cs.AISep 24, 2026
SaveAtelier: Learning Local Self-Supervised Features for CryoEM Volumes via Hypernetworks
Organizations: Biohub Chicago, IL · University of Chicago, Biohub Chicago, IL
Abstract
CryoEM map interpretation requires features that are spatially localized, consistent across samples, and informative across spatial scales. Most deep learning methods for map annotation extract features from fixed voxel grids. However, implicit neural representations (INRs) are able to model volumetric data as scale-agnostic, coordinate-conditioned functions. INRs are therefore attractive for cryoEM, but fitting a separate INR for each map is too expensive for large-scale feature extraction and produces representations that are not aligned across samples. We introduce Atelier, a self-supervised framework that amortizes INR fitting for reconstructed cryoEM maps. Pretrained on 5,439 Electron Microscopy Data Bank maps, Atelier is a transformer-based hypernetwork that generates high-fidelity reconstructions across a wide range of protein structures, including large multi-subunit assemblies. Beyond reconstruction, the INR generated by the pretrained transformer exposes a continuous, local feature field through its intermediate activations at any spatial query point, a property that voxel grid and patch-tokenizer architectures do not naturally provide. Used as auxiliary channels to a 3D nested U-Net annotation head trained from scratch, these coordinate-conditioned features improve performance on eight voxel-level property prediction tasks over a volume-only baseline. Our results demonstrate that amortized implicit neural representations are an effective primitive for geometry-aware analysis of cryoEM data.
Figures & tables
Figure 1: Overview. Atelier is a self-supervised framework for learning geometry-aware implicit neural representations of cryoEM density maps via hypernetworks. The generated representations expose local features at arbitrary spatial resolution, supporting downstream tasks on unseen protein structures.
Figure 2: Atelier architecture. (a) A cryoEM volume occupies a -shaped bounding box, split into non-overlapping patches of size . Since the volume is naturally sparse with most voxels being background (set to a value of ), we collect the active patches containing non-background voxels into a batch and discard the rest. These patches are fed through a convolutional neural network that transforms each patch into a volume token. The volume tokens are passed into a transformer along with learnable weight tokens; 3D rotary positional embeddings inform the volume tokens of the patch that they came from. The contextualized weight tokens output by the transformer are then projected to the input volume-conditioned weight matrices of an INR decoder , which can be queried at arbitrary spatial points in to produce a reconstructed volume . Density map of EMDB-0406 [ 37 ] visualized with ChimeraX. (b) The INR naturally exposes an interface for fine-grained, coordinate-conditioned feature extraction. At inference, the transformer produces weight matrices that parameterize the INR . For any query coordinate , we extract the post-activation hidden states from each of the INR layers (pink) and stack them into a query-conditioned feature . Because is continuous, features can be queried at sub-voxel resolution and are spatially localized to the geometry around without interpolation.
Figure 3: Atelier successfully generalizes to unseen volumes. Here we show ground truth (orange) and INR-rendered (blue) volumes for four representative samples in the test set, as well as the Fourier shell correlation curves between them. We also show close ups for the predicted densities overlaid over ground truth atomic models to demonstrate that Atelier is able to reconstruct near atomic-level detail. Note the biological diversity of structures; from left to right, the structures are a membrane-bound enzyme (EMDB-25368 [ 46 ] ), a nanobody bound to an HIV envelope glycoprotein (EMDB-23480 [ 47 ] ), a nucleosome–kinetochore complex (EMDB-7293 [ 48 ] ), and a fully assembled TCR/CD3/pMHC immune-recognition complex (EMDB-28571 [ 45 ] ).
Figure 4: Atelier generates meaningful local continuous features. (a) After pretraining, Atelier is able to generate local features invariant to global symmetries due to its geometrically-driven pretraining task in two structures from Figure 3 . Box denotes minimum, quartiles, and maximum. (b) PyMol visualization of the same two structures evaluated in part (a), with atoms colored by trilinearly interpolating voxel features projected and scaled to via PCA. Observe in the nucleosome (EMDB-7293) that the histones (center, green) primarily occupy a different color space than the surrounding DNA (purple) and centromere protein (upper left, teal). In the TCR/CD3/pMHC complex (EMDB-28571), there is considerably less variation within chains than across chains. Note, for example, the consistent coloring of the transmembrane region (bottom, purple).
