • PSSM
    • Atchley
    • AtomModel
  • --accelerator: Uesd device to train model (default:gpu)
    • cpu:use cpu to train model
    • gpu:use gpu to train model
  • --batchsize:batchsize (default:100)