Hydra

Hydra is a method for ab initio heterogeneous reconstruction in cryo-electron microscopy, developed by Ellen Zhong's group at Princeton University with Gordon Wetzstein's group at Stanford and collaborators at Columbia. It was presented at NeurIPS 2024.

Single-particle cryo-EM images an ensemble of molecules, and two kinds of variation coexist in a dataset: conformational heterogeneity, the continuous motion of one complex, and compositional heterogeneity, the presence of several different complexes. Neural reconstruction methods had addressed the first while assuming a single species, so a stack containing more than one protein — an unpurified lysate, an in situ sample — could not be reconstructed without first separating the particles. That separation is itself the hard part: cryo-EM images are too noisy to cluster directly, and orientations are unknown at initialization, so the two problems cannot be solved in sequence.

Hydra removes the sequencing by representing the accessible structures as a mixture of K neural fields rather than one. Each field is specialized to a single compositional state, and the poses, per-particle conformations, class probabilities and the fields themselves are all optimized together against a single likelihood. The result extends the neural-field line begun in CryoDRGN-AI, which Hydra builds on directly and whose user interface and workflow it inherits.

#Key Features

• Joint compositional and conformational inference: Structure, pose, class assignment and conformational state are estimated simultaneously, so mixtures of flexible complexes can be reconstructed without a prior classification step.
• Mixture of neural fields: The density map for class k comes from its own coordinate network, implemented as a residual MLP conditioned on a low-dimensional conformational latent, giving each species its own representational capacity.
• Likelihood-based mixture loss: The negative log-likelihood of a Gaussian mixture replaces a hard assignment, so class identity is a soft variational quantity during optimization and only collapses to an index at the end.
• Fully ab initio: No consensus refinement, initial model, or upstream pose estimate is required, which matters precisely on the compositionally mixed datasets where consensus poses are unreliable.
• Tolerates over-parameterization: Setting K above the true number of species leaves the surplus classes empty rather than fragmenting a real one.

#Technical Details

Poses are optimized with the two-stage strategy inherited from CryoDRGN-AI: hierarchical pose search over a discrete grid of orientations and translations for the first 100,000 to 2,000,000 steps, then stochastic gradient descent, with latents, class scores and network parameters trained by SGD throughout. On tomotwin3, a synthetic 3,000-image mixture of three PDB entries rendered at SNR 0.01, Hydra reaches an adjusted Rand index of 1.00 for classification with two conformational dimensions, against 0.59 for DRGN-AI, 0.36 for cryoDRGN2 and 1.00 for cryoSPARC, and the best per-image Fourier shell correlation of the four. On ribosplike — 15,000 images combining the pre-catalytic spliceosome, the 80S ribosome and the SARS-CoV-2 spike, each moving along a one-dimensional trajectory — it records an ARI of 0.997, a per-image FSC of 0.414, and a median pose error of 1.07 degrees. On an experimental red blood cell lysate of 85,656 particles, Hydra with K = 4 recovers the ryanodine receptor, respiratory chain complex III, the p97 valosin-containing protein and a fourth junk class in a single ab initio pass; DRGN-AI splits the same data only into ryanodine receptor and non-ryanodine receptor particles.

#Applications

Hydra is aimed at samples that resist purification: cellular lysates, cross-linked fractions, and in situ datasets where several complexes are imaged together. For a structural biologist, the practical gain is that a mixed stack can go straight into reconstruction instead of through an expert-driven cryoSPARC classification pipeline first. The code ships under GPL-3.0 as a Python package installed from source, and follows the DRGN-AI configuration and analysis workflow, so groups already running that software can adopt it without a new toolchain.

#Impact

Hydra is the first neural reconstruction method to handle discrete compositional mixtures and continuous motion in the same model, and it demonstrates that a neural approach can match the classification quality of voxel-based mixture methods while retaining a continuous conformational representation. Like the rest of its family, it is an optimization procedure rather than a distributed checkpoint: every dataset gets its own fitted fields, and the reported runs use hierarchical pose search over hundreds of thousands of steps, which is computationally demanding. The number of classes K remains a user-set hyperparameter, and while over-parameterization is tolerated, the method offers no automatic estimate of how many species a sample contains.