Efficient high-resolution refinement in cryo-EM with stochastic gradient descent.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Electron cryo-microscopy (cryo-EM) is vital for determining 3D molecular structures from 2D images.
• Efficient algorithms are critical for processing large cryo-EM datasets.
• Stochastic gradient descent (SGD) accelerates initial low-resolution reconstructions but struggles at high resolutions.

Purpose of the Study:

• To investigate why gradient descent methods fail at high resolutions in cryo-EM.
• To develop a more efficient algorithm for high-resolution cryo-EM structure determination.
• To improve the speed and flexibility of cryo-EM reconstruction and refinement.

Main Methods:

• Theoretical analysis of the optimization problem's condition number in cryo-EM.
• Development of a diagonal preconditioner using Hutchinson's diagonal estimator.
• Numerical experiments comparing preconditioned SGD with existing methods.

Main Results:

• Identified large condition numbers as a barrier to high-resolution gradient descent in cryo-EM.
• Demonstrated that a preconditioned SGD approach significantly improves convergence speed.
• Showcased the effectiveness of the diagonal preconditioner for enhancing SGD performance.

Conclusions:

• Preconditioned SGD offers a promising path towards unified and efficient cryo-EM reconstruction and refinement.
• This method can potentially overcome limitations of current state-of-the-art algorithms.
• The findings represent a significant step towards faster and more flexible high-resolution cryo-EM analysis.