These projects form a connected AI-for-cryo-EM ecosystem for turning noisy particle data into molecular structure discovery. Together, they cover the data infrastructure, foundation-model representations, heterogeneous reconstruction methods, and automated workflows needed to make cryo-EM analysis more scalable and reproducible.
Cryo-EM particle images are noisy, heterogeneous, and difficult to interpret one by one. Designed and released the Cryo-IEF ecosystem, pretrained on approximately 65 million cryo-EM particle images to learn transferable representations for structural classification, pose clustering, particle-quality assessment, and automated reconstruction workflows.
Real biological samples often contain multiple molecular species or continuous conformational motion rather than a single static structure. Developed a prior-guided heterogeneous reconstruction framework that uses Cryo-IEF representations to reduce random ab initio initialization and disentangle compositional classification from 3D reconstruction.
Deep learning methods can only become useful in cryo-EM when experimental data and metadata can be moved reliably into training and inference workflows. Created the open-source data-processing layer used by Cryo-IEF, CryoDECO, and CryoWizard, with support for MRC/MRCS preprocessing, LMDB-backed datasets, Fourier/Hartley feature generation, balanced sampling, data loading, and CryoSPARC-to-RELION metadata conversion.
Cryo-EM structure determination traditionally requires many expert choices across data-processing stages. Built and extended an end-to-end computational pipeline integrating CryoRanker with CryoSPARC, streamlining processing from raw movies, micrographs, or particles to high-resolution 3D volumes through command-line, web, and browser-extension interfaces.
For additional public repositories and contributions, see github.com/yanyang1998.