CryoNeRF

A cryo-EM dataset is a stack of noisy 2D projections of many copies of the same complex, imaged in random orientations. Those copies are rarely identical: a domain may hinge (conformational heterogeneity) or a subunit may be present in some particles and absent in others (compositional heterogeneity). Recovering a 3D density — and, ideally, the full distribution of states behind it — is the central computational problem of the field.

CryoNeRF attacks it with neural radiance fields, the implicit-representation technique developed for novel-view synthesis in computer vision. The analogy is close: NeRF learns a continuous function from 3D coordinates to density and color given many 2D views of a scene, which is structurally the same problem as learning a continuous electron density from many 2D particle projections. The distinguishing choice is that CryoNeRF works entirely in Euclidean 3D space. Established neural methods such as CryoDRGN-AI and the autoencoder-based OPUS-DSD2 operate in Fourier space, where discretization introduces numerical error that costs high-frequency detail; CryoNeRF avoids the transform altogether and pays for it in GPU memory instead.

The method comes from Tianlong Chen's group at UNC Chapel Hill with collaborators at the University of Washington — including William Stafford Noble — and Purdue, and was posted as a preprint in January 2025.

#Key Features

• Reconstruction without a Fourier transform: Density is queried at arbitrary 3D coordinates and rendered into 2D projections directly, removing the interpolation error that Fourier-space methods incur on high-frequency terms.
• Multi-resolution hash encoding: Coordinates are encoded through a pyramid of learnable hash tables, giving fine spatial detail without the parameter cost of a dense voxel grid.
• Heterogeneity-aware encoder: A ResNet-34 image encoder maps each particle image to an embedding that conditions the density network, so one trained model covers a whole continuum or set of states.
• Interpretable latent space: The per-particle embeddings organize into structure — a continuum for rotational motion, separated clusters for distinct assembly states — that can be inspected and used to pick out states for reconstruction.
• Unsupervised optimization: Training minimizes reconstruction loss between input images and projections of the predicted volume, requiring no labels or state annotations.

#Technical Details

CryoNeRF takes 2D particle images with their CTF parameters and poses; the rotations and translations come from prior ab initio or homogeneous refinement, so the method sits as a heterogeneity-modeling stage rather than replacing pose search. On homogeneous data the paper reports gold-standard FSC resolutions of 3.2 Å on the Plasmodium falciparum 80S ribosome (EMPIAR-10028) against 3.8 Å for cryoDRGN and 4.1 Å for cryoSPARC, and 3.8 Å on the RAG1–RAG2 complex (EMPIAR-10049) against 4.0 Å and 4.5 Å. The abstract's headline figure of 15.8% improvement is the ribosome case specifically; on the RAG complex the margin over cryoDRGN is 4.0 Å to 3.8 Å.

On the simulated CryoBench benchmarks, reconstructions of the IgG-1D dataset (100,000 images spanning a rotational continuum) reached 2.63 ± 0.01 Å across all rotation angles, and Ribosembly reached 2.06 ± 0.08 Å across classes, close to the 2 Å numerical limit of the FSC metric at that box size. On experimental data, the pre-catalytic spliceosome (EMPIAR-10180, 307,240 images) yielded a latent space tracking the known conformational transition, and the L17-depleted 50S assembly intermediates (EMPIAR-10076, 131,899 images) separated all six published states, including one accounting for roughly 2% of particles. The cost is real: 60 epochs on IgG-1D took 18 hours on four A100 80GB GPUs.

#Applications

The target user is a structural biologist who already has a refined particle stack and wants to know what states are hiding in it. Typical uses are resolving a functional cycle from a single imaging session, checking whether a sample is genuinely homogeneous before pushing for higher resolution, and identifying minority assembly intermediates in ribosome or large-complex datasets that classification would otherwise average away. The released package runs on arbitrary new particle stacks and, per the authors, on Colab.

#Impact

CryoNeRF imports the neural-radiance-field toolkit — hash encodings, volume rendering, coordinate networks — into cryo-EM heterogeneity analysis, and shows that giving up the Fourier-space formulation buys resolution rather than costing it. The code is on GitHub under GPL v3 with contact details for alternative licensing, and preprocessed particle stacks, checkpoints, and reconstructions are archived on Zenodo. The work remains a preprint and has not been peer reviewed. Its main structural limitation is stated by the authors: reconstruction quality inherits the accuracy of the input poses, so integrating pose estimation to reconstruct straight from raw particle images is left to future work. Euclidean-space rendering is also markedly more memory-hungry than Fourier-based alternatives.