Ge Liu
Talk: Generative AI for biomolecule engineering and autonomous discovery
The rational design and optimization of functional biomolecules—from therapeutic antibodies to novel enzymes—offers transformative potential but presents unique challenges. These problems inherently involve both discrete variables (protein sequences) and continuous 3D structures that exist on non-Euclidean manifolds (e.g., the rotational group SO(3)). Conventional design processes remain slow, fragmented, and unable to efficiently navigate the vast and complex molecular landscape. The grand challenge lies in making the leap to autonomous generative design: creating novel, high-performance molecules that nature has not yet imagined with the power of generative AI.
This talk will present our work on developing powerful generative AI framework for next-generation discovery engine. We will first introduce our recent advances in flow-matching and diffusion models, including Statistical Flow Matching (SFM) a novel generative framework that leverages the Riemannian geometry of statistical for high-fidelity discrete generation and our Riemannian Consistency Models (RCMs), which dramatically accelerate the design phase by enabling high-quality generation in just a few steps.
Furthermore, we will show how to make these powerful generators can be made steerable and self-improving. We present OC-Flow, a universal framework based on Stochastic Optimal Control that provides training-free guidance for solving inverse problems and optimizing for desired properties. To close the DBTL loop, we will discuss how our advanced reinforcement learning algorithms, including Wasserstein-regularized Online Reward Weighting (ORW-W2) and Adaptive Divergence Regularized Policy Optimization (ADRPO), enable our models to self-improve from feedback. By integrating these specific innovations, we are building a truly autonomous bioengineering engine, paving the way for the accelerated discovery of next-generation therapeutics and catalysts.
BIO:
Dr. Ge Liu is an Assistant Professor in the Siebel School of Computing and Data Science at the University of Illinois Urbana-Champaign. Her research develops scalable, reliable, and geometry-aware deep generative models and sequential optimization techniques to solve critical problems in synthetic biology, immunology, and molecular biology. Her work aims to move beyond predictive modeling towards the autonomous, de novo design of functional molecules.
She specializes in geometry-aware multimodal flow matching, diffusion models, and reinforcement learning to computationally design novel antibodies, enzymes, catalysts, and vaccines. Dr. Liu received her Ph.D. from MIT’s EECS department, where her thesis won the George M. Sprowls Thesis Award in AI and Decision-Making. She conducted her postdoctoral research with Dr. David Baker at the Institute for Protein Design. Her long-term vision is to create autonomous AI systems that accelerate the entire discovery cycle for next-generation therapeutics and sustainable biotechnologies.