Overcoming the Preferred-Orientation Problem in Single-Particle Cryo-EM: An Innovative Solution through Deep Learning

Background Introduction

In recent years, single-particle cryogenic electron microscopy (Single-Particle Cryo-EM) has become a core technique in structural biology due to its ability to resolve the atomic-resolution structures of biomacromolecules in near-native states. However, a major technical bottleneck—the “preferred orientation” problem—has continued to challenge researchers. This problem stems from the uneven orientation distribution of biomolecules on cryo-EM grids, leading to insufficient data sampling in certain directions. The orientation bias often arises during sample preparation due to the interaction of molecules with the air–water interface (AWI) or the support film–water interface.

The preferred-orientation problem is particularly severe in three-dimensional (3D) reconstructions, causing anisotropy that damages or even distorts structures. These issues manifest as skewed secondary structures, peptide chain breaks, and discontinuous densities, ultimately compromising molecular model accuracy. This problem is especially detrimental for low-symmetry or asymmetric molecules. Previous studies have attempted to resolve the preferred-orientation issue through biochemical and physical approaches, such as adjusting sample preparation methods, using different support films, and adding reagents to alter molecular behavior. However, these methods are often complex, time-consuming, expensive, and prone to introducing new artifacts, such as increased background noise or new preferred-orientation issues.

To overcome this challenge, a research team from the University of California, Los Angeles (UCLA), led by Yun-Tao Liu, Hongcheng Fan, Jason J. Hu, and Z. Hong Zhou, proposed a novel computational method. They developed a tool called SPISONet (Single-Particle IsoNet) based on “self-supervised deep learning” to address the anisotropy issues arising from preferred orientation. Remarkably, SPISONet achieves this without requiring changes to sample preparation workflows. This study was published in Nature Methods in January 2025.

---

Research Process

Research Goal

The development of SPISONet primarily targets two core issues caused by preferred orientation: 1. Anisotropy in reconstructions 2. Particle misalignment due to anisotropy

The research team introduced two interdependent but complementary modules: the Anisotropy Correction Module and the Misalignment Correction Module, both enabled by an innovative self-supervised deep learning framework.

---

Workflow and Experimental Design

1. Anisotropy Correction Module

This module operates on two unfiltered half-maps and a solvent mask as inputs. The key steps include:

• 3DFSC Calculation
  A 3D Fourier shell correlation (3DFSC) algorithm quantifies anisotropic resolution in reconstructions. An efficient 3DFSC implementation in SPISONet reduces computational complexity and execution time significantly.

• Deep Learning Network Training
  The module leverages a U-Net-based neural network architecture, trained with four loss functions: Consistency Loss, Equivariance Loss, and two Noise2Noise-based losses. These loss functions ensure the network fills in under-sampled regions while avoiding overfitting and artifact generation.

• Information Recovery and Denoising
  Through self-supervised learning, SPISONet compensates for density loss in anisotropic regions and denoises the reconstruction, enhancing the quality of the 3D maps.

Tests on both simulated and real datasets demonstrate that SPISONet’s anisotropy correction module significantly improves map quality for datasets with mild to moderate preferred orientation. It can also partially alleviate extreme structural damage caused by severe anisotropy.

---

2. Misalignment Correction Module

In 3D molecular reconstruction, preferred orientation leads to incorrect particle-orientation assignments, which are a major source of reconstruction artifacts. The misalignment correction module addresses this through a comprehensive workflow:

• Reference Map Generation
  Users provide reference maps (generated from tilted data or low-resolution initial structures).

• Automated Iterative Refinement
  This module integrates RELION’s 3D refinement with SPISONet’s anisotropy correction, iteratively optimizing particle orientation distributions. To avoid potential model bias, users can adopt a low-resolution mode to maintain control over the results.

• Unified Error Control
  The iterative refinement process significantly reduces artifacts arising from orientation bias.

Compared to standard 3D reconstruction workflows, SPISONet demonstrates superior error correction performance. After applying the combined misalignment and anisotropy corrections, multiple cases exhibit improved secondary structure density, such as clear α-helical pitch and amino acid side-chain detail.

---

Datasets and Experimental Validation

1. β-Galactosidase Dataset

The research team tested SPISONet on β-Galactosidase Cryo-EM data (a RELION tutorial dataset). By isolating particles with specific orientations (from 2D classifications), the team confirmed that SPISONet successfully enhanced anisotropic maps, improving secondary structure visibility in both side-view and top-view dominant reconstructions.

2. Hemagglutinin Trimer Tilted Dataset

Using a dataset collected at a 40° tilt (EMPIAR-10097), SPISONet achieved a resolution improvement to 4.1 Å, sharpening previously invisible side-chain regions. Additionally, the misalignment correction module improved isotropy, overcoming directional sampling bias.

3. Hemagglutinin Trimer Non-Tilted Dataset

In a dataset with severe preferred orientation (EMPIAR-10096), SPISONet combined misalignment and anisotropy corrections to generate a 3.5 Å reconstruction. Notably, this method resolved previously insurmountable artifacts, producing interpretable secondary structure features.

4. Pathogenic 70S Ribosome Dataset

The team applied SPISONet to an Acinetobacter baumannii 70S ribosome dataset with a curated subset of preferentially oriented particles. Using low-resolution references (70S or 80S ribosomes), SPISONet consistently generated high-quality reconstructions, demonstrating its flexibility and robustness.

5. HIV Virus-Like Particle (VLP) Dataset

In HIV VLP subtomogram averaging experiments, SPISONet improved a 3.6 Å reconstruction. All metrics, including side-chain density and isotropy, exceeded those of standard pipelines, showcasing its potential for in situ structural studies.

---

Results and Conclusion

SPISONet provides an efficient and general computational solution to the preferred-orientation problem through its self-supervised deep learning framework. Its key advantages include:

1. Fully Computational Workflow: SPISONet eliminates the need for complex sample preparation adjustments.
2. Exceptional Performance on Extreme Datasets: It successfully reconstructs high-quality structures even in datasets with severe preferred orientation.
3. Flexibility and Generality: The tool supports diverse biological molecules and integrates seamlessly with both single-particle Cryo-EM and subtomogram averaging methods.

This innovative computational tool not only expands Cryo-EM’s applicability to challenging biological samples but also paves the way for efficient, high-throughput, and high-precision structural determination.