Generative AI for Molecules: Semi-Equivariant Flows, Sketchy Diffusion, & Quantum Ground States
Generative AI for Molecules: Semi-Equivariant Flows, Sketchy Diffusion, & Quantum Ground States
Dr. Daniel Freedman
Abstract:
Generative AI has made tremendous strides over the last few years in a wide variety of fields, including text, images, audio, and video. In this talk, we discuss the use of Generative AI techniques in the realm of molecules, emphasizing the incorporation of invariances to transformation groups, and covering three applications. In the first, we show an approach to the problem of generating molecules which will bind to a particular receptor molecule, a problem common in drug design. We design specialized normalizing flows which respect the physical invariances inherent in the problem, through the use of semi-equivariant networks. In the second application, we show how to adapt diffusion models to deal with this same problem. In particular, we address the size disparity between the receptor and the generated molecule, which can be problematic for learning as the receptor can overwhelm the training; we do so by creating a small sketch of the receptor, dubbed a “virtual receptor”. In the final application, we address a problem common to chemistry, material science, and condensed matter physics: computing the quantum ground state of a molecule or material. We demonstrate an efficient method of solving the Electronic Schrodinger Equation by using a carefully designed antisymmetric normalizing flow to construct the wavefunction ansatz.
Bio:
Daniel Freedman is the Head of the Fundamental AI Research - Science group at Verily (= Google Life Sciences). In previous stints in industry, he has served as a Research Scientist at Google Research; Microsoft Research; IBM Research; and HP Labs. Prior to that, Daniel was a professor for nearly a decade, mainly at Rensselaer Polytechnic Institute (RPI), but with stops also at Bar Ilan University and as a Fulbright Fellow at the Weizmann Institute of Science. Daniel's research interests focus on AI4Science. More specifically, he is interested in: • Novel Imaging Modalities: nanoscale imaging - single molecule localization microscopy; full sound-speed inversion ultrasound; hyperspectral microscopy; theory of inverse problems. • Physics and Chemistry: design of new molecules, including for drug discovery; generating quantum correlations between photons; ab initio solution to the electronic Schrodinger equation. • Medicine: automated diagnosis and detection of disease in endoscopy and surgery; virtual staining of pathology slides; analysis of EKG and EEG signals. His older research focused primarily on the development of mathematical techniques, in particular in computational algebraic topology and mathematical methods for computer vision (PDEs, combinatorial optimization). Over the years, Daniel's work has received several honours, including the National Science Foundation CAREER Award; the Fulbright Fellowship; and several best paper awards. Additionally, he has been an investigator on a number of grants from NSF, NIH, and the US Army. Daniel received his AB in Physics from Princeton University (Magna Cum Laude, Phi Beta Kappa, Sigma Xi), and his PhD from the Division of Applied Sciences at Harvard University.
