Abstract
 

As the 3D structure prediction of proteins becomes mature after the release of AlphaFold2 from DeepMind last December, the obvious next step is to predict the movement and flexibility of the underlying protein, which will provide crucial medical insight for the design of enzymes or the development of drugs. Among the tools available, cryo-EM is a promising computational technique with high efficiency that can perform conformation analysis. However, the data characteristics of cryo-EM include strong noise, huge dimension, large sample size and unknown orientations, have made the conformation analysis very challenging. Traditional approaches address this problem at 3D level. Thus, they require the information of 3D orientations and a consensus 3D structure before starting to analyze the 3D variability, which is computationally expensive and may not be applicable to datasets whose conformations differ a lot. Therefore, there is a need to develop a new algorithm that preserves accuracy while solving the scalability issue due to the fast-growing data acquisition rate. In this talk, I will first introduce the importance of 3D structure determination of protein as well as the related background of cryo-EM image processing. Second, I will elaborate our approach to the heterogeneity problem, which utilizes the network analysis to partition the dataset into several homogeneous communities. Specifically, I will discuss several novel criteria that we used to measure the conformation similarity between cryo-EM images. Finally, I will discuss future directions in 3D conformation analysis.