CryoEM map interpretation demands features that are spatially local, consistent across samples, and informative across scales. Most deep‑learning approaches extract features on fixed voxel grids, whereas implicit neural representations (INR) model volumetric data as scale‑agnostic functions conditioned on coordinates, making them attractive for CryoEM. Fitting a separate INR for each map, however, is computationally expensive and yields representations that are not aligned across samples. Atelier introduces a self‑supervised framework that amortizes INR fitting. The framework is a transformer‑based hypernetwork pretrained on 5,439 Electron Microscopy Data Bank maps, capable of generating high‑fidelity reconstructions for diverse protein structures, including large multi‑subunit assemblies. The pretrained transformer’s INR provides a continuous local feature field via its intermediate activations at any spatial query point—a property not naturally offered by voxel‑grid or patch‑tokenizer architectures. These coordinate‑conditioned features are fed as auxiliary channels to a 3D nested U‑Net annotation head trained from scratch, improving performance on eight voxel‑level property prediction tasks compared with a volume‑only baseline. Results demonstrate that amortized implicit neural representations constitute an effective primitive for geometry‑aware CryoEM analysis.
Review: The work highlights the synergy between hypernetworks and INRs, offering a scalable path for extracting rich local features from CryoEM volumes.