Molecular Graph Generation via Geometric Scattering

Dhananjay Bhaskar, Jackson D. Grady, Egbert Castro, Michael Perlmutter, Smita Krishnaswamy

TL;DR

This paper introduces GRASSY, a novel method for generating novel molecular graphs with desired properties. By leveraging geometric scattering transforms to create a structured latent space and then using a GAN for generation, GRASSY achieves a higher proportion of chemically valid molecules compared to existing methods.

What Problem Does This Paper Address?

Generating novel molecules with specific properties is a critical challenge in drug discovery. While Graph Neural Networks (GNNs) have shown promise in predicting molecular properties, they often struggle with generating entirely new, valid molecular structures. Existing methods for molecular graph generation, such as sequential or reinforcement learning-based approaches, can be slow and produce a high percentage of invalid molecules that require extensive post-processing. This paper aims to address this by proposing a "representation-first" approach that prioritizes learning a well-organized latent space for molecular graphs.

Methods

GRASSY employs a multi-stage approach:

  1. Geometric Scattering Transform: This technique is used to generate multiscale descriptions of molecular graphs. Unlike traditional GNNs that focus on local neighborhood aggregation, the geometric scattering transform captures global graph structure information. It uses diffusion wavelets and collects statistical moments, resulting in a representation that is permutation-invariant and independent of graph size. The paper utilizes a learnable version of this transform, allowing the diffusion scales to be learned from the data, rather than being fixed.
  2. Regularized Autoencoder: The scattering moments are fed into an autoencoder. This autoencoder is trained with two key losses:
  3. Reconstruction Loss: Ensures the autoencoder can reconstruct the original scattering moments.
  4. Property Prediction Loss: A regression loss that penalizes the autoencoder's inability to predict physicochemical properties from the latent representation. This encourages the latent space to be organized by molecular properties.
  5. Adversarial Molecular Generation Network (GAN): A Generative Adversarial Network (GAN) is used to generate molecular graphs directly from the structured latent space produced by the autoencoder. This GAN is trained with:
  6. Adjacency Matrix Reconstruction Loss: Encourages the generated adjacency matrices to be similar to real ones.
  7. Adversarial Loss: Standard GAN loss where a discriminator tries to distinguish between real and generated molecules.
  8. Smoothness Loss: A novel loss that penalizes the rate of change of scattering moments with respect to interpolations in the latent space. This helps ensure that small changes in the latent space lead to small changes in the generated molecule's structure.

Datasets

The GRASSY framework was trained and evaluated on two primary datasets:

  • ZINC: A dataset of drug-like molecules. The paper used three specific tranches (FBAB, BBAB, JBCD) organized by molecular weight. For each molecule, 10 physicochemical properties were obtained from the ChEMBL database.
  • BindingDB: A database of drug-target interactions, used to train models for predicting binding affinities to specific protein targets (P14416 and P00918).

Key Insights and Ideas

  • Representation-First Generation: The core idea is that by first learning a highly structured latent space that encodes both graph structure and molecular properties, molecule generation becomes more tractable and controllable. This contrasts with methods that try to generate molecules step-by-step.
  • Geometric Scattering for Global Structure: Using geometric scattering transforms allows the model to capture global graph properties, which is crucial for understanding and generating molecules, overcoming limitations of purely local GNN operations.
  • Latent Space Smoothness for Properties: The property prediction loss and the smoothness loss in the GAN training encourage a latent space where similar points correspond to molecules with similar properties and structures, enabling smoother interpolation and property prediction.
  • Learned Scattering Scales: Adapting the scattering transform to learn its scales from the data (rather than using fixed dyadic scales) allows for better adaptation to different molecular datasets.

Key Results

  • Higher Validity: GRASSY generated a significantly higher fraction of "molecule-like" structures compared to other state-of-the-art methods like GSAE, GraphAF, and MolGAN. For example, on the BBAB tranche, GRASSY achieved 0.86 validity, while GSAE achieved 0.22 and GraphAF 0.79.
  • Accurate Property Prediction: Models with the regression penalty (e.g., GRASSY-AE + REGR) showed strong performance in predicting physicochemical properties in the latent space, with lower absolute errors compared to models without this penalty.
  • Smooth Latent Space: The paper demonstrates that physicochemical properties vary smoothly in the learned latent space, suggesting that the model can be used for property prediction and interpolation. Tables 2 and 5 show smoothness metrics across different models and datasets.
  • Effective GAN Generation: The GAN component successfully generated valid molecular adjacency matrices that have similar structures to nearby points in the latent space.

Limitations and Future Work

  • Dataset Size and Diversity: While ZINC and BindingDB are large, the paper focused on specific tranches. Future work could explore training on the entire ZINC dataset or other diverse molecular databases.
  • Computational Cost: The geometric scattering transform and the GAN training can be computationally intensive.
  • Generalizability to Novel Scenarios: While GRASSY shows promise for generating molecules with known properties, its ability to discover entirely novel chemical spaces or properties might require further investigation.
  • Future Directions: The authors suggest that the structured latent space could be used for more sophisticated goal-directed drug synthesis, exploring inverse design problems, and potentially for de novo design of proteins or other biomolecules.

Summary updated

Abstract

Although existing deep learning models perform remarkably well at predicting physicochemical properties and binding affinities, the generation of new molecules with optimized properties remains challenging. Inherently, most GNNs perform poorly in whole-graph representation due to the limitations of the message-passing paradigm. Furthermore, step-by-step graph generation frameworks that use reinforcement learning or other sequential processing can be slow and result in a high proportion of invalid molecules with substantial post-processing needed in order to generate valid molecules. To address these issues, we propose a representation-first approach to molecular graph generation. We guide the latent representation of an autoencoder by capturing graph structure information with the geometric scattering transform and apply penalties that organize the representation by molecular properties. We show that this highly structured latent space can be directly used for molecular graph generation by the use of a GAN. We demonstrate that our architecture learns meaningful representations of drug datasets and provides a platform for drug synthesis using publicly available ZINC and BindingDB datasets.
BibTeX
@inproceedings{bhaskar2021molecular,
  author = {Dhananjay Bhaskar and Jackson D. Grady and Egbert Castro and Michael Perlmutter and Smita Krishnaswamy},
  title = {{Molecular Graph Generation via Geometric Scattering}},
  booktitle = {MLSP},
  year = {2021},
  doi = {10.1109/mlsp55214.2022.9943379},
}
Figure