Self-supervised cryo-EM map encoder whose transformer hypernetwork emits a neural field, giving local features at any 3D coordinate.
No providers recorded yet. Browse all providers
Most deep learning tools for annotating a cryo-EM map treat it as a voxel grid: a 3D U-Net reads the lattice and predicts a label per voxel. An implicit neural representation (INR), a small neural network that maps any 3D coordinate to a density value, describes the same map as a continuous function, and its internal activations at a point are a natural description of the local geometry there. But fitting one INR per map takes hundreds to thousands of optimization steps, and independently fitted INRs share no common feature space.
Atelier, developed at Biohub with the University of Chicago and released as a preprint in September 2026, removes both problems by amortizing the fit. A transformer hypernetwork reads a cryo-EM map and outputs, in a single forward pass, the weights of an INR that re-renders that map. Because every INR is produced by the same shared hypernetwork, their hidden-layer activations live in one aligned space across volumes. Querying the generated INR at any coordinate and stacking its hidden activations yields a feature vector tied to the density around that point, at whatever resolution the caller asks for.
Atelier is designed to augment voxel-grid annotators rather than replace them. Unlike neural-field methods such as CryoDRGN-AI, which reconstruct density from particle images, it starts from a finished map and learns transferable features from it.
Maps are denoised, scaled to [0, 1], resampled to 3 Å voxels and padded to a 96³ box. A 3D convolutional tokenizer encodes each non-empty 6³ patch as a volume token; empty patches are dropped, so a median map uses 236 of 4,096. These tokens, together with learned weight tokens (one per INR output neuron), pass through a 24-layer transformer encoder (latent dimension 256, 8 heads) with 3D rotary position embeddings. The transformed weight tokens are linearly projected into the weight matrices of a ReLU MLP that takes Fourier-encoded coordinates and returns density; biases are learned but not generated. The feature for a point is the stack of post-activation hidden states across the INR's layers, each 256-dimensional. The model has 38.3 million trainable parameters. The loss adds a high-frequency-weighted Fourier error to real-space mean squared error.
Atelier was pretrained on 5,439 maps from the Electron Microscopy Data Bank, drawn from the Cryo2StructData set of maps with fitted atomic models (the data behind Cryo2Struct), with 684 validation and 689 test maps. Pretraining took about 720 H100-hours. On the 689 held-out maps the mean normalized area under the Fourier shell correlation curve was 0.634. Frozen Atelier features were then added as input channels to EMNUSS, a nested 3D U-Net retrained from scratch on eight voxel-level tasks including secondary structure, amino acid identity, solvent-accessible surface area and B-factor. The augmented model beat the volume-only baseline on all eight, for example raising secondary-structure balanced accuracy from 0.729 to 0.797, and beat a capacity-matched transformer-only autoencoder on seven of eight.
Atelier is a feature extractor for structural biologists and method developers who build automated map-interpretation tools: secondary-structure and residue assignment, solvent-exposure estimation, and segmentation of a map into chains and domains. Features are precomputed once per map and appended to an existing voxel network. Because features at matched coordinates stay similar across related complexes, including mouse and human C3aR–Go structures absent from training, the representation also supports comparing structures within a protein family. The authors also point to interface and binding-site prediction in large assemblies.
Atelier is an arXiv preprint and has not been peer reviewed. The preprint links no code or pretrained weights. Its contribution is showing that a hypernetwork can serve as a reusable pretrained encoder for reconstructed maps whose neural-field activations carry structural information beyond the density itself. The authors' ablation attributes most of the downstream gain to self-supervised transformer pretraining, with the INR adding a smaller further improvement plus the ability to query at arbitrary coordinates. Splits are random over EMDB IDs, so homologs can straddle train and test, and maps larger than a 96³ box at 3 Å were excluded, limiting claims about new folds and very large assemblies.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.