| layer | trained near/far ratio | untrained near/far ratio |
|---|---|---|
| layer 1 | 18.6 [13.7, 29.9] | 1.01 [0.98, 1.02] |
| layer 2 | 25.1 [16.3, 38.3] | 0.99 [0.93, 1.03] |
| layer 3 | 34.4 [23.6, 39.9] | 0.99 [0.96, 1.03] |
| layer 4 | 29.3 [17.7, 38.4] | 1.00 [0.98, 1.03] |
Table 1: Per-layer median near/far average corruption effect (see Equation ( 3 )) over 16 holdout volumes (95% bootstrap confidence interval) for trained and untrained Atelier model. For the trained model, perturbations local to a point change intermediate activations significantly more than distant perturbations, while for the untrained model, local and distant perturbations have nearly identical effect. This indicates that Atelier activations capture local information, and that this locality is learned.
| classification (balanced accuracy ) | regression (relative error ) | |||||||
| network input | sec. struc. | amino | mol. type | mol. class. | hydro. | abs. SASA | rel. SASA | B-fact. |
| volume alone | ||||||||
| w/ coarse feats. | 0.249 | |||||||
| w/ rand. feats. | ||||||||
| w/ fine feats. | 0.797 | 0.300 | 0.976 | 0.752 | 0.386 | 0.331 | 0.363 | |
Table 2: Comparison on property prediction performance on the test set for (a) vanilla EMNUSS that takes the single-channel volume alone as input, (b) EMNUSS with inputs augmented with coarse features as additional channels, and (c) EMNUSS with inputs augmented with fine features as extra channels. Metrics are averaged over all labeled voxels in the test set; a more detailed, per-volume statistical analysis is reported in § B.4 . Augmentation with fine features outperforms the volume-alone model in eight out of eight tasks, and outperforms the coarse model in seven out of eight tasks.
Figure 5: Illustration of the family of four complexes compared in Table 3 .
| comparison | fine feature cosine similarity between volumes (mean std.) |
|---|---|
| base vs. same | 0.949 0.059 |
| base vs. homolog | 0.926 0.095 |
| base vs. different | 0.774 0.321 |
Table 3: Cosine similarities between Atelier fine features at corresponding coordinates between four different structures (mean std. over 80 sampled points). The results suggest that at corresponding locations, the activations for more similar structures are more similar than those of distantly related structures, even after spatial alignment. As negative controls, we also computed the cosine similarities between the 80 points in the base structure and the same 80 points in a completely unrelated decoy volume (EMDB-28571, a TCR/CD3 complex, cosine similarities 0.450 0.417, as well as computing the cosine similarity between a base point and one mismatched point each from the same/homologous/different structures (0.021 0.441).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Illustration of the effect of our denoising method on four uncurated cryoEM volumes. By thresholding out the “dust” present in raw deposited maps, it becomes straightforward for the tokenizer component of Atelier to only tokenize patches containing real signal.
Figure 7: The transformer-only ablation architecture. (a) As an ablation, we train a transformer-only model that directly maps per-patch volume tokens to patch reconstructions through a learnable linear projection. (b) Coarse patch-level features can be extracted by returning the transformed volume token corresponding to the patch containing a query point . Note that all query points within the same patch will have the same coarse feature.
| network input | sec. struc. | amino | mol. type | mol. class. |
|---|---|---|---|---|
| volume alone | ||||
| w/ coarse feats. | 0.227 | |||
| w/ rand. feats | ||||
| w/ fine feats. | 0.705 | 0.294 | 0.950 |
Table 4: Matthews correlation coefficient (MCC ) on the classification tasks from Table 2 . Metrics are averaged over all labeled voxels in the test set.
| Task | Comparison | (Wilcoxon) | (Holm–Bonferroni) | Win rate | Hodges–Lehmann Effect (95% CI) | |
|---|---|---|---|---|---|---|
| 1 | sec. struc. | fine > coarse | (****) | (****) | 0.642 | , [0.0069, 0.0115] |
| 2 | sec. struc. | fine > random | (****) | (****) | 0.869 | , [0.0343, 0.0389] |
| 3 | sec. struc. | fine > volume alone | (****) | (****) | 0.938 | , [0.0575, 0.0651] |
| 4 | amino | fine > coarse | (****) | (****) | 0.212 | , [ 0.0138, 0.0109] |
| 5 | amino | fine > random | (****) | (***) | 0.527 | , [0.0013, 0.0040] |
| 6 | amino | fine > volume alone | (****) | (****) | 0.954 | , [0.0255, 0.0279] |
Table 5: Per-volume comparison on fine Atelier features against baselines. We see that for amino acid prediction, coarse significantly outperforms fine (row 4, note the negative effect size), while for B-factor, fine does not significantly outperform volume-alone (row 24). For all other comparisons, the improvement is modest but significant, supporting the efficacy of Atelier-derived features.
More in this topic
3D Representation Learning
3D representation learning builds geometric features for recognizing, reconstructing, generating, and reasoning about objects and scenes.
See all 187 papers