The Open Biomedical Network Benchmark (OBNB) is a comprehensive resource for setting up benchmarking graph datasets using biomedical networks and gene annotations. Our goal is to accelerate the adoption of advanced graph machine learning techniques, such as graph neural networks and graph embeddings, in network biology for gaining novel insights into genes' function, trait, and disease associations using biological networks. To make this adoption convenient, OBNB also provides dataset objects compatible with popular graph deep learning frameworks, including PyTorch Geometric (PyG) and Deep Graph Library (DGL).
A comprehensive benchmarking study with a wide-range of graph neural networks and graph embedding methods on OBNB datasets can be found in our benchmarking repository obnbench.
We provide a high-level dataset constructor to help users easily set up benchmarking graph datasets for a combination of network and label. In particular, the dataset will be set up with study-bias holdout split (6/2/2), where 60% of the most well-studied genes according to the number of associated PubMed publications are used for training, 20% of the least studied genes are used for testing, and the rest of the 20% genes are used for validation. For more customizable data loading and processing options, see the customized dataset construction section below.
from obnb.dataset import OpenBiomedNetBench
from obnb.util.version import get_available_data_versions
root = "datasets" # save dataset and cache under the datasets/ directory
version = "current" # use the last archived version
# Optionally, set version to the specific data version number
# Or, set version to "latest" to download the latest data from source and process it from scratch
# Download and process network/label data. Use the adjacency matrix as the ML feature
dataset = OpenBiomedNetBench(root=root, graph_name="BioGRID", label_name="DisGeNET",
version=version, graph_as_feature=True, use_dense_graph=True)
# Check the specific archive data version used
print(dataset.version)
# Check all available stable archive data versions
print(get_available_data_versions())Users can also load the dataset objects into ones that are compatible with PyG or DGL (see below).
from obnb.dataset import OpenBiomedNetBenchPyG
dataset = OpenBiomedNetBenchPyG(root, "BioGRID", "DisGeNET")Note: requires installing PyG first (see installation instructions)
from obnb.dataset import OpenBiomedNetBenchDGL
dataset = OpenBiomedNetBenchDGL(root, "BioGRID", "DisGeNET")Note: requires installing DGL first (see installation instructions)
Evaluation of simple machine learning methods such as logistic regression and label propagation can be done easily using the trainer objects.
from obnb.model_trainer import SupervisedLearningTrainer, LabelPropagationTrainer
sl_trainer = SupervisedLearningTrainer()
lp_trainer = LabelPropagationTrainer()Then, use the fit_and_eval method of the trainer to evaluate a given ML model over all tasks
in a one-vs-rest setting.
from sklearn.linear_model import LogisticRegression
from obnb.model.label_propagation import OneHopPropagation
# Initialize models
sl_mdl = LogisticRegression(penalty="l2", solver="lbfgs")
lp_mdl = OneHopPropagation()
# Evaluate the models over all tasks
sl_results = sl_trainer.fit_and_eval(sl_mdl, dataset)
lp_results = lp_trainer.fit_and_eval(lp_mdl, dataset)Training and evaluation of Graph Neural Network (GNN) models can be done in a very similar fashion.
from torch_geometric.nn import GCN
from obnb.model_trainer.gnn import SimpleGNNTrainer
# Use onehot encoded log degress as node feature by default
dataset = OpenBiomedNetBench(root=root, graph_name="BioGRID", label_name="DisGeNET",
auto_generate_feature="OneHotLogDeg", version=version)
# Train and evaluate a GCN
gcn_mdl = GCN(in_channels=1, hidden_channels=64, num_layers=5, out_channels=dataest.num_tasks)
gcn_trainer = SimpleGNNTrainer(device="cuda", metric_best="apop")
gcn_results = gcn_trainer.train(gcn_mdl, dataset)from obnb import data
# Load processed BioGRID data from archive.
g = data.BioGRID(root, version=version)
# Load DisGeNET gene set collections.
lsc = data.DisGeNET(root, version=version)from obnb.util.converter import GenePropertyConverter
from obnb.label.split import RatioPartition
# Load PubMed count gene property converter and use it to set up
# 6/2/2 study-bias based train/val/test splits
pubmedcnt_converter = GenePropertyConverter(root, name="PubMedCount")
splitter = RatioPartition(0.6, 0.2, 0.2, ascending=False,
property_converter=pubmedcnt_converter)# Apply in-place filters to the labelset collection
lsc.iapply(
filters.Compose(
# Only use genes that are present in the network
filters.EntityExistenceFilter(list(g.node_ids)),
# Remove any labelsets with less than 50 network genes
filters.LabelsetRangeFilterSize(min_val=50),
# Make sure each split has at least 10 positive examples
filters.LabelsetRangeFilterSplit(min_val=10, splitter=splitter),
),
)from obnb import Dataset
dataset = Dataset(graph=g, feature=g.to_dense_graph().to_feature(), label=lsc, splitter=splitter)OBNB can be installed easily via pip from PyPI:
pip install obnbOBNB provides interfaces with several other packages for network feature extractions, such as
PecanPy and GraPE.
To enable those extensions, install obnb with the ext extra option enabled:
pip install obnb[ext]Follow installation instructions for PyG or DGL to set up the graph deep learning library of your choice.
Alternatively, we also provide an installation script that helps you installthe graph deep-learning dependencies in a new conda environment obnb:
git clone https://github.com/krishnanlab/obnb && cd obnb
source install.sh cu117 # other options are [cpu,cu118]