AutoGRN

The implementation code of the paper "AutoGRN: An Automated Graph Neural Network Framework for Gene Regulatory Network Inference".

Requirements

python==3.7.12
torch==1.12.1+cu116
dgl==1.0.1.cu116    if CPU dgl==1.0.1
scikit-learn==1.0.2
numpy==1.19.5
networkx==2.6.3
requests==2.31.0
psutil==6.1.0
tqdm==4.67.1
pandas==1.3.5
tables==3.7.0
if use GPU torch==2.0.0+cu118
pip install dgl==2.0.0+cu118 -f https://data.dgl.ai/wheels/cu118/repo.html

Dataset Preparation

download the datasets from Google Drive:

https://drive.google.com/file/d/1nZai2lTdVmb-WwZIsY1e_mjxqoB4gkod/view?usp=sharing

then unzip this folder and post it in the root directory of the project (AutoGRN/)

unzip data_evaluation.zip

Quick Start

A quick start example is given by:

$ python auto_main.py

By default, auto_main.py uses PPO-based RL search (--search_method rl --rl_algorithm ppo).

To reproduce the original genetic search in the paper:

$ python auto_main.py --search_method genetic

RL examples:

# Default: PPO + Critic, faster search-phase training
$ python auto_main.py --search_method rl --rl_algorithm ppo --search_train_epoch 80

# Legacy REINFORCE baseline
$ python auto_main.py --search_method rl --rl_algorithm reinforce

# Fast smoke test
$ python auto_main.py --search_method rl --rl_warmup_num 5 --search_epoch 2 --sharing_num 3 --search_train_epoch 20

Optimal GNN Architectures

The following table summarizes the optimal GNN architectures searched by AutoGRN for each dataset:

Dataset	Conv Func	Bi Conv Func	Activation	Hidden Dim	Fusion Type
boneMarrow	GATv2Conv	BiGCNConv	leaky_relu	512	concat
mESC_1	SAGEConv	BiNoneConv	relu	256	concat
mESC_2	TAGConv	BiGraphConv	leaky_relu	256	sum
mHSC_E	SAGEConv	BiSAGEConv	leaky_relu	256	abs_difference
mHSC_GM	GATv2Conv	BiSAGEConv	relu	1024	max
mHSC_L	SAGEConv	BiSAGEConv	leaky_relu	256	concat

Run auto_test.py to directly verify the result of the optimal GNN searched by AutoGRN

$ python auto_test.py