Source-linked AI summary

DGL-LifeSci: An Open-Source Toolkit for Deep Learning on Graphs in Life Science

Mufei Li, Jinjing Zhou, Jiajing Hu, Wenxuan Fan, Yangkang Zhang, Yaxin Gu, George Karypis

arXiv:2106.14232v1cs.LGq-bio.QM

TL;DR

GNN-based life-science modeling is difficult because it combines graph preprocessing, programming, and deep-learning expertise. DGL-LifeSci packages this workflow into an open-source toolkit with command-line interfaces, modular components, and pre-trained models, and reports up to 6x speedup over previous implementations. It supports molecular property prediction, molecule generation, and reaction prediction, while its current focus remains small molecules.

  • Problem

    GNN modeling in life science requires programming, machine learning, and graph-modeling skills, creating a need for ready-to-run tools for custom datasets.

  • Method

    DGL-LifeSci is an open-source toolkit providing scripts, programming APIs, pre-trained models, and datasets for three graph-learning tasks.

  • Results

    Up to 6x speedup is reported versus previous implementations, with comparable WLN reaction-prediction performance and 76.4% JTVAE reconstruction accuracy.

  • Takeaways & Limitations

    Users can model custom datasets without programming, while advanced users can adapt modular building blocks for development.

  • Takeaways & Limitations

    The current implementations primarily focus on small molecules, with broader support for proteins and biological networks left for future work.

Abstract

from arXiv · show

Graph neural networks (GNNs) constitute a class of deep learning methods for graph data. They have wide applications in chemistry and biology, such as molecular property prediction, reaction prediction and drug-target interaction prediction. Despite the interest, GNN-based modeling is challenging as it requires graph data pre-processing and modeling in addition to programming and deep learning. Here we present DGL-LifeSci, an open-source package for deep learning on graphs in life science. DGL-LifeSci is a python toolkit based on RDKit, PyTorch and Deep Graph Library (DGL). DGL-LifeSci allows GNN-based modeling on custom datasets for molecular property prediction, reaction prediction and molecule generation. With its command-line interfaces, users can perform modeling without any background in programming and deep learning. We test the command-line interfaces using standard benchmarks MoleculeNet, USPTO, and ZINC. Compared with previous implementations, DGL-LifeSci achieves a speed up by up to 6x. For modeling flexibility, DGL-LifeSci provides well-optimized modules for various stages of the modeling pipeline. In addition, DGL-LifeSci provides pre-trained models for reproducing the test experiment results and applying models without training. The code is distributed under an Apache-2.0 License and is freely accessible at https://github.com/awslabs/dgl-lifesci.

1 INTRODUCTION

Life-science data often forms attributed graphs, making GNNs useful across molecular and biological tasks, but applying them requires programming, machine-learning, and graph-modeling expertise. DGL-LifeSci addresses this gap with ready-to-run tools, custom-dataset support, and pre-trained models.

  • Motivation: Chemical and biological tasks can be formulated as graph learning problems, including property prediction, molecule generation, and reaction prediction.These tasks map graphs to values, distributions, or product graphs.
  • Motivation: GNN-based modeling is challenging because users must build clean data-processing and modeling pipelines requiring programming, machine learning, and GNN expertise.The paper identifies this combined skill set as time-consuming to obtain.
  • Prior work: Prior packages lower the barrier but still require programming for some or all life-science modeling tasks.DeepChem, Chainer Chemistry, and PiNN are cited as examples with remaining coding requirements.
  • Contribution: DGL-LifeSci provides implementations and command-line scripts for molecular property prediction, molecule generation, and chemical reaction prediction on custom datasets.It also supplies pre-trained models for the experiments.
  • Contribution: Up to 6x speedup is reported relative to previous implementations.The result is stated as a comparison with previous implementations without specifying a single benchmark in this passage.

2 GRAPH NEURAL NETWORKS OVER MOLECULES

Molecular GNNs use iterative message passing to update atom representations, then read out a whole-molecule representation for prediction. Increasing message-passing rounds expands the local molecular context available to each node.

  • Graph neural networks: GNNs combine graph topology with node and edge features by iteratively aggregating neighboring representations.More iterations allow nodes to incorporate information from increasingly larger local subgraphs.
  • Molecular GNN pipeline: Molecular GNNs have separate message-passing and readout phases.The message-passing phase updates node representations, while readout produces a graph representation.
  • Message passing: Each message-passing round updates nodes using learnable functions of their own, neighboring-node, and incident-edge representations.After k rounds, each node aggregates information from nodes and edges within k hops.
  • Readout: The readout phase applies a function to all node representations to form a graph representation, which can then feed an MLP for prediction.Summation is given as one example of the readout function.

3 PACKAGE FEATURES

DGL-LifeSci combines ready-to-run scripts, programmable modules, pre-trained models, and built-in datasets for three major molecular and reaction tasks. Its command-line workflows support custom datasets after users provide standardized input files.

