CryoGAN: A New Reconstruction Paradigm for Single-Particle Cryo-EM Via Deep Adversarial Learning

We present CryoGAN, a new paradigm for single-particle cryo-electron microscopy (cryo-EM) reconstruction based on unsupervised deep adversarial learning. In cryo-EM, the structure of biomolecule needs to be reconstructed from large set noisy tomographic projections with unknown orientations. Current techniques are marginalized maximum-likelihood formulation that requires calculations over all possible poses each projection image, computationally demanding procedure. Our approach is seek 3D has simulated match real data in distributional sense, thereby sidestepping pose estimation or marginalization. prove that, an idealized mathematical model this results recovery correct structure. Motivated by distribution matching, we propose specialized GAN consists structure, cryo-EM physics simulator, and discriminator neural network. During reconstruction, optimized so its obtained through simulator resemble (to discriminator). Simultaneously, trained distinguish projections. CryoGAN takes as input only images imaging parameters. It involves neither prior training nor initial currently achieves 10.8 Å resolution realistic synthetic dataset. Preliminary experimental β-galactosidase 80S ribosome demonstrate ability exploit statistics under standard conditions. believe opens door family novel likelihood-free algorithms reconstruction.