Can ML enable 100-Plex cryo-EM structure determination? (www.owlposting.com)

Princeton CS professor Ellen Zhong, a pioneer in applying deep learning to cryo-electron microscopy (cryo-EM), discusses in a recent podcast how ML methods—most notably her VAE-based CryoDRGN family—are pushing cryo-EM from static 3D reconstructions toward reconstructing distributions and continuous motions of macromolecules. A central theme is the possibility of multiplex or “100‑plex” cryo-EM: scaling structure determination to analyze many distinct proteins or conformational states from mixed samples, enabling high-throughput structural screens and richer dynamic ensembles. Zhong also described work during a sabbatical at Generate Biomedicines and expanding efforts into cryo-ET and NMR to broaden experimental contexts. Technically, the core challenge remains inferring unknown particle poses and heterogeneous conformations from thousands of noisy 2D projections; ML models like CryoDRGN treat the problem as learning a latent distribution over 3D volumes, enabling ab initio reconstruction and continuous heterogeneity modeling. Practical limits include sample prep quality, dataset availability, and validation/interpretability for wet‑lab users. Zhong highlights synergy with protein prediction models (AlphaFold family) as priors but stresses experiments remain essential for large complexes and dynamics. The implications: scalable ML-driven cryo-EM could accelerate discovery, improve training data for next‑gen structure predictors, and move structural biology toward routine, high‑throughput mapping of functional machinery.