An adaptive Expectation-Maximization algorithm with GPU implementation for electron cryomicroscopy.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Maximum-likelihood (ML) estimation offers optimal properties for 3D cryo-electron microscopy (cryo-EM) volume reconstruction from noisy single particle images.
• Current ML implementations rely on the computationally intensive Expectation-Maximization (EM) algorithm, requiring extensive calculations over all particle orientations and positions.

Purpose of the Study:

• To present a novel strategy for accelerating the EM algorithm in ML-based cryo-EM reconstruction.
• To enhance the computational efficiency of 3D volume reconstruction from cryo-EM data.

Main Methods:

• Introduced a domain reduction strategy to optimize the EM algorithm's performance.
• Domain reduction involves using a coarse grid to identify key integration regions, followed by fine-grid evaluation within those regions.

Main Results:

• Achieved significant speedups for the EM algorithm using domain reduction.
• Simulations demonstrated speedups exceeding 10x in early iterations and over 60x in terminal iterations.

Conclusions:

• Domain reduction is an effective method for accelerating ML-based cryo-EM 3D reconstruction.
• This approach substantially reduces the computational burden of the EM algorithm, making it more practical for analyzing large cryo-EM datasets.