RESEARCH · RESEARCH · #1110
Atelier: transformer hypernetwork that amortizes INRs for cryoEM volumes (arXiv:2609.30569v1)
Atelier is a self-supervised, transformer-based hypernetwork pretrained on 5,439 EMDB maps that amortizes fitting implicit neural representations (INRs) for cryoEM volumes. The generated INRs provide continuous, coordinate-conditioned local feature fields; when used as auxiliary channels for a 3D nested U-Net, these features improve performance on eight voxel-level property prediction tasks versus a volume-only baseline.
KEY POINTS
- Atelier is a self-supervised, transformer-based hypernetwork pretrained on 5,439 EMDB maps that amortizes fitting implicit neural representations (INRs) for cryoEM volumes.
- The generated INRs provide continuous, coordinate-conditioned local feature fields; when used as auxiliary channels for a 3D nested U-Net, these features improve performance on eight voxel-level property prediction tasks versus a volume-only baseline.
- This work demonstrates that amortized INRs via a transformer hypernetwork can yield aligned, coordinate-conditioned local features that scale across diverse cryoEM structures and improve voxel-level annotation.
WHY IT MATTERS
This work demonstrates that amortized INRs via a transformer hypernetwork can yield aligned, coordinate-conditioned local features that scale across diverse cryoEM structures and improve voxel-level annotation.