SMBp Group Meeting: Minhuan Li (Harvard)

Tuesday Jul 11, 2023, 12:00 PM → 1:00 PM America/New_York

3rd Floor Classroom/3-Flatiron Institute (162 5th Avenue)

Speaker: Minhuan Li  (Harvard)

 

Topic: Towards automatic crystallographic structure refinement with deep generative models

Abstract: Proteins mediate most biochemical transformation in our bodies. Like man-made machines, they do so by moving through cycles of configurations. To intervene in protein function when things go wrong, we need to understand these configurations. Machine learning is making rapid strides in the prediction of biomolecular structures and their conformational ensembles. X-ray crystallography, on the other hand, has enabled experimental structure determination for many biomolecules. Spurred on by powerful new X-ray Free Electron Lasers and synchrotron beamlines, X-ray crystallography has also begun to enable visualization of the dynamics of proteins on timescales from femtoseconds to seconds. However, the lack of interfaces between crystallographic data and machine learning methods prevents the application of modern deep learning frameworks to crystal structure determination. I will present SFCalculator, a differentiable pipeline to generate crystallographic observables from atomistic molecular structures with a bulk solvent correction, bridging the rich crystallography data and fast-evolving neural-network- based molecular sampling methods. We validate SFCalculator by benchmarking against conventional calculations and illustrate its potential using two examples: performing molecular replacement in rotation and translation space, and optimizing atomic models in the physics-informed latent space constructed by a deep generative model.