Cryogenic electron microscopy (cryoEM) has revolutionized structural biology, enabling the atomic-level visualization of macromolecular complexes in near-native states. However, a persistent challenge—the "preferred orientation" problem—has continued to hinder structural determination using this powerful technique. In our latest study, we introduce spIsoNet, a self-supervised deep-learning method that offers a computational solution to this challenge, paving the way for high-throughput cryoEM structure determination.

The Challenge of the Preferred Orientation Problem

In a cryoEM experiment, one aims to capture many 2D images of individual molecules in random orientations, uniformly distributed in all possible directions. This randomness is crucial because it samples all the different views of the molecule, necessary to reconstruct a complete and accurate 3D model of the molecules.

In practice, however, truly random distribution of particle orientations is rarely the case. During sample preparation, a small amount of solution is spread onto a cryoEM grid and quickly frozen in liquid ethane, forming a thin, vitreous (glass-like) layer of ice. While this ice preserves molecules in their near-native states, it also introduces two air-water interfaces (AWIs)—where the liquid meets the air. These interfaces can interact with the surface of the molecules, which often have regions that are more electrostatically charged or hydrophobic (water-repelling). As a result, certain parts of the molecule tend to stick to the interfaces, causing the molecules to settle in preferred orientations.

The preferred orientation problem is a significant issue in cryoEM because it leads to biased data: one ends up collecting many images showing the same view of the molecule, while other orientations are underrepresented or missing entirely. This bias can prevent investigators from accurately reconstructing the 3D structure, leaving artifacts in our structures.

It’s like trying to assemble a 3D puzzle with pieces that all show the same side of the object—you can’t see the full picture.

Traditional Approaches and Their Limitations

Several traditional approaches have been explored to mitigate this issue, including modifying molecules to reduce their interactions with the AWI, occupying the AWI with detergents (or surfactants), or adding a thin support film to keep the sample away from the AWI. Other strategies include speeding up the freezing process to limit the time molecules spend near the AWI, or tilting the sample stage to capture images from different angles. However, these methods are often labor-intensive, ineffective, or sample-specific. Despite these efforts, the preferred orientation problem remains a significant hurdle in cryoEM, highlighting the need for a more general and straightforward solution to this challenge.

Deep Learning Approaches for Scientific Discovery

In recent years, deep learning tools have reshaped many traditional fields and opened new frontiers. Deep learning methods have mastered complex board games like Go, excelled in medical image segmentation, and generated highly realistic images and content. Notably, deep learning has also made significant strides in predicting molecular structures, exemplified by this year's Nobel-winning AlphaFold and RosettaFold. These breakthroughs have opened new doors for advancements in structural biology and drug discovery and highlight the transformative potential of deep learning in science and technology.

When it comes to scientific data, the key challenge of deep learning is ensuring that the content generated is both reliable and capable of contributing to genuine discovery. While deep learning models are trained on vast existing datasets, an important consideration is whether predictions based on this learned data can lead to novel insights. If discoveries are made from previously learned patterns, how can we ensure that the predictions are truly innovative? This presents a fundamental dilemma: can deep learning be used not only to inform but also to drive scientific discovery?

Tackling the Preferred Orientation Problem with Self-Supervised Deep Learning

The preferred orientation problem in cryoEM can be framed as the challenge of predicting under-sampled molecule views, ensuring isotropic reconstructions and avoiding artifacts. This turns the problem into a machine learning challenge: how can we reliably infer the missing orientations in cryoEM data?

spIsoNet (single-particle IsoNet) is a self-supervised deep learning-based software designed to restore angular isotropy in cryoEM structures suffering from preferred orientations. The core idea behind spIsoNet is to exploit the rich, recurring information present in well-sampled orientations of the macromolecule to reconstruct under-sampled views. The self-supervised nature of spIsoNet means it does not rely on external training data or predefined structural models. Instead, it learns directly from the cryoEM maps provided by the user, ensuring that the resulting reconstructions are accurate and faithful representations of the molecular structures.

spIsoNet comprises two primary modules. Anisotropy Correction: this module addresses the anisotropic nature of reconstructions by correcting uneven sampling in Fourier space, which is a consequence of the biased orientation distribution. Misalignment Correction: this module improves particle alignment accuracy by fixing errors introduced during iterative refinement processes in cryoEM workflows. These errors are often amplified by the structure distortions caused by preferred orientation.

By integrating these two modules, spIsoNet enhances both the angular isotropy of the reconstructed maps and the accuracy of particle alignments. This computational solution presents a novel approach to addressing the preferred orientation issue, eliminating the need for additional experiments or biases from existing datasets.

Looking ahead, future enhancements may include integrating spIsoNet with existing cryoEM pipelines, further optimizing its algorithms for even broader applicability, and extending its capabilities to tackle other computational challenges in structural biology. We hope that spIsoNet should serve as an inspiration for further innovations at the intersection of deep learning and structural biology.

The software and documentation are available at https://github.com/IsoNet-cryoET/spIsoNet.