  • Package components: DGL-LifeSci has four components: scripts, programming APIs, pre-trained models, and built-in datasets.Pre-trained models can be fine-tuned or used directly for inference.
  • Supported tasks: The toolkit supports molecular property prediction, molecule generation, and chemical reaction prediction.Property prediction includes regression and classification for single molecules.
  • Usage: Command-line scripts invoke training and evaluation pipelines for each task using standardized user-provided data.For molecular property prediction, users supply SMILES strings and target properties in a CSV file.
  • Usage: Regression and classification training scripts expose command-line workflows for molecular property prediction.The supplied command examples are formatted as Python training-script invocations.
  • Usage: Optional hyperparameter search uses Bayesian optimization through hyperopt instead of only default settings.This feature is presented as an alternative to using defaults.

4 IMPLEMENTATION

DGL-LifeSci provides modular GNN tooling for molecular datasets, including graph construction, featurization, model implementations, and pre-trained representations. Its interfaces support multiple molecular learning tasks and custom dataset workflows.

  • Implementation: DGL-LifeSci uses PyTorch, DGL, and RDKit to provide a toolkit for graph-based modeling in life science.DGL supplies GNN functionality, PyTorch provides deep learning infrastructure, and RDKit supports cheminformatics utilities.
  • Modeling Pipeline and Modules: Its modeling pipeline modularizes dataset preparation, model initialization, and training into reusable stage components.Dataset preparation includes loading, graph construction, featurization, and dataset interfaces; training includes updates, metrics, and early stopping.
  • Dataset Preparation: Dataset interfaces support built-in and custom datasets, while graph construction includes molecular, distance-based, and complete molecular graphs.Distance-based graphs require molecular conformations and connect atoms within a cutoff; complete graphs connect every atom pair.
  • Dataset Preparation: DGL-LifeSci supports node and edge feature initialization from atom and bond descriptors, including raw numeric or one-hot representations.Non-numeric discrete descriptors use one-hot encoding, while numeric discrete descriptors can use raw values or one-hot encoding.
  • Models Included: The toolkit implements GCN, GAT, NF, Weave, MPNN, AttentiveFP, JTVAE, and WLN for molecular prediction, generation, and reaction prediction.JTVAE uses junction-tree and molecular-graph representations, while WLN identifies bond changes before ranking candidate products.
  • Pre-trained Models: Four pre-trained GIN models address limited molecular-property labels through context prediction, deep graph infomax, edge prediction, and attribute masking.The models were pre-trained using millions of unlabeled molecules and can be fine-tuned for particular properties.

5 MODELING PERFORMANCE

DGL-LifeSci was evaluated on molecular property prediction, reaction prediction, and molecule generation benchmarks. It achieved comparable WLN reaction-prediction performance and improved JTVAE reconstruction accuracy over the authors’ released code.

  • Molecular Property Prediction: Molecular property prediction uses six MoleculeNet binary classification datasets, scaffold splits, and ROC-AUC averaged across tasks.Models are trained with 80% training, 10% validation, and 10% test data to assess performance on unseen molecular structures.
  • Evaluation Procedure: The evaluation selects test performance after 32-trial hyperparameter searches with validation-based early stopping.Each trial trains a randomly initialized model and stops when validation performance fails to improve for 30 epochs.
  • Reaction Prediction: WLN achieves comparable performance to the original work on both reaction-center identification and candidate-product ranking on USPTO.The two stages identify candidate bond changes, then enumerate and rank products.
  • Molecule Generation: 76.4% reconstruction accuracy is achieved by JTVAE on a ZINC subset, compared with 74.4% from the authors’ released code.The evaluation reconstructs input molecules.

6 TRAINING SPEED

DGL-LifeSci’s modeling efficiency is compared with prior implementations, including original implementations and DeepChem. Training time is measured per epoch on a fixed V100-based AWS EC2 testbed.

  • Training Speed: Training efficiency is compared using averaged training time for one epoch across DGL-LifeSci, original implementations, and DeepChem.The comparison uses one AWS EC2 p3.2xlarge instance with one NVIDIA V100 GPU and 8 VCPUs.

7 CONCLUSIONS

DGL-LifeSci supports GNN-based modeling for three life-science tasks and enables no-code modeling on custom datasets. Its current implementations primarily focus on small molecules, while broader graph support remains future work.

  • DGL-LifeSci supports GNN-based modeling for molecular property prediction, reaction prediction, and molecule generation.
  • Command-line interfaces enable efficient modeling on custom datasets without programming, while advanced users can adapt modular building blocks.
  • Current implementations primarily focus on small molecules, with proteins and biological networks identified as future extensions.
  • Table 6 reports epoch training time in seconds.

8 DATA AND SOFTWARE AVAILABILITY

DGL-LifeSci’s datasets, models, and experiment-reproduction scripts are publicly available through the project repository. Examples cover molecular property prediction, reaction prediction, and molecule generation.

  • The datasets and models are publicly available through the DGL-LifeSci GitHub repository.
  • Scripts for reproducing experiments are provided in the project examples.
  • Examples cover molecular property prediction, reaction prediction, and molecule generation.
Loading 2106.14232v1…