cryoSPHERE minimal reproducible dataset

Gabriel Ducrocq, Lukas Grunewald, Sebastian Westenhoff, Fredrik LindstenarXiv 2024 · DOI · arXiv

The field of structural biology is undergoing a quiet revolution. While tools like AlphaFold have mastered predicting static protein "snapshots," the real magic of biology happens in motion. Proteins are dynamic machines that twist, bend, and shift to perform their functions. Capturing this "conformational heterogeneity" is the next great challenge, and a new paper from ICLR 2025 introduces a powerful tool to do just that: cryoSPHERE.

TL;DR

cryoSPHERE is a deep learning framework that reconstructs the full range of a protein's movement from noisy cryo-EM images. By learning to divide a protein into rigid segments and moving them to fit 2D data, it achieves unprecedented resilience to noise and recovers realistic structural ensembles where traditional methods see only a blur.

What Problem Does This Paper Address?

Single-particle cryo-electron microscopy (cryo-EM) is the gold standard for seeing large protein complexes. Scientists freeze millions of protein copies in ice and take 2D "micrographs." However, these images are incredibly noisy—imagine trying to reconstruct a 3D movie of a dancer from thousands of grainy, static photos taken from random angles in a blizzard.

Traditional reconstruction methods usually produce a single "consensus" structure—essentially a long-exposure photograph that blurs out all the interesting movement. While newer "continuous heterogeneity" methods exist, they often struggle with the extreme noise of cryo-EM. They either produce unrealistic, "broken" proteins by trying to move every atom independently or get stuck in local minima, failing to capture large-scale domain movements.

Methods

The authors introduce cryoSPHERE, a Variational Autoencoder (VAE) designed specifically for structural reconstruction. Its workflow is elegant:

  1. The Input: It starts with a "nominal structure" (a starting guess, often from AlphaFold) and a set of cryo-EM images.
  2. Learned Segmentation: Instead of moving every atom, the model uses a Gaussian Mixture Model (GMM) to learn how to divide the protein chain into a user-defined number of segments. Crucially, this segmentation is learned during training based on how the protein actually moves in the data.
  3. Rigid-Body Deformations: For every single image, the encoder maps the 2D projection to a latent space. The decoder then translates this into a specific rotation and translation for each segment.
  4. Differentiable Physics: The model "poses" the 3D structure, projects it into a 2D image (accounting for microscope physics like the Contrast Transfer Function), and compares it to the original noisy image.

By constraining the motion to rigid segments, cryoSPHERE effectively "regularizes" the problem, preventing the model from fitting to noise and ensuring the resulting structures remain physically plausible.

Datasets

The authors validated cryoSPHERE on both simulated and experimental data:

  • Synthetic Toy Dataset: A simple two-domain system to prove the model can learn exact rotation angles.
  • Bacterial Phytochrome (MD Simulation): 10,000 structures sampled from a molecular dynamics trajectory, corrupted with extreme noise (SNR as low as 0.001).
  • EMPIAR-10180 (Spliceosome): A massive real-world dataset of a pre-catalytic spliceosome (over 327,000 images).
  • EMPIAR-12093 (Bacterial Phytochrome): A recently published real dataset featuring complex "scissoring" motions.

Key Insights and Ideas

  • Constraints are a Feature, Not a Bug: The core "aha!" moment is that by limiting the degrees of freedom (moving segments instead of atoms), the model becomes much more robust. This allows it to work at Signal-to-Noise Ratios (SNR) where other methods fail.
  • Automatic Domain Discovery: Although the "segments" are defined mathematically (via GMMs on residue indices), the model often naturally discovers actual biological domains because those are the parts that move together.
  • End-to-End Differentiability: Because the entire pipeline—from segmentation to 2D projection—is differentiable, the model can backpropagate errors directly from the 2D pixels to the 3D segment positions.

Key Results

  • Noise Resilience: In synthetic tests with an SNR of 0.001 (where the protein is barely visible to the human eye), cryoSPHERE recovered the correct distribution of domain distances with a correlation of 0.73, significantly outperforming state-of-the-art methods like cryoDRGN and cryoStar.
  • Real-World Success: On the EMPIAR-12093 dataset, cryoSPHERE successfully captured a complex "scissoring" motion in the bacterial phytochrome. Other methods either failed to reconstruct the moving parts or produced "clashing" residues where parts of the protein physically overlapped.
  • Resolution Gains: The method showed consistent improvements in Fourier Shell Correlation (FSC) at critical cutoffs, indicating more accurate and higher-resolution reconstructions of the moving parts.

Limitations and Future Work

While powerful, cryoSPHERE has a few hurdles to clear:

  • Compositional Heterogeneity: The current model assumes the protein chain is always complete. If parts of the protein are missing in some images (compositional change), the model tries to "hide" those residues rather than deleting them.
  • Dependency on a Base Structure: It requires a good starting PDB file. If the AlphaFold prediction is significantly wrong, the reconstruction will inherit those biases.
  • Future Directions: The authors suggest moving toward learning the amplitude of Gaussian modes to handle missing density better and integrating chemical force fields to ensure even higher biological accuracy.

