Supervised cryo-EM tomography denoising enabled by physics-based image simulation

Abstract

Cryo-electron tomography enables three-dimensional visualization of biomolecules in their native aqueous state but suffers from low signal-to-noise ratio (SNR) due to electron dose limitations. Subtomogram averaging improves resolution but averages out structural heterogeneity, while self-supervised denoising risks over-smoothing fine details. Supervised learning has not been widely applied to cryo-ET because of the challenge of generating realistic training data. Here, we present a physics-informed supervised denoising framework based on a dedicated simulation pipeline. Microscope parameters, detector noise, and electron dose were extracted from experimental data, and clean images were generated through multislice calculations that solve the quantum-mechanical propagation of the electron wave function. Paired with noisy counterparts, these images were used to train a U-Net convolutional neural network. Applied to simulated and experimental tilt-series of liposomes, the method improved SNR by up to 132% and enhanced bilayer contrast without averaging, enabling recovery of fine structural details under low-dose conditions.