Cryo-Electron Microscopy (cryo-EM) is an increasingly popular experimental
technique for estimating the 3D structure of macromolecular complexes such as
proteins based on 2D images. These images are notoriously noisy, and the pose
of the structure in each image is unknown \textit{a priori}. Ab-initio 3D
reconstruction from 2D images entails estimating the pose in addition to the
structure. In this work, we propose a new approach to this problem. We first
adopt a multi-head architecture as a pose encoder to infer multiple plausible
poses per-image in an amortized fashion. This approach mitigates the high
uncertainty in pose estimation by encouraging exploration of pose space early
in reconstruction. Once uncertainty is reduced, we refine poses in an
auto-decoding fashion. In particular, we initialize with the most likely pose
and iteratively update it for individual images using stochastic gradient
descent (SGD). Through evaluation on synthetic datasets, we demonstrate that
our method is able to handle multi-modal pose distributions during the
amortized inference stage, while the later, more flexible stage of direct pose
optimization yields faster and more accurate convergence of poses compared to
baselines. Finally, on experimental data, we show that our approach is faster
than state-of-the-art cryoAI and achieves higher-resolution reconstruction.

我们提出了一种新的方法来通过 Cryo-Electron Microscopy 技术从 2D 图像中估计蛋白质等大分子复合物的 3D 结构，该方法采用多头架构作为一种姿态编码器，以一种摊销的方式推断出每个图像中多个可能的姿态，并通过自动解码方式对姿态进行修正，实验证明我们的方法能够处理任务的初期高不确定性，快速并准确地收敛。