Atelier Learns Local Self-Supervised Features for CryoEM Volumes with Hypernetworks

Summary

Cryo-electron microscopy map interpretation needs features that remain spatially localized, consistent across samples, and useful at multiple spatial scales. Existing deep-learning methods commonly rely on fixed voxel grids, while implicit neural representations can describe a volume as a coordinate-conditioned function. Fitting a separate implicit representation for every map, however, is computationally expensive and does not naturally align representations across samples. The paper introduces Atelier, a self-supervised framework that amortizes this fitting process with a transformer-based hypernetwork. Pretrained on 5,439 Electron Microscopy Data Bank maps, the system generates high-fidelity reconstructions across a broad range of protein structures, including large multi-subunit assemblies. Its intermediate activations also expose a continuous local feature field that can be queried at arbitrary spatial points. When supplied as auxiliary channels to a 3D nested U-Net annotation head trained from scratch, these features improve results over a volume-only baseline on eight voxel-level property-prediction tasks. The authors present the work as evidence that amortized implicit representations can support geometry-aware cryoEM analysis.