Separating signal from noise: a self-distillation approach for amortized heterogeneous cryo-EM reconstruction

ICLR 2026 Conference Submission21169 Authors

Keywords: Cryo-EM, Generalizable 3D Reconstruction, Self-supervised learning

Abstract: Cryogenic electron microscopy (cryo-EM) has become an essential tool in structural biology for determining dynamic biomolecular structures at high resolution. However, state-of-the-art VAE-based reconstruction methods such as CryoDRGN do not generalize to new particle images: its encoder overfits to the training data due to the presence of high amounts of noise. In this work, we propose a simple yet effective strategy to generalize to images not in the training set: learning noise invariant representations. We propose Cryo-No-Overfit (CryoNOO), which extends CryoDRGN via self-distillation by leveraging the reconstruction method itself as a denoiser to generate augmented views of training images. We then learn noise-invariant representations via self-supervised learning, enabling reconstruction methods to amortize inference to unseen images. Extensive empirical evaluations on both synthetic and experimental datasets demonstrate that our method dramatically improves reconstruction quality on unseen test data, marking a key step towards robust, generalizable cryo-EM reconstruction.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21169