Installation | Data | Usage | Contributors | Citation | References

This repository contains end-to-end training and evaluation of the SEAL [1] and Graph Auto-Encoder [2] link prediction algorithms on a Chemical Reaction Knowledge Graph built on reactions from USPTO. This code has been used to generate the results in [3].

In [3], a novel de novo design method is presented in which the link prediction is used for predicting novel pairs of reactants. The link prediction is then followed by product prediction using a transformer model, Chemformer, which predicts the products given the reactants. This repository covers the link prediction (reaction prediction) and for the subsequent product prediction we refer to the original Chemformer repository.

Link Prediction in this setting is equivalent to predicting novel reactions between reactant pairs. The code presented here is based on the implementation by Zhang et al. [1] of SEAL.

Figure 1. Overview of the method. (top) Step 1, link prediction in a Chemical Reaction Knowledge Graph (CRKG) using SEAL, and (bottom) Step 2, product prediction for highly predicted novel links using Chemformer.

After cloning the repository, the recommended way to install the environment is to use conda:

$ conda env create -f environment.yaml

Download the USPTO reaction graph from here and place it inside the data/ folder.

Use this repository to train and evaluate our proposed model with,

$ python main.py --graph_path [PATH-TO-GRAPH] --name [EXPERIMENT-NAME]

Optional settings can be provided as additional arguments. The training script generates the following files,

• data/: Processed data files.
• results/: Individual folders containing all relevant results from a GraphTrainer, including
  • Checkpoints of model and optimizer parameters, based on best validation AUC and best validation loss separately.
  • Log file of outputs from training, including the number of datapoints in train/valid/test split, number of network parameters, and more.
  • Pickle files of all results from training and testing separately.
  • Some preliminary plots.
  • Test metrics and test predictions in csv format.
  • A csv settings file of the hyperparameters used for training.

Once a SEAL model has been trained the probability of novel links can be predicted as follows,

$ python predict_links.py --model_dir_path [PATH-TO-TRAINED-SEAL] --save_path [SAVE-PATH] --graph_path [PATH-TO-GRAPH] --edges_path data/negative_links_uspto.csv

Exchange data/negative_links_uspto.csv with your potential links.

Most optimally, run with GPU available. In addition, SEAL-based link prediction is parallelizable on CPUs. Negative links generation by default uses a node degree distribution-preserving sampling function (sample_degree_preserving_distribution) which can take a long time depending on graph size. However, it only needs to be run once for a given link-sampling seed after which it is stored in data/. Alternatively, an approximating function (sample_distribution) can be used with quicker runtime.

torch_trainer.py contains the main trainer class and is called by the main.py, optimize.py and predict_links.py individually.

The main script initializes and runs a GraphTrainer from the torch_trainer.py file. The training process utilizes the following modules:

• datasets/reaction_graph.py: Importing graph and setting up training/validation/test positive edges.
• datasets/seal.py: Dynamic dataloader for SEAL algorithm, including sub-graph extraction and Double Radius Node Labelling (DRNL).
• datasets/GAE.py: Dataloader for GAE algorithm.
• models/dgcnn.py: The Deep Graph Convolutional Neural Networks used for the prediction of the likelihood of a link between the source and target nodes in the given subgraph.
• models/autoencoder.py: Graph Autoencoder used for prediction of the likelihood of a link between the source and target nodes, implemented using Torch Geometric library.
• utils/: various related functions used throughout the project.

The software is licensed under the Apache 2.0 license (see LICENSE), and is free and provided as-is.

• Emma Rydholm
• Tomas Bastys
• Emma Svensson

Please cite our work using the following reference.

@article{rydholm2024expanding,
    author = {Rydholm, Emma and Bastys, Tomas and Svensson, Emma and Kannas, Christos and Engkvist, Ola and Kogej, Thierry},
    title  =  {{Expanding the chemical space using a chemical reaction knowledge graph}},
    journal  = {Digital Discovery},
    year  = {2024},
    pages  = {-},
    publisher  = {RSC},
    doi  = {10.1039/D3DD00230F}
}

This work was partially supported by the Wallenberg Artificial Intelligence, Autonomous Systems, and Software Program (WASP), funded by the Knut and Alice Wallenberg Foundation. Additionally, this work was partially funded by the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska-Curie Innovative Training Network European Industrial Doctorate grant agreement No. 956832 “Advanced machine learning for Innovative Drug Discovery”.

[1] M. Zhang and Y. Chen, "Link prediction based on graph neural networks," Advances in neural information processing systems 31, 2018.

[2] T. N. Kipf and M. Welling, "Variational Graph Auto-Encoders", Neural Information Processing Systems 2016.

[3] E. Rydholm, T. Bastys, E. Svensson, C. Kannas, O. Engkvist and T. Kogej, "Expanding the chemical space using a Chemical Reaction Knowledge Graph," ChemRxiv. 2023

[4] R. Irwin, S. Dimitriadis, J. He and E. Bjerrum, "Chemformer: a pre-trained transformer for computational chemistry," Machine Learning: Science and Technology. 2022, 31 Jan. 2022.

Link prediction, chemical reactions, synthesis prediction, forward synthesis prediction, transformer, chemical space, de novo design, knowledge graph, reaction graph