CryoGaussian: Heterogeneous Cryo-EM Reconstruction Made Explicit with Gaussian Splatting

Ziwen Wang, Jiajun Fan, Chaoran Cheng, Ruihan Guo, Yanru Qu, Juexiao Zhou, Ge LiuarXiv 2026 · arXiv

TL;DR

CryoGaussian introduces an explicit 3D Gaussian Splatting framework for reconstructing dynamic biomolecular structures from noisy cryo-EM images. By combining a differentiable, additive rasterizer with latent-conditioned Gaussian modulation, it achieves state-of-the-art accuracy in resolving rare structural states and continuous motions while being up to 28x faster than previous Gaussian-based methods.

What Problem Does This Paper Address?

Single-particle cryo-electron microscopy (cryo-EM) is a powerful tool for determining the near-atomic structures of proteins and viruses. However, molecules are not static; they exist in multiple "compositional" states (where parts are added or removed) and exhibit "conformational" heterogeneity (continuous flexibility).

Traditional reconstruction methods often assume a single static structure or struggle to capture rare, transient states. While newer neural-field-based approaches can model flexibility, they are often "implicit" (hard to interpret directly) and computationally expensive. Previous attempts to use "explicit" Gaussian mixtures often required a high-quality reference structure to start, limiting their ability to discover entirely new or rare structural variations from scratch.

Methods

The authors present CryoGaussian, which represents molecular density as a collection of 3D Gaussians. Unlike standard 3D Gaussian Splatting used in computer vision, CryoGaussian is tailored for the physics of cryo-EM:

  1. Additive Differentiable Rasterizer: Cryo-EM images are orthographic line integrals rather than perspective projections with occlusion. The authors developed a custom rasterizer that accumulates Gaussian contributions additively and accounts for the Contrast Transfer Function (CTF), which blurs and distorts the raw images.
  2. Data-Derived Constraints: To enable "ab initio" (from scratch) reconstruction without a reference map, the method uses the data itself to set constraints. The number of Gaussians is derived from the estimated spatial support of the molecule, and their minimum size is bounded by the Nyquist frequency of the images to prevent the model from simply "fitting the noise."
  3. Latent-Conditioned Modulation: To handle heterogeneity, each particle image is assigned a latent vector. A set of small neural networks (MLPs) takes these latents and modulates the position, amplitude, and scale of every Gaussian in the model. This allows a single set of Gaussians to represent an entire landscape of structural variations.
Overview of the CryoGaussian pipeline, showing data-derived constraints and latent-conditioned modulation.

Datasets

The method was evaluated on both synthetic and real-world experimental data:

  • CryoBench: A comprehensive synthetic benchmark featuring five datasets (IgG-1D, IgG-RL, Ribosembly, Tomotwin-100, and Spike-MD) designed to test various types of structural heterogeneity.
  • EMPIAR-10076: Experimental data of the assembling 50S ribosomal subunit, known for complex compositional changes.
  • EMPIAR-10180: Experimental data of the pre-catalytic spliceosome, which exhibits significant continuous conformational motion.

Key Insights and Ideas

  • Explicit is Efficient: By using an explicit Gaussian representation instead of an implicit neural field, the model is more interpretable and significantly faster to train and render.
  • Random Initialization Works: The paper demonstrates that with the right physical constraints (like the Nyquist bound), you don't need a starting reference structure. The Gaussians can "find" the molecule starting from random positions.
  • Linearity Enables Speed: Because cryo-EM projection is linear (additive), the authors could optimize the backward pass of their rasterizer to accumulate gradients per-Gaussian rather than per-pixel, leading to massive speedups on modern GPUs.

Key Results

  • State-of-the-Art Accuracy: CryoGaussian outperformed all compared methods (including CryoDRGN and 3DFlex) on the CryoBench benchmark for both compositional and conformational tasks.
  • Resolving Rare States: On experimental ribosome data, it successfully resolved a rare assembly intermediate (class C4) that made up only 1.3% of the total particles.
  • Visualizing Motion: It captured the complex, continuous "swinging" motions of subcomplexes within the spliceosome.
  • Massive Speedup: The optimized rasterizer reduced training time by 12.8x to 28.0x compared to the previous state-of-the-art Gaussian splatting baseline (CryoSplat).
Reconstructed densities and latent space projections for the CryoBench datasets.

Limitations and Future Work

While CryoGaussian is a major step forward, it currently relies on pre-calculated particle orientations (poses) and CTF parameters from other software. The authors suggest that future versions could jointly optimize these parameters alongside the structure.

Additionally, they envision integrating CryoGaussian into "agentic workflows"—automated systems that can iteratively select particles and refine structures to support rapid drug and vaccine design.

Compositional heterogeneity resolved in the 50S ribosomal subunit, highlighting the rare C4 state.

Summary updated

Abstract

Single-particle cryo-electron microscopy (cryo-EM) determines macromolecular structures from noisy particle images. However, rare, transient compositional states and subtle, continuous conformational variation are difficult to resolve. We present CryoGaussian, a Gaussian splatting method for heterogeneous cryo-EM reconstruction that fits an explicit Gaussian density representation directly to the particle images through a differentiable rasterizer rendering orthographic line integrals modulated by the contrast transfer function. The Gaussians are initialized randomly, with their number derived from the estimated spatial support of the molecular density and their minimum size bounded by the Nyquist frequency of the particle images. To capture the compositional and conformational variation across particle images, we learn particle-image and per-Gaussian latents that jointly modulate the base position, amplitude, and scale of each Gaussian. Each structural state is then reconstructed by fitting the Gaussians conditioned on its inferred latent. On CryoBench, CryoGaussian achieves state-of-the-art reconstruction accuracy for both compositional and conformational heterogeneity. On experimental data, it resolves a rare compositional state of the 50S ribosomal subunit (class C4, 1.3% of particles) with rRNA helix 68 docked in the absence of the central protuberance, and continuous conformational motion of the SF3b and helicase subcomplexes of the pre-catalytic spliceosome. We further optimize the rasterizer for additive cryo-EM projection, reducing training time by 12.8-28.0$\times$.
BibTeX
@misc{wang2026cryogaussian,
  author = {Ziwen Wang and Jiajun Fan and Chaoran Cheng and Ruihan Guo and Yanru Qu and Juexiao Zhou and Ge Liu},
  title = {{CryoGaussian: Heterogeneous Cryo-EM Reconstruction Made Explicit with Gaussian Splatting}},
  year = {2026},
  eprint = {2611.26128},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26128},
}
Figure