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

Abstract

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

UVTomo-GAN: An adversarial learning based approach for unknown view X-ray tomographic reconstruction

“…We compare the results of UVTomo-GAN with unknown PMF on four baselines, 1) UVTomo-GAN with known PMF, 2) UVTomo-GAN with unknown PMF but fixing it with a Uniform distribution during training, 3) TV regularized convex optimization, 4) expectation-maximization (EM). In the first baseline, similar to [17], we assume that the ground truth PMF of the projection angles is given. Thus, in Alg 1, we no longer update p (step 9).…”

“…Similar to [17], we choose Wasserstein GAN [19] with gradient penalty (WGAN-GP) [20]. Our loss function and the mini-max objective for I, p and φ are defined as, Algorithm 1 UVTomo-GAN Require: α φ , α I , α p : learning rates for φ, I and p. n disc : the number of iterations of the discriminator (critic) per generator iteration.…”

“…Our approach benefits from the proven potential of generative adversarial networks (GAN) [16] to recover the image and projection angle distribution that match the given projection measurements in a distribution sense. Our approach is mainly inspired by CryoGAN [17]. Unlike CryoGAN, we have a more challenging setting, as we assume that the distribution of the projection angles is unknown.…”

MSR-GAN: Multi-Segment Reconstruction via Adversarial Learning

“…Our method is unsupervised as it only relies on the given observations and does not use large paired datasets for training. Similar to [1], our method aims to find x and p such that the distribution of the partial noisy measurements generated from (1) matches the real measurements {ξ j real } N j=1 . To this end, we use a generative adversarial network (GAN) [17].…”

“…Therefore, we train a generator discriminator pair, where the discriminator tries to distinguish between the measurements output by the generator and the real ones. Our approach is inspired by CryoGAN [1] in which the goal is to reconstruct a 3D structure given 2D noisy projection images from unknown projection views. Unlike CryoGAN, we assume the distribution of the latent variables, i.e.…”