Installation

Required environment: Python 3.8, PyTorch 2.0.1

For MIGU and NANA algorithm:

conda activate MIGU conda env update --file environment_pygdemo.yml

For generating biophysical & other features:

conda activate hydrogen_bond conda env update --file environment_bio.yml

Dataset Preparation

Please download the dataset XXX from *** and put the file under the folder XXX

Run Augmentation

Generate Augmented Dataset

Please run the following command

python hydrogen_bond.py

Train the Model

Command Line Options

• --training_title: Name of the training, type: string
• --epochs: Training epochs, type: float
• --lr: Learning rate, type: Float
• --optimizer: Name of the optimizers, type: string, options: [adam]
• --dim_h: the width of the layers, type: integer
• --batch_size: Training batch size, type:
• --model: Model types, type: string, options: [GIN_Attribute]
• --num_workers: Number of dataloader workers, type: integer
• --dataset_path: The path of loaded dataset, type: string
• --dataset: The used dataset for the training, type: string
• --edge: Whether to use co-embedding residual learning, type: boolean
• --eval_batch_size: The batch size in the evaluation stage, type: integer
• --version: Use augmented or non-augmented data, "mine" means augmented data, "their" means non-augmented data, type: string, options: [their, mine]
• --type: The types of layers, type: string, options: [GCN, GIN, MPNN]
• --bond: Use edge attributes or not; if the flag is true, it would represent the MiGu augmentation; otherwise it is NaNa augmentation, type: boolean

Run with the Following Command

For example, to train a GCN model with augmented EC dataset and residual learning framework, you can use the following command

CUDA_LAUNCH_BLOCKING=1 python graph_hao_retry.py --training_title 226_MPNN_true_edge --epochs 4000 --lr 0.0005 --optimizer adam --dim_h 128 --batch_size 50 --model GIN_Attribute --num_workers 8 --dataset_path /home/ysl_0128/DIG/examples/threedgraph/dataset --dataset EC --edge true --eval_batch_size 20 --version mine --type GCN --bond false