A0472
Title: Contrastive modeling for Cryo-EM 3D orientation estimations
Authors: Szu-Chi Chung - National Sun Yat-sen University (Taiwan) [presenting]
Abstract: Learning methods are becoming popular to determine the 3D structure of a protein. The most remarkable achievement is arguably the recent release of AlphaFold 2, which significantly advances the understanding of protein by performing highly accurate structure prediction from chains of amino acid sequences. However, these methods are mainly used to predict a static structure that resembles the X-ray crystal, which may not be the native state of a protein. It is noted that Cryogenic electron microscopy (cryo-EM) can capture the protein in its native states, but providing a robust initial volume for cryo-EM reconstruction is still time-consuming. The data include heavy noise, huge dimensions, and many unlabeled samples (no clean target is available for training) with unknown orientations, making it challenging to reach a robust computation. We explore the possibility of using the recent advances in contrastive modeling to directly estimate the orientations for each cryo-EM image. We elaborate on our approach, which represents the orientation using unit quaternion instead of the traditional Euler angle and several loss functions we choose. Finally, we will discuss some ongoing research, including utilizing the framework to estimate and refine the contrast transfer function parameters and the end-to-end orientation estimation procedure.
