DGL-LifeSci

DGL-LifeSci is a Python package designed to apply Graph Neural Networks (GNNs) to tasks in chemistry and biology. It is built on top of PyTorch, DGL (Deep Graph Library), and RDKit.

Key application areas include:

• Molecular property prediction
• Generative models
• Reaction prediction
• Protein-ligand binding affinity prediction

DGL-LifeSci is a toolkit built on top of DGL (Deep Graph Library) designed for applying graph neural networks to life science applications, such as molecular graphs and biological networks.

Key functionalities include:

• Graph construction
• Featurization
• Evaluation methods
• Model architectures
• Training scripts
• Pre-trained models

DGL-LifeSci provides specialized utilities for constructing and featurizing graphs representing protein-ligand complexes. These utilities are designed to facilitate deep learning tasks involving molecular docking or binding affinity prediction by converting complex 3D structures into graph representations.

Available utilities include:

• dgllife.utils.ACNN_graph_construction_and_featurization: For ACNN-style graph construction.
• dgllife.utils.PN_graph_construction_and_featurization: For PN-style graph construction.

DGL-LifeSci provides several graph construction methods to transform molecules (from SMILES or RDKit molecule objects) into DGL graphs. You can choose from three primary topologies:

• bigraph: Bi-directed graphs that correspond exactly to molecular graphs.
• complete_graph: Graphs where every atom is connected to every other atom.
• nearest_neighbor_graph: Graphs where each atom is connected to its $k$ closest neighbors based on 3D coordinates.

Available conversion utilities include:

• mol_to_graph / smiles_to_bigraph / mol_to_bigraph (and variants for complete/nearest neighbor graphs).
• Class-based interfaces: ToGraph, MolToBigraph, and SMILESToBigraph.

For molecular generation tasks, the Model Zoo provides:

• DGMG: A generative model for molecules.
• JTNNVAE: Junction Tree Variational Autoencoder.

The Weisfeiler-Lehman Network (WLN) implementation in the Model Zoo supports reaction-related tasks:

• WLN for Reaction Center Prediction: To identify the reaction center in a chemical reaction.
• WLN for Ranking Candidate Products: To rank potential products in a reaction.

Link prediction in DGL-LifeSci is the task of estimating the probability of links between nodes in a graph. A typical Graph Neural Network (GNN) workflow for link prediction follows these five steps:

1. Graph Construction: Construct graphs based on biological networks.
2. Feature Preparation: Prepare initial node and edge features for the graphs.
3. Node Representation Update: Use GNNs to update the node representations within the graphs.
4. Link Representation Computation: Compute the representation for a potential link by taking the product of its two updated nodes.
5. Prediction: Pass the resulting link representations through a Multi-Layer Perceptron (MLP) for training and final link prediction.

The DGL-LifeSci Model Zoo provides modular building blocks for constructing complex models. These components can be used to build custom architectures for graph-based tasks.

Available building blocks include:

• MLP Predictor: Multi-Layer Perceptron for property prediction.
• Hadamard Link Predictor: For link prediction tasks using Hadamard products.

When using the Alchemy dataset in DGL-LifeSci:

• Graph Construction: By default, a complete graph is constructed for each molecule, meaning every pair of atoms is connected.
• Data Splitting: The dataset uses a stratified split for training, validation, and test sets to ensure each set covers the full range of provided labels.
• Featurization: Specific details for node and edge featurization can be found in the DGL-LifeSci API documentation.