Exploring the GDB-13 chemical space using deep generative models

Josep Arús‐Pous, Thomas Blaschke, Silas Ulander, Jean‐Louis Reymond, Hongming Chen, Ola Engkvist

TL;DR

This paper explores how well deep generative models, specifically Recurrent Neural Networks (RNNs), can learn and sample from the vast chemical space. By training an RNN on a small fraction of the GDB-13 database, they demonstrate it can reproduce a significant portion of the entire database and introduce a novel method using negative log-likelihood plots to benchmark generative models.

What Problem Does This Paper Address?

The chemical space is astronomically large, making it impossible to explore exhaustively for discovering new molecules, especially for drug discovery. Traditional methods involve searching existing databases or using implicit models to navigate this space. While deep learning has shown promise in generating novel molecules, a key question remains: how much of the chemical space can these models truly learn and represent from limited training data? This paper aims to quantify this learning capability and understand the limitations imposed by molecular representations like SMILES.

Methods

The core of this research involves training Recurrent Neural Networks (RNNs) using SMILES strings as molecular representations. Specifically, they employ a Gated Recurrent Unit (GRU) architecture, which is a type of RNN known for its efficiency. To assess the training progress and the model's ability to explore chemical space, they introduce two key methods:

  • Negative Log-Likelihood (NLL) Plots: These plots track the probability of sampled SMILES sequences, providing insights into how well the model has learned the underlying distribution of the training data. By analyzing NLL distributions over training epochs, they can identify when the model is optimally trained.
  • "Coupon Collector Problem" Analogy: This mathematical framework is adapted to compare the trained model's sampling performance against an "ideal model" that samples uniformly from the entire GDB-13 database. This allows for a quantitative assessment of how much of the target chemical space the generative model has covered.

Datasets

  • GDB-13: This is the primary dataset used. It's a large, enumerated database containing approximately 975 million drug-like molecules with up to 13 heavy atoms.
  • Training Set: A random sample of 1 million molecules from GDB-13 was used to train the RNN model. This represents only about 0.1% of the entire GDB-13 database.
  • Validation Set: A separate sample of 100,000 molecules (not from the training set) was used to evaluate the model's performance during training.

Key Insights and Ideas

  • Impressive Coverage from Limited Data: Even training on a tiny fraction (0.1%) of GDB-13, the RNN model could reproduce a substantial portion (68.9%) of the entire database when sampling a large number of molecules. This highlights the power of deep learning in learning complex patterns from sparse data.
  • NLL Plots as a Training Guide: The negative log-likelihood plots proved to be a valuable tool for monitoring the training process. They allowed the researchers to identify the optimal training epoch where the model achieved a good balance between learning the training data and generalizing to unseen molecules, avoiding both underfitting and overfitting.
  • SMILES Syntax as a Bottleneck: The study revealed that the SMILES representation itself introduces biases. Molecules with complex ring systems and numerous heteroatoms are more difficult for the RNN to sample correctly, likely due to the sequential and context-sensitive nature of SMILES.
  • Benchmarking Generative Models: The "coupon collector problem" approach offers a novel and computationally efficient way to benchmark generative models. It provides an upper bound for coverage and allows for direct comparison of different architectures and training strategies.

Key Results

  • A model trained on 1 million GDB-13 molecules (0.1% of the database) reproduced 68.9% of the entire GDB-13 database when sampling 2 billion molecules.
  • The NLL analysis indicated that optimal training for a uniform and complete chemical space exploration was achieved around epochs 60-90.
  • When sampling 2 billion molecules, the model generated 85.6% of SMILES directly in their canonical form at epoch 70.
  • Analysis of molecules generated outside of GDB-13 showed that the most common broken constraints were "not allowed functional groups" (26.2%) and "not allowed graph" (19.8%), suggesting these are the hardest for the model to learn.

Limitations and Future Work

  • SMILES Limitations: The paper acknowledges that the SMILES format inherently limits the model's ability to learn complex molecular structures, particularly those with many rings and heteroatoms. This bias is significant for exploring the full chemical space.
  • Future Work:
    • Researchers are encouraged to use the proposed benchmarking method to compare different generative model architectures and input formats on GDB-13.
  • Investigating alternative molecular representations that are less ambiguous than SMILES (e.g., graph-based models) is suggested to overcome current limitations.
  • Exploring how training with smaller datasets impacts model performance and identifying optimal architectures and hyperparameters for minimizing overfitting is a promising avenue.
  • The authors suggest tailoring training sets to include more complex molecules or exploring modified SMILES syntaxes to address the bias towards simpler structures.

Summary updated

Abstract

Recent applications of recurrent neural networks (RNN) enable training models that sample the chemical space. In this study we train RNN with molecular string representations (SMILES) with a subset of the enumerated database GDB-13 (975 million molecules). We show that a model trained with 1 million structures (0.1% of the database) reproduces 68.9% of the entire database after training, when sampling 2 billion molecules. We also developed a method to assess the quality of the training process using negative log-likelihood plots. Furthermore, we use a mathematical model based on the "coupon collector problem" that compares the trained model to an upper bound and thus we are able to quantify how much it has learned. We also suggest that this method can be used as a tool to benchmark the learning capabilities of any molecular generative model architecture. Additionally, an analysis of the generated chemical space was performed, which shows that, mostly due to the syntax of SMILES, complex molecules with many rings and heteroatoms are more difficult to sample.
BibTeX
@article{aruspous2019exploring,
  author = {Josep Arús‐Pous and Thomas Blaschke and Silas Ulander and Jean‐Louis Reymond and Hongming Chen and Ola Engkvist},
  title = {{Exploring the GDB-13 chemical space using deep generative models}},
  journal = {Journal of Cheminformatics},
  year = {2019},
  doi = {10.1186/s13321-019-0341-z},
}
Figure