New Algorithms for Drug Design by Combining Chemistry, Algorithm, and Data
Currently, more than one million people die each year due to drug-resistant infectious diseases. The United Nations predicts that if current trends continue, this number will increase to ten million by 2050. Drug-resistance is also a central concern for cancer therapeutics: 90% of metastatic cancers become resistant to the drugs we use to treat them. It is critical that we not only find new drugs, but also find new drug design strategies. Computational approaches have the potential to let us design drugs digitally. But realizing this potential will require developing new algorithms that intelligently combine experimental data with chemical physics to characterize possible drugs and the proteins they target. Here, I discuss two such algorithms my collaborators and I have developed.
First, I describe a new family of neural networks for machine learning on molecular graphs that we call Automorphism-Based Graph Neural (Autobahn) Networks. Unlike prior neural networks which build neurons on individual atoms, Autobahn networks build neurons directly on chemical substructures such as functional groups. These networks are provably more powerful than the existing message-passing networks typically used for molecular learning tasks, and achieve results comparable with state-of-the-art architectures. Second, I present an approach to analyzing cryo-electron microscopy data to characterize how proteins move. Our work combines molecular dynamics with cryo-electron microscopy data using a Bayesian reweighting algorithm, allowing us to recover the full Boltzmann ensemble of a protein from experiment. By integrating chemical principles, molecular simulation, and machine learning, we hope that these algorithms will enable efficient discovery of new drug families.
Hosted by Professor Wilma Olson
~Coffee/tea will be served prior to the lecture~