Need help with dgl-lifesci?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

178 Stars 51 Forks Apache License 2.0 184 Commits 14 Opened issues


Python package for graph neural networks in chemistry and biology

Services available


Need anything else?

Contributors list


Documentation | Discussion Forum

We also have a slack channel for real-time discussion. If you want to join the channel, contact [email protected]


Deep learning on graphs has been an arising trend in the past few years. There are a lot of graphs in life science such as molecular graphs and biological networks, making it an import area for applying deep learning on graphs. DGL-LifeSci is a DGL-based package for various applications in life science with graph neural networks.

We provide various functionalities, including but not limited to methods for graph construction, featurization, and evaluation, model architectures, training scripts and pre-trained models.

For a list of community contributors, see here.

For a full list of work implemented in DGL-LifeSci, see here.



DGL-LifeSci should work on

  • all Linux distributions no earlier than Ubuntu 16.04
  • macOS X
  • Windows 10

DGL-LifeSci requires python 3.6+, DGL 0.5.2+ and PyTorch 1.5.0+.

Install pytorch

Install dgl

Additionally, we require

RDKit 2018.09.3
for utils related to cheminformatics. We recommend installing it with
conda install -c rdkit rdkit==2018.09.3

For other installation recipes for RDKit, see the official documentation.

Pip installation for DGL-LifeSci

pip install dgllife

Conda installation for DGL-LifeSci

pip install hyperopt
conda install -c dglteam dgllife

Installation from source

If you want to try experimental features, you can install from source as follows:

git clone
cd dgl-lifesci/python
python install

Verifying successful installation

Once you have installed the package, you can verify the success of installation with

import dgllife



If you are new to DGL, the first time you import dgl a message will pop up as below:

DGL does not detect a valid backend option. Which backend would you like to work with?
Backend choice (pytorch, mxnet or tensorflow):

and you need to enter


Example Usage

To apply graph neural networks to molecules with DGL, we need to first construct

-- the graph data structure in DGL and prepare initial node/edge features. Below gives an example of constructing a bi-directed graph from a molecule and featurizing it with atom and bond features such as atom type and bond type.
from dgllife.utils import smiles_to_bigraph, CanonicalAtomFeaturizer, CanonicalBondFeaturizer

Node featurizer

node_featurizer = CanonicalAtomFeaturizer(atom_data_field='h')

Edge featurizer

edge_featurizer = CanonicalBondFeaturizer(bond_data_field='h')

SMILES (a string representation for molecule) for Penicillin

smiles = 'CC1(C(N2C(S1)C(C2=O)NC(=O)CC3=CC=CC=C3)C(=O)O)C' g = smiles_to_bigraph(smiles=smiles, node_featurizer=node_featurizer, edge_featurizer=edge_featurizer) print(g) """ DGLGraph(num_nodes=23, num_edges=50, ndata_schemes={'h': Scheme(shape=(74,), dtype=torch.float32)} edata_schemes={'h': Scheme(shape=(12,), dtype=torch.float32)}) """

We implement various models that users can import directly. Below gives an example of defining a GCN-based model
for molecular property prediction.

from dgllife.model import GCNPredictor

model = GCNPredictor(in_feats=1)

For a full example of applying

, run the following command
python examples/property_prediction/ -m GCN -d Tox21

For more examples on molecular property prediction, generative models, protein-ligand binding affinity prediction and reaction prediction, see


We also provide pre-trained models for most examples, which can be used off-shelf without training from scratch. Below gives an example of loading a pre-trained model for

on a molecular property prediction dataset.
from import Tox21
from dgllife.model import load_pretrained
from dgllife.utils import smiles_to_bigraph, CanonicalAtomFeaturizer

dataset = Tox21(smiles_to_bigraph, CanonicalAtomFeaturizer()) model = load_pretrained('GCN_Tox21') # Pretrained model loaded model.eval()

smiles, g, label, mask = dataset[0] feats = g.ndata.pop('h') label_pred = model(g, feats) print(smiles) # CCOc1ccc2nc(S(N)(=O)=O)sc2c1 print(label_pred[:, mask != 0]) # Mask non-existing labels

tensor([[ 1.4190, -0.1820, 1.2974, 1.4416, 0.6914,

2.0957, 0.5919, 0.7715, 1.7273, 0.2070]])

Similarly, we can load a pre-trained model for generating molecules. If possible, we recommend running the code block below with Jupyter notebook.

from dgllife.model import load_pretrained

model = load_pretrained('DGMG_ZINC_canonical') model.eval() smiles = [] for i in range(4): smiles.append(model(rdkit_mol=True))






If you are running the code block above in Jupyter notebook, you can also visualize the molecules generated with

from IPython.display import SVG
from rdkit import Chem
from rdkit.Chem import Draw

mols = [Chem.MolFromSmiles(s) for s in smiles] SVG(Draw.MolsToGridImage(mols, molsPerRow=4, subImgSize=(180, 150), useSVG=True))

Speed Reference

Below we provide some reference numbers to show how DGL improves the speed of training models per epoch in seconds.

| Model | Original Implementation | DGL Implementation | Improvement | | ---------------------------------- | ----------------------- | -------------------------- | ---------------------------- | | Neural Fingerprint on HIV | 5.8 (DeepChem 2.3.0) | 2.5 | 2.3x | | AttentiveFP on Aromaticity | 6.0 | 1.2 | 5x | | JTVAE on ZINC | 1826 | 743 | 2.5x | | WLN for reaction center prediction | 11657 | 858 (1 GPU) / 134 (8 GPUs) | 13.6x (1GPU) / 87.0x (8GPUs) | | WLN for candidate ranking | 40122 | 22268 | 1.8x |

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.