Development of Mathematical Methods for 3D Reconstruction in Cryogenic Electron-Microscopy

For the past several years, advances in cryogenic-electron microscopy (cryo-EM)have led to a revolution in structural biology, for which the Nobel Prize in Chemistrywas awarded in 2017. To enable the 3D reconstruction of atomic structures, thedevelopment of mathematical methods has been instrumental to address thesample size, noise, and geometric complexity inherent with cryo-EM data. Tosignificantly improve the 3D reconstruction and reduce its computational burden,we propose to perform variational inference of the rotation associated with eachimage, by implementing a variational auto-encoder neural network. A specificity ofthe network is that it will include differential geometric constraints on the latentspace characterizing the action of the 3D rotation group. For this project, thestudent will first get familiar with an earlier implementation of the method for 2Drotation space developed by our collaborators, Dr. Miolane (UCSB) and Dr. Poitevin(Stanford), before extending it to 3D space.

Faculty Supervisor:

Khanh Dao Duc

Student:

Partner:

CentraleSupélec

Discipline:

Mathematics

Sector:

Education

University:

The University of British Columbia

Program: