Generative Priors for Cryo-EM Image Reconstruction

Single-particle cryo-electron microscopy (cryo-EM) is the premier technique for determining 3D biomolecular structures, yet its reliance on hundreds of thousands of high-fidelity 2D images creates substantial data throughput bottlenecks. To address this, we formulate the recovery of full-resolution cryo-EM images from compressively sampled measurements as an inverse problem, solved via posterior sampling under a learned generative prior. By training a denoising diffusion probabilistic model (DDPM) on EMPIAR datasets, we capture the low-dimensional manifold of protein images and accurately reconstruct them from both spatial and Fourier-domain undersampled data. Our approach successfully recovers 2D images at compression factors up to 2× while strictly preserving the biological signal required for downstream structural analysis, including conformational heterogeneity identification and atomic model building. Ultimately, this work demonstrates that generative diffusion priors can decode highly compressed measurements without sacrificing the high-resolution biological signal necessary for structural biology, offering a robust computational pathway to accelerate cryo-EM workflows.