Atelier hypernetwork amortizes INR fitting across 5,439 EMDB maps for cryoEM feature extraction

The framework replaces per-map INR optimization with a single pretrained transformer that emits weights for a coordinate-based decoder. This yields aligned, scale-agnostic feature fields at any query point, a property absent from fixed voxel grids or patch tokenizers used in prior cryoEM annotation models. Training used only reconstruction loss on deposited maps; no task labels were required.

Evaluation compared the auxiliary INR features against a volume-only baseline on eight downstream voxel classification and regression tasks. The hypernetwork-derived channels produced consistent gains, confirming that amortized implicit representations capture local geometry more effectively than discrete grids. The approach directly addresses the cost and alignment problems that have limited INR adoption in large-scale EMDB analysis.

Related work on hypernetworks for shape generation and on implicit representations for protein density maps shows the same amortization pattern reduces per-sample fitting time by orders of magnitude while preserving reconstruction fidelity. Atelier extends this pattern to cryoEM by exposing intermediate activations as spatially continuous features rather than global latents.

Operational deployment on new depositions requires only a forward pass through the frozen transformer followed by standard U-Net training, lowering the barrier to geometry-aware annotation pipelines in structure-guided drug design workflows.