ACE-EM: Boosted ab initio Cryo-EM 3D Reconstruction with Asymmetric Complementary Autoencoder

Lin Yao, Ruihan Xu, Guolin Ke, Zhifeng Gao, Yuhang Wang

Keywords: autoencoder, electron microscopy, cryo-EM, 3D reconstruction, pose inference, asymmetric complementary autoencoder

Abstract: Cryo-electron microscopy (cryo-EM) is an imaging technique for obtaining high-resolution biomolecular structures. The central problem in cryo-EM is to recover the underlying 3-dimensional (3D) objects from 2-dimensional (2D) projection images. Aside from signal corruptions and extremely low signal-to-noise ratio, a major challenge in cryo-EM 3D reconstruction is to estimate the poses of 3D objects during the projection image formation, which are missing from experimental measurements. Recent methods attempted to solve the pose estimation problem using the autoencoder architecture. A key issue with this approach is that the latent vector is only indirectly updated through the decoder. The encoder's learning of the pose space can be easily trapped in a local subspace, resulting in suboptimal pose inferences and inferior 3D reconstruction quality. Here we present a modified autoencoder architecture called ACE (asymmetric complementary autoencoder) and designed the ACE-EM method to solve this issue, which consists of two tasks. The first task takes projection images and outputs predicted images using an image encoder followed by a pose decoder. The second task reverses the order of encoder and decoder, which takes randomly sampled poses and outputs predicted poses. The two tasks complement each other and can achieve a more balanced training of the encoder-decoder parameters. Compared to other methods, ACE-EM can reach higher pose space coverage within the same training time and has achieved state-of-the-art 3D reconstruction results for several benchmark datasets.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )

TL;DR: 3D cryo-EM reconstruction with ACE-EM