Amortized Inference for Heterogeneous Reconstruction in Cryo-EM.

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryo-electron microscopy (cryo-EM) offers critical insights into biomolecular dynamics.
• Accurate 3D structure and conformational heterogeneity determination from noisy 2D projections remains computationally challenging.
• Existing methods struggle with efficient pose estimation and analysis of dynamic biological molecules.

Purpose of the Study:

• To develop a computationally efficient method for *ab initio* heterogeneous reconstruction in cryo-EM.
• To enable joint estimation of poses and conformational heterogeneity without computationally expensive pose search.
• To analyze dynamic information from experimental cryo-EM datasets.

Main Methods:

• Introduced cryoFIRE, an amortized framework for *ab initio* heterogeneous reconstruction.
• Employed an encoder-decoder architecture for joint pose and conformation estimation.
• Utilized a physics-based decoder to create an implicit neural representation of conformational space.

Main Results:

• Achieved a one order of magnitude speedup on large cryo-EM datasets (millions of images) without accuracy loss.
• Validated that pose and conformation estimation can be amortized across dataset size.
• Demonstrated the capability of an amortized method to extract interpretable dynamic information from experimental data.

Conclusions:

• cryoFIRE significantly enhances computational efficiency in cryo-EM data processing.
• The method successfully addresses the challenge of analyzing conformational heterogeneity and unknown poses.
• This work represents a breakthrough in extracting dynamic biomolecular information using amortized deep learning approaches in cryo-EM.