💻 bioinformatics

Ab initio CryoEM maps minimizing preferred orientation problems in heterogeneous mixtures: ReconSIREN

ReconSIREN is a novel deep learning framework utilizing a Mixture-of-Experts architecture to perform fully \textit{ab initio} CryoEM single-particle analysis that simultaneously resolves preferred orientation artifacts, particle heterogeneity, and conformational flexibility without relying on traditional optimization methods.

Original authors: Herreros, D., Sorzano, C. O. S., Carazo, J. M.

Published 2026-10-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Herreros, D., Sorzano, C. O. S., Carazo, J. M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

To see the intricate machinery of life, scientists often turn to a technique called cryo-electron microscopy. Imagine trying to photograph a tiny, fragile machine that is constantly shifting and changing shape. To capture it, researchers flash-freeze the machine in a thin layer of ice, preserving it in a state that is as close to life as possible. They then shoot electrons through the ice to create thousands of two-dimensional shadows of the machine from every possible angle. The challenge lies in the fact that these shadows are incredibly faint and noisy, like trying to recognize a face in a heavy snowstorm. Furthermore, the machines often have a habit of lying down in the ice in just one or two specific ways, making it difficult to get a complete view from all sides. To build a clear, three-dimensional picture, computers must figure out exactly how each shadow was oriented and then piece them together. If the computer guesses the angles wrong, the final image becomes a blurry, distorted mess, hiding the very details scientists need to understand how these biological machines work.

A team of researchers at the Centro Nacional de Biotecnología in Madrid has developed a new tool called ReconSIREN to solve these specific problems. Their work addresses the two biggest hurdles in this process: the tendency of particles to stick to the ice in preferred orientations, and the difficulty of figuring out the correct angles without getting lost in the noise. In the past, when computers tried to align these images, they often fell into a trap known as the "attraction problem." Because some views of the particle are clearer than others, the computer would mistakenly drag all the images toward those clearer views, ignoring the ones that were actually different. This resulted in a 3D model that looked like a flattened pancake, missing crucial parts of the structure. ReconSIREN avoids this by using a deep learning system that does not rely on a single guess. Instead, it employs a team of independent experts, each starting with a different idea of where the image might be pointing. By comparing these different possibilities, the system can find the true orientation without being tricked by the noise or the lack of certain views.

The researchers tested this new system on a variety of difficult biological samples, including a protein involved in cell metabolism, a complex virus spike, and a membrane protein that is notoriously hard to study. In every case, the system was able to build a high-quality 3D map from scratch, without needing any prior information about how the particles were arranged. For the samples that suffered from strong preferred orientations, the new method succeeded where older techniques failed. While traditional software produced models that were missing entire sections of the protein, ReconSIREN recovered the full shape, including the parts that were previously invisible. The system also proved capable of detecting when the biological machines were flexing or changing shape, separating the different states of the molecule to show a full landscape of its movements. This is a significant step forward because it allows scientists to see the dynamic nature of these proteins right from the start of their analysis, rather than having to guess or refine the data later.

One of the most striking aspects of this work is that the system works entirely on its own, learning from the raw images without needing human intervention to correct the angles or center the particles. The researchers demonstrated that even when they fed the system images that had been simplified to speed up the process, the final results were accurate enough to be used as a starting point for further, more detailed analysis. In some tests, the system produced maps that were so precise they could be refined further to reveal structural details that rivaled those obtained after hours of traditional processing. The team also showed that the method works even when the input data is not the raw images, but rather pre-averaged groups of similar images, offering a faster route to a solution when time is critical. By distributing the workload across multiple independent experts within the software, the system avoids the common pitfall of getting stuck in a local minimum, a state where the computer thinks it has found the answer but has actually settled for a poor approximation.

The implications of this work extend beyond just making better pictures. By solving the problem of preferred orientations and attraction artifacts, ReconSIREN opens the door to studying biological specimens that were previously considered too difficult to analyze. Many important proteins, such as those embedded in cell membranes, naturally lie flat on the ice, making it nearly impossible to get a complete 3D view with older methods. With this new approach, scientists can now recover the full structure of these elusive molecules, including the parts that were previously hidden. The system also provides a clear view of how these molecules move and change, which is essential for understanding how they function and how they might be targeted by drugs. The researchers have made the software freely available, allowing other scientists to apply this powerful tool to their own challenging datasets. This democratization of the technology means that the barrier to entry for high-resolution structural biology is lowering, potentially accelerating discoveries across the field.

In the end, ReconSIREN represents a shift in how we approach the reconstruction of the microscopic world. It moves away from the rigid, step-by-step optimization of the past toward a more flexible, intelligent system that can navigate the complexities of noisy data and uneven distributions. The researchers did not just improve the resolution of the maps; they fundamentally changed the way the angles are determined, ensuring that the final image reflects the true shape of the molecule rather than the biases of the data collection. As the field of structural biology continues to push toward understanding more complex and dynamic systems, tools like this will be essential for turning the faint, scattered shadows of the electron microscope into clear, actionable insights into the machinery of life.

Technical Summary: ReconSIREN

Problem Statement
Cryo-Electron Microscopy Single Particle Analysis (CryoEM SPA) relies heavily on accurate ab initio estimation of relative orientations, in-plane shifts, and initial volume reconstructions. These parameters are critical for subsequent refinement and heterogeneity analysis. However, current workflows face significant challenges:

Preferred Orientation and the "Attraction Problem": Many specimens exhibit non-uniform projection distributions (preferred orientations). This imbalance often leads to an "attraction problem" where images are incorrectly assigned to projection directions with higher Signal-to-Noise Ratios (SNR), even if those directions are geometrically incorrect. This results in artefactual maps and local minima that obscure high-resolution features.
Convergence and Robustness: Existing deep learning approaches for ab initio reconstruction (e.g., CryoAI, CryoSPIN) often require a priori information, such as pre-centered images, preventing fully ab initio application. Others rely on hybrid approaches combining classical grid searches with local optimization (e.g., CryoDRGN-AI), which prevents full amortization of parameter estimation.
Heterogeneity: Accurately capturing conformational landscapes and compositional heterogeneity ab initio without trapping the model in artefactual states remains difficult.

