Owl Posting Can machine learning enable 100-plex cryo-EM structure determination? (Ellen Zhong, Ep #5)
14 snips
Nov 10, 2025 
Ellen Zhong, a computer science professor at Princeton University, is revolutionizing molecular imaging through deep learning applications in cryo-EM. She discusses how her lab addresses challenges in reconstructing 3D structures from 2D images. Ellen also explores the use of machine learning to uncover previously hidden structures and the ongoing issues of heterogeneity in protein analysis. She emphasizes the importance of improving algorithms alongside experimental data collection for effective results, and reflects on the future of cryo-EM in biomedical applications.
AI Snips
Chapters
Transcript
Episode notes
Deep Learning Fits Cryo-EM Dynamics
- Ellen Zhong saw cryo-EM continuous-motion reconstruction as a juicy ML problem and applied deep learning early in her PhD.
- She believed deep models could learn complex distributions from large-scale particle-image data to infer 3D structure and dynamics.
Reconstruction Requires Joint Pose And Structure
- The central computational challenge is jointly inferring 3D structure and unknown camera poses from noisy 2D projections.
- Modeling continuous conformational ensembles lets you reconstruct molecular movies rather than a single static volume.
Match Algorithms To Sample Quality
- For well-behaved, purified samples, existing ab initio methods can produce structures quickly given high-quality data.
- But when samples are dynamic or mixed, jointly designing experiments and algorithms becomes necessary.