Interpreting Molecule Generative Models for Interactive Molecule Discovery

Discovering novel molecules with desired properties is crucial for advancing drug discovery and chemical science. Recently deep generative models can synthesize new molecules by sampling random vectors from latent space and then decoding them to a molecule structure. However, through the feedforward generation pipeline, it is difficult to reveal the underlying connections between latent space and molecular properties as well as customize the output molecule with desired properties. In this work, we develop a simple yet effective method to interpret the latent space of the learned generative models with various molecular properties for more interactive molecule generation and discovery. This method, called Molecular Space Explorer (MolSpacE), is model-agnostic and can work with any pre-trained molecule generative models in an off-the-shelf manner. It first identifies latent directions that govern certain molecular properties via the property separation hyperplane and then moves molecules along the directions for smooth change of molecular structures and properties. This method achieves interactive molecule discovery through identifying interpretable and steerable concepts that emerge in the representations of generative models. Experiments show that MolSpacE can manipulate the output molecule toward desired properties with high success. We further quantify and compare the interpretability of multiple state-of-the-art molecule generative models. An interface and a demo video are developed to illustrate the promising application of interactive molecule discovery.