GraphCPP: A state-of-the-art graph convolutional neural network for cell-penetrating peptides

Installation

We recommend installing GraphCPP with mambaforge. Mamba is a fast, robust, and cross-platform package manager. Mambaforge is a Python-based CLI conceived as a drop-in replacement for conda, offering higher speed and more reliable environment solutions.

mamba env create -f environment.yml
mamba activate graphcpp

You can deactive the environment later with mamba deactivate

The repository has been tested on CUDA 11.7.

Repository structure

The following folder&file structure describes the modules used in this project.

├── main.py # Quick helper file to reproduce our results.
├── hyperparameter.py # Perform bayesian hyperparameter search using arm-mango.
├── dashboard.py # Single-page streamlit server for easy prediction locally.
├── cv.py # k-fold cross validation of the selected model.
├── config.py # Configuration file. This is where we specify the best architecture.
├── graphcpp # GraphCPP module.
│   ├── act.py # Activation layers.
│   ├── dataset.py # Dataset loader and featurizer functions.
│   ├── generalconv.py # Modified version of the general convolutional layers from Design Space for Graph Neural Networks https://arxiv.org/abs/2011.08843.
│   ├── lightning.py # Lightning modules.
│   ├── model.py # Model architecture.
│   ├── pooling.py # Global graph pooling layers.
│   └── utils.py # Misc utils.
├── assets # Misc static assets.
├── model # Saved model file.
├── README.md # Readme file.
└── .gitignore # gitignore file with general python ignores.

Prediction

GraphCPP utilizies the streamlit python package for locally hosting a web-server. You can start the prediction dashboard with the following command in the root directory of the project:

streamlit run dashboard.py

Hyperparameter optimization

We have utilized arm-mango to search the vast hyperparameter space of our model. In previous iterations of the model we have determined that the GraphSAGE convolutional layer performs best for our data. To perform hyperparameter optimization run the following script:

python hyperparameter.py

k-fold Cross-validation

To cross-validate the best model determined by hyperparameter optimization run the following script:

python cv.py

It defaults to 10-fold cross-validation.

Model

(model): GCN(
(pre_mp): GeneralMultiLayer(
    (Layer_0): GeneralLayer(
    (layer): Linear(
        (model): Linear(in_features=32, out_features=256, bias=False)
    )
    (post_layer): Sequential(
        (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (1): ReLU()
    )
    )
    (Layer_1): GeneralLayer(
    (layer): Linear(
        (model): Linear(in_features=256, out_features=256, bias=False)
    )
    (post_layer): Sequential(
        (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (1): ReLU()
    )
    )
    (Layer_2): GeneralLayer(
    (layer): Linear(
        (model): Linear(in_features=256, out_features=256, bias=False)
    )
    (post_layer): Sequential(
        (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (1): ReLU()
    )
    )
)
(mp): GNNStackStage(
    (layer0): GeneralLayer(
    (layer): SAGEConv(
        (model): SAGEConv(256, 256, aggr=max)
    )
    (post_layer): Sequential(
        (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (1): ReLU()
    )
    )
)
(post_mp): GNNGraphHead(
    (layer_post_mp): MLP(
    (model): Sequential(
        (0): GeneralMultiLayer(
            (Layer_0): GeneralLayer(
                (layer): Linear(
                (model): Linear(in_features=256, out_features=256, bias=False)
                )
                (post_layer): Sequential(
                (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
                (1): ReLU()
                )
            )
            (Layer_1): GeneralLayer(
                (layer): Linear(
                (model): Linear(in_features=256, out_features=256, bias=False)
                )
                (post_layer): Sequential(
                (0): BatchNorm1d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
                (1): ReLU()
                )
            )
        )
        (1): Linear(
            (model): Linear(in_features=256, out_features=1, bias=True)
        )
    )
    )
)

attilaimre99 / GraphCPP