Graph Regulatory Autoencoder Network in python (GRANpy) for Gene Regulatory Network completion using scRNA-Seq datasets

Installation

Conda

The environment required can be installed via conda with the environment.yml file available in this repository:

conda env create -f environment.yml

Docker

Alternatively, a docker image is available on docker hub that includes all requirements:

https://hub.docker.com/repository/docker/marcostock/granpy/

Requirements

• TensorFlow (1.0 or later) with TensorBoard -- tested with tf1.14.0
• python 2.7
• networkx
• scikit-learn
• scipy

Run the demo (Yeast scRNA-Seq dataset)

After activating the conda GRANpy environment with:

conda activate GRANpy

the main algorithm with a yeast scRNA-Seq sample dataset can be started with:

python main.py

Options

Input data

--dataset

Default: gasch_GSE102475

In order to use your own data, you have to provide

• an N by N adjacency matrix (N is the number of nodes), and
• an N by D feature matrix (D is the number of features per node) -- optional

Have a look at the load_data() function in input_data.py.

--ground_truth

Default: yeast_chipunion_KDUnion_intersect

Gold standard edges file name.

Training

--model

Default: gcn_ae

You can choose between the following models:

• gcn_ae: Graph Auto-Encoder (with GCN encoder)
• gcn_vae: Variational Graph Auto-Encoder (with GCN encoder)

--features

Default: 1

Whether to use features (1) or not (0).

--random_prior

Default: 0

Wether prior adjacency matrix should be set to random matrix (1) or not (0).

--learning_rate

Default: 0.00001

Initial learning rate.

--epochs

Default: 1000

Number of epochs to train.

--hidden1

Default: 64

Number of units in hidden layer 1.

--hidden2

Default: 48

Number of units in hidden layer 2.

--weight_decay

Default: 0

Weight for L2 loss on embedding matrix.

--dropout

Default: 0

Dropout rate (1 - keep probability).

--early_stopping

Default: 5

Tolerance for early stopping (# of epochs).

Evaluation

--ratio_val

Default: 0.2

Ratio of edges used for validation metrics.

--ratio_test

Default: 0.1

Ratio of edges used for test metrics.

--balanced_metrics

Default: 1

Whether to use balanced metrics (1) or not (0).

BEEline

--inFilePath

Default: None

Input Files path.

--outFilePath

Default: None

Output Files path.

Others

--verbose

Default: 1

Verbosity of output from low (0) to high (1)

--crossvalidation

Default: 0

Whether to use crossvalidation (1) or not (0).

--hp_optimization

Default: 0

Whether to start the hyperparameter optimization run (1) or not (0).

Original paper by Kipf et. al. 2016 (Graph Auto-Encoders)

@article{kipf2016variational,
  title={Variational Graph Auto-Encoders},
  author={Kipf, Thomas N and Welling, Max},
  journal={NIPS Workshop on Bayesian Deep Learning},
  year={2016}
}

T. N. Kipf, M. Welling, Variational Graph Auto-Encoders, NIPS Workshop on Bayesian Deep Learning (2016)

Graph Auto-Encoders (GAEs) are end-to-end trainable neural network models for unsupervised learning, clustering and link prediction on graphs.

GAEs have successfully been used for:

• Link prediction in large-scale relational data: M. Schlichtkrull & T. N. Kipf et al., Modeling Relational Data with Graph Convolutional Networks (2017),
• Matrix completion / recommendation with side information: R. Berg et al., Graph Convolutional Matrix Completion (2017).

GAEs are based on Graph Convolutional Networks (GCNs), a recent class of models for end-to-end (semi-)supervised learning on graphs:

T. N. Kipf, M. Welling, Semi-Supervised Classification with Graph Convolutional Networks, ICLR (2017).

A high-level introduction is given in our blog post:

Thomas Kipf, Graph Convolutional Networks (2016)