A generative model for molecule generation based on chemical reaction trees

Dai Hai Nguyen, Koji TsudaarXiv 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel generative model that creates not just new molecules with desired properties, but also provides a complete, multi-step synthetic route to create them. This addresses a critical gap in current molecule generation, where synthesized molecules are often difficult or impossible to make in practice.

What Problem Does This Paper Address?

The discovery of new molecules with specific chemical properties is crucial for fields like drug discovery. While deep generative models have become adept at designing novel molecules, they often fall short when it comes to practical synthesis. Existing methods either generate molecules without considering how they'd be made, or they are limited to single-step reactions, restricting the diversity of achievable compounds. This paper tackles the challenge of generating molecules and their entire multi-step synthetic pathways, ensuring practical synthesizability.

Methods

The core innovation lies in representing chemical synthesis as a "reaction tree." This tree structure alternates between molecule nodes (starting materials, intermediates, or final products) and reaction template nodes (representing specific chemical transformations). The model uses a variational autoencoder (VAE) framework, but with a twist: it jointly learns representations for both the molecule's structure (via a "junction tree") and its synthetic pathway (the "reaction tree").

The process involves:

  1. Encoding: A junction tree encoder captures the structural essence of a molecule, and a reaction tree encoder captures its synthesis. These are mapped to latent codes.
  2. Decoding: A decoder takes these latent codes and reconstructs both the junction tree (molecule structure) and the reaction tree (synthetic pathway). This is done in a top-down, recursive manner, ensuring that each step in the synthesis is guided by both the desired molecular structure and the learned patterns of chemical reactions.

Datasets

The model was trained on a dataset of 21,218 valid multi-step chemical reaction trees. This dataset was generated using the Retro* algorithm on the USPTO reaction data. The training data included vocabularies of 9,766 starting molecules and 5,567 reaction templates. A vocabulary of 275 valid chemical substructures was also extracted for junction tree generation.

Key Insights and Ideas

  • Reaction Trees as a Unified Representation: Representing synthesis as a tree structure allows for a more holistic approach to molecule generation, integrating structure and synthesis.
  • Joint Learning of Structure and Synthesis: By jointly encoding and decoding junction trees (for structure) and reaction trees (for synthesis), the model can better guide the generation process, leading to more plausible and synthesizable molecules.
  • Bridging the Gap to Practical Synthesis: The model's explicit generation of multi-step reaction pathways directly addresses the "synthesizability" problem, a major bottleneck in current AI-driven molecule design.
  • Bayesian Optimization for Property Tuning: The model's latent space is amenable to optimization techniques like Bayesian Optimization, allowing for the targeted generation of molecules with specific desired chemical properties.

Key Results

  • Generation Metrics: The model achieved acceptable validity (64.5%), uniqueness (73.87%), and novelty (66.44%) in random molecule generation. Notably, it achieved a high quality score (95.80%) and a low Frechet ChemNet Distance (0.66), indicating stable and novel molecules.
  • Property Optimization: Bayesian optimization successfully guided the model to generate molecules with significantly higher QED and penalized logP scores compared to random sampling.
  • Synthesizability: The model demonstrated a superior synthetizability rate of 97.81%, meaning nearly all generated molecules could be successfully synthesized using the Retro* tool, a significant improvement over existing methods like JT-VAE (50.5%) and MOLECULE CHEF (65.6%).

Limitations and Future Work

The authors acknowledge that a significant limitation is the reliance on a large dataset of existing chemical reaction trees, which may not always be available or comprehensive for all desired molecules. Future work will focus on developing machine learning models that can handle reaction trees and molecules even when complete synthetic routes are not readily available in the training data. They also suggest exploring explainable AI methods to better understand the model's decision-making process.

Summary updated

Abstract

Deep generative models have been shown powerful in generating novel molecules with desired chemical properties via their representations such as strings, trees or graphs. However, these models are limited in recommending synthetic routes for the generated molecules in practice. We propose a generative model to generate molecules via multi-step chemical reaction trees. Specifically, our model first propose a chemical reaction tree with predicted reaction templates and commercially available molecules (starting molecules), and then perform forward synthetic steps to obtain product molecules. Experiments show that our model can generate chemical reactions whose product molecules are with desired chemical properties. Also, the complete synthetic routes for these product molecules are provided.
BibTeX
@misc{nguyen2021generative,
  author = {Dai Hai Nguyen and Koji Tsuda},
  title = {{A generative model for molecule generation based on chemical reaction trees}},
  year = {2021},
  eprint = {2106.03394},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.03394},
}
Figure