George Mason University,
Keywords: AI, novel molecules
Summary:Designing novel molecules in silico is a key objective to advance cheminformatics, drug discovery, biotechnology, and material science but remains challenging for a variety of reasons. The chemical space is vast and complex, and the relationship between chemical structures and biological properties is non-trivial. Deep graph generative models that learn directly from data present a great opportunity to advance small molecule generation, but current state-of-the-art models are unable to link the chemical and biological space. In this talk I will present several novel disentangled graph variational autoencoder models that provide us with molecular property control. I will first show how we leverage inductive biases to connect learned latent factors to molecular properties. I will additionally present monotonically-regularized models that learn and exploit the correspondence between latent variables and molecular properties parameterized by polynomial functions. Extensive experimental evaluation on diverse benchmark datasets demonstrates the superiority of these models on accuracy, novelty, disentanglement, and control towards desired molecular properties. This work opens up an exciting line of research on controllable molecule generation with broad scientific and translational impact.