Abstract
Cryo-EM reconstruction algorithms seek to determine a molecule's 3D density map from a series of noisy, unlabeled 2D projection images captured with an electron microscope. Although reconstruction algorithms typically model the 3D volume as a generic function parameterized as a voxel array or neural network, the underlying atomic structure of the protein of interest places well-defined physical constraints on the reconstructed structure. In this work, we exploit prior information provided by an atomic model to reconstruct distributions of 3D structures from a cryo-EM dataset. We propose Cryofold, a generative model for a continuous distribution of 3D volumes based on a coarse-grained model of the protein's atomic structure, with radial basis functions used to model atom locations and their physics-based constraints. Although the reconstruction objective is highly non-convex when formulated in terms of atomic coordinates (similar to the protein folding problem), we show that gradient descent-based methods can reconstruct a continuous distribution of atomic structures when initialized from a structure within the underlying distribution. This approach is a promising direction for integrating biophysical simulation, learned neural models, and experimental data for 3D protein structure determination.
Introduction
Single-particle cryo-electron microscopy (cryo-EM) has become a cornerstone of structural biology, enabling the visualization of proteins at near-atomic resolution. However, a significant "representation gap" exists in current computational pipelines. Standard reconstruction algorithms typically represent proteins as generic 3D volumes—either through voxel grids or implicit neural representations like Multi-Layer Perceptrons (MLPs). These models lack inherent knowledge of atomic physics, such as bond lengths and connectivity.
This lack of physical constraints often leads to manual, labor-intensive post-processing to fit atomic models into the resulting density maps. Furthermore, when dealing with flexible (heterogeneous) proteins, unconstrained models can produce artifacts like "ghost" densities where mass appears or disappears inconsistently across different states. Cryofold addresses this by incorporating a coarse-grained atomic model directly into the reconstruction loop, effectively turning the reconstruction process into a physically constrained "folding" problem.
Methods: The Cryofold Architecture
Cryofold parameterizes the 3D density map using a coarse-grained atomic model instead of voxels. The core of the approach relies on two main components:
- RBF Volume Model: Each amino acid is represented by two Gaussian Radial Basis Functions (RBFs), corresponding to the backbone and sidechain. The total 3D density is calculated as the sum of these RBFs.
- Physics-Inspired Constraints: To maintain structural integrity, harmonic bond terms are applied between backbone RBFs, alongside distance constraints for sidechains.
For heterogeneous datasets, Cryofold employs a Variational Autoencoder (VAE) framework. An image encoder maps 2D projections to a latent space ($z$), while a decoder predicts $z$-dependent offsets for the atomic coordinates. The entire system is trained end-to-end using Stochastic Gradient Descent (SGD) to minimize the difference between projected atomic models and observed 2D images, incorporating physical constraint penalties as regularization.
Results and Discussion
The authors evaluated Cryofold using synthetic datasets based on a 141-residue protein (PDB ID: 5NI1). These datasets included a homogeneous set of 50,000 noisy images and a heterogeneous set simulating a 1D continuous backbone motion.
The Power of Physical Priors
By enforcing an atomic representation, Cryofold naturally regularizes the reconstruction. This leads to sharper 3D volumes compared to unconstrained neural models like cryoDRGN, which often produce blurred "clouds" of motion. Cryofold's ability to reconstruct specific atomic coordinates moving through a manifold allows it to accurately interpolate transition states, even when data is sparse. Ablation studies confirmed that internal bond terms are essential; their removal led to a dramatic degradation in accuracy, proving that backbone connectivity is the most critical prior.
Comparison with cryoDRGN
In heterogeneous reconstruction tasks, Cryofold demonstrated significant advantages. While both models can capture the general range of motion in latent space, Cryofold's physics-constrained RBF representation maintains a consistent, high-resolution structure across the entire conformational landscape. Unlike unconstrained models, Cryofold ensures that mass is conserved and physically connected throughout the reconstructed motion.
Challenges in Initialization
A key finding is that reconstructing from atomic coordinates is a highly non-convex optimization problem, akin to the protein folding problem. Cryofold is sensitive to initialization; starting from random coordinates often leads to failure. However, when initialized with a related reference structure—a common scenario in practical structural biology—the model converges reliably. For instance, initialization with an approximate structure from molecular dynamics yielded a backbone RMSE of 3.81 Å, while an exact structure baseline reached 0.843 Å.
Conclusion and Future Directions
Cryofold represents a promising step toward integrating biophysical knowledge into cryo-EM reconstruction. While currently tested on synthetic data with known poses, future work will focus on handling the high noise and unknown orientations of real-world experimental data. Additionally, addressing compositional heterogeneity—where parts of a complex may be missing in certain images—and scaling the model to larger macromolecular complexes remain important areas for development. More advanced optimization techniques, such as Hamiltonian Monte Carlo, may also be explored to mitigate the sensitivity to initialization.
Summary updated