Cryo-EM image processing of amyloid filaments in RELION-5.1

This paper introduces a suite of new tools in RELION-5.1 for the automated processing and structural determination of amyloid filaments from cryo-EM data, including automated picking, on-the-fly pre-processing, filament type selection via clustering, and a specialized denoising network, which are demonstrated to successfully identify new filament types in human islet amyloid protein datasets.

The Gist

Imagine you are trying to solve a massive, 3D jigsaw puzzle, but the pieces are microscopic, invisible to the naked eye, and they are all mixed together in a blurry, noisy photograph. This is the challenge scientists face when trying to understand amyloid filaments—long, stringy protein structures that are often linked to diseases like Alzheimer's and Type 2 diabetes.

This paper introduces a new, super-smart toolkit inside a software program called RELION-5.1 that acts like a "digital detective" to solve these puzzles faster and more accurately than ever before.

Here is how the new tools work, explained with everyday analogies:

1. The "Rhythm Detector" (The New Auto-Picker)

The Problem: In a photo of a cell, amyloid filaments look like long, thin threads. But there are also other threads (like actin) that look similar but aren't the ones you want. Also, the photos are very grainy (noisy), making it hard to see the threads.
The Solution: The authors realized that amyloid filaments have a very specific "beat" or rhythm. Just like a drumbeat repeats every few seconds, the structure of an amyloid filament repeats every 4.75 Angstroms (a tiny unit of measurement).
The Analogy: Imagine you are in a crowded room trying to find people wearing red hats. Instead of looking at every face, you put on special glasses that only light up when they see a specific pattern of red. The new software does this: it scans the blurry photos looking only for that specific 4.75 Å "beat." If it finds the rhythm, it knows, "Aha! That's an amyloid filament!" and marks it. If it doesn't hear the beat, it ignores it. This stops the computer from getting confused by other types of threads.

2. The "Sorter" (Bi-Hierarchical Clustering)

The Problem: Sometimes, a single sample contains a mix of different types of amyloid filaments (like having a bag of mixed Lego sets). If you try to build the model using all the pieces at once, you get a messy, broken structure. You need to separate the "Star Wars" sets from the "Harry Potter" sets before building.
The Solution: The new tool looks at the 2D "shadows" (class averages) of the filaments and groups them based on who they hang out with.
The Analogy: Imagine a giant dance floor where everyone is dancing. Some people are dancing the Waltz, others are doing the Salsa, and some are breakdancing. If you just look at the crowd, it's a mess. But if you look at who is dancing next to whom over time, you can see clear groups forming. This tool creates a "seating chart" for the filaments. It says, "Okay, all the Waltz dancers go to Table A, the Salsa dancers to Table B." This allows scientists to separate the different types of filaments automatically, so they can build a perfect model for each type.

3. The "Noise-Canceling Headphones" (Amyloid-Specific Blush)

The Problem: When scientists try to build the 3D model, the computer sometimes gets "too creative." Because the data is so noisy, the computer might invent details that aren't really there (like seeing a face in the clouds). This is called "overfitting."
The Solution: They used a "denoising" neural network (a type of AI) to clean up the image. But the old AI was trained on round, ball-shaped proteins. Amyloids are long and flat, like spaghetti. The old AI didn't know how to handle them well.
The Analogy: Think of the old AI as a music producer who only knows how to mix pop songs. If you give them a heavy metal track, they might mess it up. The authors took that producer and gave them a crash course specifically on "Spaghetti Rock" (amyloids). They re-trained the AI using thousands of amyloid examples. Now, when the AI hears the noise, it knows exactly how to filter it out without accidentally deleting the real "music" (the protein structure) or inventing fake notes.

4. The "Assembly Line" (Automated Pre-processing)

The Problem: Usually, processing these images requires a human to sit there, click buttons, check results, and adjust settings for hours or days.
The Solution: They built a fully automated assembly line.
The Analogy: Instead of a human chef chopping, frying, and plating every dish one by one, they built a robotic kitchen. You dump the raw ingredients (the raw microscope photos) into the machine, press "Start," and the robot automatically cleans the photos, finds the filaments, sorts them, and serves up the final 3D model. You can even set it up to work while the microscope is still taking pictures, so you get results in real-time.

The Results: What did they find?

Using these new tools, the team tested two different protein samples:

1. Tau (Alzheimer's): The tools worked perfectly, quickly sorting the filaments and building a crystal-clear 3D model.
2. hIAPP (Diabetes): This was much harder. The "Rhythm Detector" found that some filaments were "out of tune" (they didn't have the strong 4.75 Å beat). The team realized these "out of tune" filaments were actually damaged or less organized. By ignoring the bad ones and focusing on the good ones, and by using the "Sorter" to separate the different shapes, they discovered two brand new structures of amyloid filaments that had never been seen before.

The Bottom Line

This paper isn't just about new math; it's about giving scientists a better set of glasses, a smarter sorting machine, and a more experienced editor. These tools make it easier, faster, and more reliable to see the hidden structures of disease-causing proteins, which is a huge step forward in understanding and eventually treating diseases like Alzheimer's and diabetes.