Methodology: ReconSIREN
ReconSIREN is a novel deep learning framework designed to perform fully ab initio reconstruction, addressing alignment, shifts, consensus volume generation, and conformational heterogeneity simultaneously. Its architecture and optimization strategy include:

Mixture-of-Experts (MoE) Architecture: The framework distributes the estimation workload across independent expert modules, each with its own optimizer and cost function. This includes:
- Image Alignment Module: Uses a VGG16-based encoder to predict poses and shifts. To overcome local minima, it employs an ensemble of heads (default 18) initialized with "anchor" rotations evenly distributed across the projection sphere. This forces the network to explore different regions of the solution space simultaneously, reducing the risk of collapsing into a single, potentially incorrect, local minimum.
- Consensus Volume Module: Reconstructs the volume using Gaussian Splatting. It learns the positions and amplitudes of a set of Gaussians (initialized randomly within a sphere) to represent the 3D density.
- Conformational Variability Module: An autoencoder that maps images to a latent space and decodes them into heterogeneous 3D volumes using a SIREN-based MLP decoder with Feature-wise Linear Modulation (FiLM). This allows the model to learn continuous conformational changes and compositional heterogeneity.
Optimization and Priors:
- Pose-Ensemble Optimization: The network utilizes a "Winner-Takes-All" (WTA) strategy over the ensemble of predicted poses to select the best alignment for each image.
- Geometric Priors: To regularize the Gaussian reconstruction, the model applies losses for nearest-neighbor distances (to ensure signal continuity and prevent collapse), amplitude similarity between neighbors, and a spherical boundary to prevent outliers.
- Uniformity Prior: A Sliced Wasserstein Distance loss is applied to the predicted viewing directions to penalize deviations from a uniform distribution on the sphere ( $S^2$ ). This specifically mitigates the attraction problem by discouraging the network from collapsing all images into high-SNR directions.
- Gaussian Recycling: A mechanism to replace "dead" Gaussians (those with near-zero amplitude) with new ones near active Gaussians, ensuring the model retains expressivity throughout training.
- Heterogeneity Stabilization: To prevent the heterogeneity module from collapsing into the consensus map, the model enforces variance and decorrelation priors on the latent space dimensions.
Symmetry Handling: ReconSIREN can handle symmetry by internally symmetrizing predicted poses during training to enforce a symmetric consensus volume, while keeping the conformational variability module in $C_1$ (asymmetric) mode to capture true structural variations.

Key Results
The authors evaluated ReconSIREN on five experimental datasets (PKM2, SARS-CoV-2 spike, GR:Hsp90:FKBP51, HER2, ABCB11) and two challenging datasets with strong preferred orientations (a simulated CAHRA challenge dataset and Hemagglutinin from EMPIAR 10096).

Mitigation of Attraction Artifacts: In datasets with strong preferred orientations (e.g., Hemagglutinin), traditional methods like CryoSPARC failed to recover the full structure, often missing the bottom part of the protein due to attraction artifacts. ReconSIREN successfully recovered these missing regions and produced accurate ab initio maps without assuming symmetry.
Convergence and Accuracy: ReconSIREN demonstrated robust convergence to correct angular assignments and high-resolution consensus volumes. When used as an initial guess for local refinement in CryoSPARC, the resulting maps recovered structural details comparable to or better than those obtained from tilted datasets, despite using only untilted particles.
Heterogeneity Detection: The model successfully identified both compositional heterogeneity (in the channel dataset) and continuous motions (SARS-CoV-2 RBD, ABCB11, GR:Hsp90:FKBP51). Control experiments on the rigid PKM2 dataset showed the network did not hallucinate motions, correctly identifying the protein's rigidity.
Resolution and Input Flexibility: ReconSIREN can operate on downsampled images for speed or full-resolution images for maximum accuracy. When run on full-resolution data, it achieved ab initio reconstructions comparable to or surpassing results typically obtained after 3D refinement steps. It also demonstrated the ability to generate initial volumes from 2D class averages, offering a faster alternative when speed is prioritized.

Significance and Claims
The paper claims that ReconSIREN represents the first neural network capable of tackling the full ab initio estimation of projection directions, in-plane shifts, and heterogeneous states relying entirely on deep learning principles, without requiring a priori image centering or hybrid grid-search steps.

Its primary significance lies in:

Solving the Attraction Problem: By leveraging ensemble-based pose optimization and uniformity priors, it effectively minimizes the "attraction problem" caused by preferred orientations, a major bottleneck in CryoEM.
Robustness: It provides robust ab initio estimates across diverse specimens with varying flexibility and experimental conditions.
Workflow Integration: The method generates outputs compatible with standard software (CryoSPARC, STAR files), allowing it to serve as a powerful starting point for downstream refinement pipelines.
Efficiency: The MoE architecture improves optimization efficiency and simplifies the process of handling complex, heterogeneous datasets.

The authors emphasize that while ReconSIREN can produce high-quality final maps, it is designed to complement existing workflows, potentially reducing the need for manual intervention or repicking strategies to correct for preferred orientations.

More like this