Cryo-electron microscopy (cryo-EM) builds 3D models of proteins from hundreds of thousands of 2D snapshots. Each snapshot is extremely noisy, with signal often below one percent of the noise, and each is taken from an unknown viewing angle. Many samples also contain a mix of molecular shapes. Current methods describe that mix with a flat, low-dimensional code space, the "latent space", and assign every image a code. When an image is too noisy to tell two closely related shapes apart, a flat space has no sensible place to put it. The image lands at some arbitrary point between or inside clusters, and its wrong shape assignment then corrupts the estimate of its viewing angle.
We propose to place the shape code on a hyperbolic space instead, while the viewing angle stays on its natural rotation group. Hyperbolic space is a negatively curved geometry in which distance from the center behaves like depth in a tree. Points near the center stand for "ancestor" shapes shared by many states, and points near the edge stand for specific sub-states. The core bet is that, under heavy noise, this depth coordinate becomes a learned confidence. Images that only carry enough signal to fix the coarse assembly state should settle at internal tree nodes, and only clean images should reach the leaves. For assembly processes that really are tree-shaped, such as ribosome biogenesis, this should keep clustering and per-image reconstruction from collapsing as noise grows.
Core Claim
The flat-latent assumption is shared by the continuous heterogeneity methods in the file, cryoDRGN and CryoDRGN-AI. Both decode a shape code $z \in \mathbb{R}^D$ through a coordinate network into a 3D density. The CryoDRGN-AI authors state that giving these latents a geometric interpretation is still open. CryoBench documents a "noise wall": most methods collapse near SNR 0.001. It also reports that, under extreme noise, methods confuse differences in viewing angle with differences in shape.
Our claim is specific and falsifiable. On datasets whose states form a hierarchy, a shape code on the Lorentz hyperboloid, trained jointly with viewing-angle search, will do three things. First, hyperbolic radius will order images by how much shape information they carry, and this ordering should track per-image SNR. Second, it will degrade more gracefully than a flat code in both clustering accuracy and per-image resolution as SNR drops toward 0.001. Third, it will reduce angle errors caused by shape mis-assignment, because ambiguous images are reconstructed against a near-consensus "ancestor" volume rather than a wrong leaf. On data whose variation is not tree-shaped, we expect no gain or a small loss, and we treat that as a required control rather than an embarrassment.
The hyperbolic machinery comes from the representation-learning literature. Poincaré embeddings showed that trees embed in very few hyperbolic dimensions: a 5-dimensional ball reached 0.823 mean average precision on WordNet, against 0.168 for 200 Euclidean dimensions. The Lorentz-model work supplies the parameterization we actually use. It has closed-form distances and avoids the numerical blow-up near the edge of the Poincaré ball. These papers motivate the geometry and supply the method, but they never touched images, physics-based forward models or unknown poses.
APo-VAE reported that a hyperbolic latent space avoided "posterior collapse", where a generative model learns to ignore its code. It kept 32 active latent units where a Euclidean model kept 2. We treat this only as weak motivation. That collapse came from strong text decoders, not from measurement noise, so we do not assume it transfers.
Research Program
The model keeps the parts of CryoDRGN-AI that are known to work on experimental data and changes one component.
The viewing angle stays on SO(3), the group of 3D rotations, together with an in-plane shift. It is found with CryoDRGN-AI's hierarchical pose search, a coarse grid search refined by gradient descent. We deliberately avoid a fully amortized pose encoder, the approach in which a single network predicts each image's pose directly. ACE-EM reports that such an encoder, as in cryoAI, failed to converge on EMPIAR-10028. This respects the geometric split noted in the file: viewing angle is positively curved and compact, while assembly heterogeneity is plausibly negatively curved. Unlike the quaternion-embedding approach, we do not force everything onto one curved space.
The shape code lives on a $d$-dimensional Lorentz hyperboloid with learnable curvature, and we test $d$ from 2 to 8. An image encoder outputs a mean point on the hyperboloid and a spread in the tangent space at that point. The approximate posterior is a wrapped normal: a Gaussian drawn in the flat tangent space and mapped onto the curved surface. This is the same construction APo-VAE used on the Poincaré ball. Its density is available in closed form, so we can train with an ordinary evidence lower bound instead of APo-VAE's adversarial objective.
The prior is a wrapped normal centered at the hyperboloid origin. The decoder is the cryoDRGN coordinate network, fed the hyperboloid's ambient coordinates. The network is unchanged, so it can represent anything it could before. The hyperboloid changes only the geometry of the code that sets distances, the prior and interpolation paths.
The one new coupling ties hyperbolic depth to the frequency curriculum. Training starts on low-pass-filtered images, following ACE-EM's warm-up, so early on only coarse shape information is available. We cap the hyperbolic radius at a schedule that grows with the frequency cutoff. The tree is thus forced to resolve from root to leaves in step with the data resolution. We also test an uncapped version, to see whether depth-as-confidence emerges without the cap.
For analysis, we read off the learned hierarchy by hierarchical clustering on hyperbolic distances. We then check radius against per-image SNR, against the defocus value that controls the microscope's image distortion, and against ground-truth angle error.
Baselines
The naive reference point is a single homogeneous reconstruction with CryoDRGN-AI's pose search and no shape code. Any heterogeneity method must beat this in per-image resolution, otherwise its code is only fitting noise.
