Amortized template matching of molecular conformations from cryoelectron microscopy images using simulation-based...

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Understanding biomolecular functions requires characterizing conformational ensembles.
• Cryo-electron microscopy (cryo-EM) provides 2D snapshots but suffers from noise and unknown orientations, complicating conformation identification.
• Traditional methods struggle with computational cost and parameter estimation.

Purpose of the Study:

• To introduce cryo-EM simulation-based inference (cryoSBI) for inferring biomolecular conformations and their uncertainties from individual cryo-EM images.
• To enable efficient and accurate analysis of large cryo-EM datasets.
• To provide interpretable machine learning models for reliable structural analysis.

Main Methods:

• CryoSBI integrates physics-based simulations with probabilistic deep learning for Bayesian inference.
• A neural network is trained on simulated cryo-EM images derived from structural hypotheses (templates).
• The trained network infers conformations from experimental images rapidly, bypassing explicit likelihood calculations and particle pose estimation.

Main Results:

• CryoSBI accurately infers biomolecular conformations and associated uncertainties from cryo-EM images.
• The method significantly enhances computational speed, processing images in milliseconds after a one-time training.
• Interpretable machine learning models are generated by combining physics-based simulations with deep neural networks.

Conclusions:

• CryoSBI offers a computationally efficient and reliable approach for analyzing single-particle cryo-EM data.
• The method facilitates direct analysis on micrographs, suitable for large-scale structural biology studies.
• CryoSBI advances the interpretation of cryo-EM data by providing transparent and trustworthy conformational insights.