Generative Models of Molecular Structure

Abstract: Generative models draw samples from distributions given by data, and have transformed how we generate text, images, audio, and video. Transferring this capability to molecular structure is critical, because progress on global challenges in health, energy, and the environment is gated by the discovery of new molecules and materials. Generative models of molecular structure already address central problems in chemistry such as inverse design, structure elucidation, and accelerated simulation; however, they remain limited in their molecular representations, problem domains, and architecture design choices. This thesis presents six works that advance molecular generative modelling along each of these axes. We introduce a robust fragment-based representation of molecular structure, the first generative models for structure elucidation from rotational spectroscopy, a generative model that predicts organic crystal structure in seconds, and autoregressive architectures for 3D molecules which greatly increase the flexibility of molecular generative models. Taken together, these contributions enable molecular generative models that generalize to more prompts, produce faster responses, and solve a broader range of chemistry problems.

The seminar will be followed by a closed oral examination.  

Meeting Link: Zoom
Meeting ID: 862 0099 7619
Passcode: FOESeminar