The strongest existing competitor on the main benchmark is RECOVAR. CryoBench reports an adjusted Rand index of 0.968 for it on Ribosembly, a clustering-agreement score where 1 is perfect. We run it with the CryoBench protocol. A clearly reported RECOVAR value does not exist for the low-SNR variants, so we rerun RECOVAR ourselves on every noise level. Beating it means at least matching 0.968 at the standard CryoBench noise and holding a statistically clear margin, over three seeds, at SNR 0.01 and 0.001.
The closest architectural baseline is CryoDRGN-AI itself, with its flat Euclidean code at the same latent dimension and at 4 times that dimension. No Ribosembly number for it appears in the file. For this and the other methods without quoted figures, no reported result was found on Lacuna, so all are run by us. All methods get the same particle stacks and the same viewing-angle initialization protocol (ab initio, no ground-truth poses). They also get the same GPU-hour budget and an equal hyperparameter search of twenty configurations each. These shared conditions are what make the comparison fair.
Three ablations isolate the key ingredient. In the first, we set curvature to zero in the same code, which recovers the Euclidean model exactly. Any gain then comes from curvature, not implementation. In the second, we keep a Euclidean code but use the encoder's posterior variance as the confidence signal, so the confidence is explicit rather than geometric. In the third, we train the Euclidean model and only afterwards fit a Lorentz embedding to its latents. This separates "hyperbolic geometry shapes learning" from "hyperbolic geometry helps visualization." The radius-cap curriculum is ablated separately.
Why not just…?
Why not just increase the Euclidean latent dimension? The WordNet results in Poincaré embeddings and the Lorentz model suggest that extra flat dimensions are a poor substitute for curvature when the structure is a tree. But those results concern clean graph embedding, not noisy images, so we do not take them as settled. The 4-times-wider CryoDRGN-AI baseline tests this directly. If it matches the hyperbolic model on Ribosembly at SNR 0.001, the geometry claim fails.
Why not just use the posterior variance a Euclidean VAE already outputs as confidence? This is the most serious threat. A flat model with a variance head may express uncertainty perfectly well. Our argument is that a flat code has no natural "ancestor" location: an uncertain image's mean still sits near some specific state, and the decoder renders that state. The second ablation settles the question. If variance-weighted Euclidean codes give the same ARI, per-image FSC and angle-error reduction, the proposal reduces to "report uncertainty," and we would say so.
Why not cluster the Euclidean latent hierarchically after training, or use discrete 3D classification? CryoBench reports that continuous methods "significantly outperformed" traditional 3D classification on Ribosembly. Discrete classification also needs the number of classes fixed in advance. Post-hoc hierarchical clustering of flat latents is cheap and plausible, which is why it is the third ablation. It cannot repair images whose codes were already pushed into the wrong cluster during training, and per-image FSC will show whether that matters.
Why not just rely on low-pass curricula like ACE-EM for robustness? We use that curriculum in every method, so it is a shared ingredient rather than a competitor. ACE-EM was evaluated only on homogeneous samples. Any remaining gap at low SNR is therefore attributable to the shape code.
Proposed Experiments
The primary benchmark is CryoBench Ribosembly: 16 assembly states of a bacterial ribosome. We regenerate it at SNRs from the standard level down to 0.001. Because the states differ by which subunits are present, we derive a ground-truth hierarchy from shared subunit composition. Beyond ARI and per-image Fourier shell correlation (FSC, a standard resolution measure), this lets us score tree recovery as agreement between the learned and true hierarchies. We also report viewing-angle error against ground truth and its breakdown by whether an image's shape was mis-assigned.
Tomotwin-100, a mixture of 100 distinct complexes on which CryoBench says most ab initio methods fail, tests whether a coarse-to-fine shape tree eases angle search in large mixtures. IgG-1D and IgG-RL serve as non-hierarchical controls, since their motions are continuous and not tree-like. The prediction there is no gain, and a clear loss would bound where the method should be used. CryoBench's neighborhood-similarity metric checks whether hyperbolic codes distort these continuous motions.
On experimental data, we run EMPIAR-10076, the E. coli large ribosomal subunit. cryoDRGN found sub-clusters within the major states there, so it is a natural test of whether those sub-clusters appear as children of shared parents. EMPIAR-10028 serves as a homogeneous sanity check, where ACE-EM reports 7.54 Å at Nyquist. A hyperbolic code should not degrade that result, and its points should stay near the origin.
Three outcomes would be decisive. The first is whether, at SNR 0.001 on Ribosembly, the hyperbolic model beats both the zero-curvature model and the variance-as-confidence ablation in ARI and per-image FSC. The second is whether hyperbolic radius correlates with per-image SNR more strongly than Euclidean posterior variance does. The third is whether the angle-error gap shrinks for images the model places at internal nodes. If the second holds but the first does not, depth works as an uncertainty readout but does not improve reconstruction, a weaker but still publishable result. If neither holds, the hierarchy hypothesis for noise-limited cryo-EM is refuted.
References
- Reconstructing continuous distributions of 3D protein structure from cryo-EM images
- CryoDRGN-AI: neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM
- Poincaré Embeddings for Learning Hierarchical Representations
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- APo-VAE: Text Generation in Hyperbolic Space
- Boosted ab initio Cryo-EM 3D Reconstruction with ACE-EM
- CryoAI: amortized inference of poses for ab initio reconstruction of 3D molecular volumes from experimental cryo-EM images
- Learning to recover orientations from projections in single-particle cryo-EM