Physics-Inspired Modeling for Molecular and Protein Structure: Prediction, Ranking, and Generation
Speaker:
Dr. XU Yuzhi
Ph.D. Candidate, Department of Chemistry, New York University
Title:
Physics-Inspired Modeling for Molecular and Protein Structure: Prediction, Ranking, and Generation
Abstract:
As deep learning is increasingly deployed to molecular and protein systems, it must operate beyond the training distribution and respect the physical laws that govern real-world chemistry, calling for new algorithmic principles that integrate domain knowledge into AI pipelines. In this talk, I will present my research on physics-inspired deep learning methods for predicting, ranking, and generating molecular and protein structures.
This talk will highlight three representative lines of my work. The first work focuses on feature-level integration, where multi-level chemical priors are encoded into an E(3)-equivariant network, EnviroDetaNet, for infrared spectra prediction, yielding substantial gains in data efficiency and out-of-distribution generalization on molecules up to three times larger than those in the training set. The second work discusses how physics-inspired architecture-level design can capture the weak, accumulated interactions that shape protein interfaces, and introduces SAKE-PP, a kinetic graph network inspired by Langevin dynamics. SAKE-PP is an independent ranker that outperforms AlphaFold 3’s ranking score on protein–protein complex selection. The third work introduces inference-level guidance, where ProSteer injects a true atomic force field into diffusion sampling without retraining, enabling physically realistic and condition-dependent generation, such as pH-driven conformational transitions that AlphaFold 3 alone cannot capture. This talk will also discuss future directions in agentic molecular dynamics, closed-loop AI-driven materials discovery, and a broader vision for trustworthy, physics-grounded AI4Science.
Biography:
Yuzhi Xu is a final-year Ph.D. candidate at New York University, supervised by Prof. John Zhang, with a Ted Keusseff Fellowship nomination. He also serves as the AI Project Lead in a biology laboratory at the National University of Singapore and as a member of OpenADMET at Memorial Sloan Kettering Cancer Center. His research focuses on physics-inspired deep learning for molecular and protein modeling, spanning spectra prediction, protein complex ranking, and structure generation. He ranked Top-10 in the 6th CACHE Drug Discovery Challenge, and his AI4Science active-learning U.S. patent was featured by Xinhua News Agency.
Enquiries:
Ms. FUNG Wing Chi Mary (maryfung@cse.cuhk.edu.hk)
Mr. WONG O Bong (obong@cse.cuhk.edu.hk)
** ALL ARE WELCOME **