Reference: Ducrocq, G., et al. (2025). CRYOSPHERE: Single-Particle Heterogeneous Reconstruction from Cryo EM. Published as a conference paper at ICLR 2025. arXiv:2407.01574

Summary updated

Abstract

About cryoSPHERE: This dataset is a minimal reproducible example for cryoSPHERE published at ICLR 2025, see github: https://github.com/Gabriel-Ducrocq/cryoSPHERE and the project page: https://gabriel-ducrocq.github.io/cryosphere.github.io/ and the paper: https://arxiv.org/pdf/2407.01574 How to reproduce the results: These results have been obtained running cryoSPHERE as explained on github, with the paramters described in parameters_two_segmentation.yaml, and taking the 1000th epoch to perform the analysis step. To reproduce the results, you can start with the particles.star, particles.mrcs, base_structure_minimal_reproducible.pdb, parameters_two_segmentation.yaml and images.yaml. From that, you can follow the steps 1 and 1 bis: obtain a consensus reconstruction with the tool of your choice, fit base_structure_minimal_reproducible.pdb in that reconstruction and center it, following the steps described on github. Alternatively, you can skip the steps 1 and 1 bis and run cryoSPHERE directly on the base_structure_minimal_reproducible_centered.pdb structure (step 2 on github). The configuration files you need to run cryoSPHERE are parameters_two_segmentation.yaml and images.yaml. They are in setup_files.zip. How has the dataset been generated: We sampled 10000 structures from a molecular dynamics simulation, see appendix B.2 of the paper. You can find the structures in ground_truth_structures.zip. We then sampled 2 poses per structure, uniformly distributed on the set of poses. We posed the structures and projected them according to our image formation model, see equation (1) and (2) in our paper. Finally, we added Gaussian noise to the picture to get a SNR of 0.1. This procedure results in 20000 images, aranged in the same order as the structures in ground_truth_structures.zip: the first two images correspond to the first structure, images 3 and 4 to structure 2, images 5 and 6 to structure 3 and so on... Plots: After running cryoSPHERE and up to 1000 epochs, we predicted 10000 structures (one structure every two pictures, so we actually get one predicted structure per one ground truth structure). We then computed the distance between domain 1 (residues 321-503 of chain A) and domain 2 (residues 321-503 of chain B) for each ground truth and predicted structure. We plot two things, available in the plots.zip: scatter_plot.png is a scatter plot of the predicted vs true distances (Å) where we see that the points are along the diagonal, showing that cryoSPHERE recovers the right conformation for a given image. histogram.png is two histograms, one of the true distances and the other one of the predicted distances. We see that they match. To reproduce these results, you can use the compute_distances.py script to compute all the distances of a set of images, within the cryoSPHERE environment you created when installing it: python compute_distances.py --path path/to/folder/structures/ --chain_information where --chain_information indicates that the chain numbers have been kept in the pdb files. If --no-chain_information, the opposite is assumed. And --path is set to the path of the folder containing the structures. This command will output a .npy file containing the distances in the same order as the numbering of the structures, in the structure folder. Once you have the distances, you can create the plots using: python plotting.py --true_distances /path/to/true_distances.npy --predicted_distances /path/to/predicted_distances.npy ---output_path /path/to/output/folder/ where --true_distances is the path to the npy file containing the true distances where --predicted_distances is the path to the npy file containing the predicted distances where --output_path is the path to the folder where the plot will be saved. The outputs of this command are one histogram and one scatter plot, like the ones in plots.zip How is this dataset organized: There are 5 files: particles.mrcs is the mrcs file containing the particles. setup_files.zip contains multiple files: particles.star is the star file containing the poses and ctf for each particle. base_structure_minimal_reproducible.pdb is the base structure obtained via AlphaFold. base_structure_minimal_reproducible_centered.pdb is the same as base_structure_minimal_reproducible.pdb but centered, following the steps explained on github. back_projection.mrc is the volume obtained using a consensus reconstruction method (here the backprojection algorithm available with cryoDRGN). back_projection_centered.mrc is the same volume as back_projection.mrc but centered following the setps explained on github. base_structure_minimal_reproducible_centered_mrc.mrc is the volume obtained from base_structure_minimal_reproducible_centered.pdb as explained on github. parameters_two_segmentation.yaml is the configuration file used to run cryoSPHERE. images.yaml is the configuration file with the informations about the images in particles.mrcs cryosphere_results.zip contains the results produce by cryosphere at training epoch 1000: pc0, folder containing: pca.png, picture of the first two axes of the PCA decomposition of the latent space, with their explained variance and 20 points corresponding to the traversal of the first axis. 20 structures, named structure_z_x.pdb where x goes from 0 to 19 corresponding to the traversal points of pca.png. The file structure_z_0.pdb corresponds to the point with the lowest value of the PCA axis 1 and the value of x corresponds to increasing values of the PCA axis 1. pc1, is the same as the folder pc0, but for PC 2. pc2, is the same as the folder pc0, but for PC 3. predicted_structures, folder containing the predicted structure for each image (20000 in total, in the same order as the images). z.npy the 20000 latent variable, one for each image, in the same order at the images in the particles.star file. ground_truth_structures.zip contains the 10000 structures used to generate the dataset, in the same order as the images. It also contains a file called true_distances.npy containing the distances of the ground truth structures, see above. plots.zip, containing the plots of the results: One histogram, see above. One scatter plot, see above. compute_distances.py to compute the distances of the structures output by cryoSPHERE, see above. plotting.py to recreate the histogram and scatter plot, see above.
BibTeX
@misc{ducrocq2024cryosphere,
  author = {Gabriel Ducrocq and Lukas Grunewald and Sebastian Westenhoff and Fredrik Lindsten},
  title = {{cryoSPHERE minimal reproducible dataset}},
  year = {2024},
  eprint = {2407.01574},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.01574},
}
Figure