Technical Summary

1. Problem Statement

Cryo-electron microscopy (cryo-EM) has revolutionized the structural biology of amyloid filaments, yet their reconstruction remains more challenging than that of globular proteins due to specific structural and algorithmic hurdles:

• Alignment Issues: Amyloid filaments lack low-resolution features along their helical axis, making alignment prone to errors and trapping refinements in incorrect local minima.
• Structural Heterogeneity: A single sample often contains mixtures of distinct filament types (polymorphs) that must be separated to achieve high-resolution reconstructions.
• Overfitting: Difficult datasets with low signal-to-noise ratios (e.g., small ordered cores) are susceptible to overfitting during refinement, leading to artifacts like "dotty" densities or radial streaks.
• Workflow Fragmentation: Existing workflows often require significant manual intervention for filament picking, type separation, and initial model generation, hindering high-throughput and objective analysis.

2. Methodology

The authors introduce a suite of integrated tools within RELION-5.1 designed to automate and improve amyloid processing:

A. Amyloid-Specific Automated Filament Picker

• Principle: Exploits the characteristic 4.75 Å repeat signal (cross-strand spacing) inherent to $\beta$-sheet amyloid structures in the Fourier domain.
• Algorithm:
  1. Micrographs are downsampled to 2.1 Å/pixel.
  2. 1D arrays are extracted in 36 directions; 1D Fourier transforms calculate accumulated power in the 4.65–4.85 Å range.
  3. A Figure-of-Merit (FOM) image and a PSI (phase shift/in-plane rotation) image are generated.
  4. Ice Rejection: Pixels with high power in the 4.2–4.4 Å range (characteristic of ice crystals) are filtered out.
  5. Neural Network: A modified U-Net (using instance normalization and GELU activation) processes the FOM/PSI pairs to output a segmentation mask.
  6. Post-processing: The mask is skeletonized and pruned to generate filament paths stored in a RELION STAR file.

B. Automated Pre-processing Pipeline

Two integrated "Schemes" (workflows) are introduced:

• amyprep: Runs on CPU. Iteratively imports movies, corrects motion, estimates CTF, runs the auto-picker (generating FOM/PSI maps), and selects micrographs with strong amyloid signals (skewness > 1).
• amyproc: Runs on GPU. Reads FOM/PSI maps, traces filaments via the neural network, extracts particles, and performs 2D classification in batches.

C. Bi-Hierarchical Clustering for Filament Type Selection

• Concept: Based on the observation that segments from a single filament usually belong to the same structural type.
• Method: Instead of standard k-means/PCA (used in the CHEP algorithm), the authors use bi-hierarchical clustering on a matrix of (Filament ID $\times$ 2D Class ID) particle counts.
• Visualization: Uses seaborn's clustermap with cosine distance to visualize blocks of filament types, allowing users to visually select subsets of data corresponding to specific polymorphs.

D. Amyloid-Specific Blush Regularisation

• Background: "Blush" is a denoising neural network used to regularize reconstructions. The original version was trained on general globular proteins.
• Improvement: The authors re-trained the network on 318 high-quality amyloid reconstructions (from the Amyloid Atlas).
• Training Constraints:
  • Restricted rotation augmentation to 90-degree steps around the helical axis to learn the 4.75 Å periodicity.
  • Removed anisotropy in map filters (amyloids are uniformly sampled around the axis).
  • Ensured masks extended to the box edges in the Z-direction.

3. Key Results

The tools were validated on two experimental datasets:

Dataset 1: Recombinant PAD12 Tau (Straightforward)

• Performance: The amyprep and amyproc pipelines processed 651 micrographs and ~477k segments in ~3 hours.
• Outcome: Automated picking was highly accurate. Bi-hierarchical clustering confirmed a single filament type.
• Resolution: Final 3D refinement using amyloid-specific Blush regularisation achieved 3.2 Å resolution.

Dataset 2: Recombinant hIAPP S20G (Challenging)

• Complexity: A time-resolved dataset containing multiple filament polymorphs.
• Discovery: Bi-hierarchical clustering identified eight distinct clusters.
  • Confirmed known structures (4PFCU, 3PFCU, 2PFCU, 2PFL).
  • Novel Structures: Identified two new 3-protofilament structures: 3PFLU (L+L+U shape) and 3PFLJ (L+L+J shape), achieving resolutions of 3.0–3.6 Å.
• FOM Signal Analysis: Filaments lacking strong 4.75 Å FOM signals (missed by the auto-picker) were manually traced. While they yielded good 2D classes, their 3D refinements stalled at 4.8 Å, failing to resolve amyloid layers. This suggests the auto-picker effectively filters out disordered or damaged filaments that would otherwise degrade high-resolution reconstruction.

Regularisation Comparison

• Overfitting Prevention: In difficult datasets (e.g., Pick-seeded tau), standard Blush regularisation resulted in "dotty" main-chain densities. The amyloid-specific Blush network produced continuous main-chain density and better side-chain definition, effectively preventing overfitting without losing signal.

4. Significance and Impact

• Automation and Objectivity: The integration of these tools into RELION-5.1 enables fully automated, "on-the-fly" pre-processing, reducing human bias and enabling high-throughput screening of assembly conditions.
• Robustness to Heterogeneity: The bi-hierarchical clustering tool provides a robust, visual method to separate complex mixtures of filament types, facilitating the discovery of rare polymorphs without multiple rounds of 3D classification.
• Improved Resolution: The amyloid-specific Blush network addresses the specific challenge of overfitting in low-signal amyloid cores, pushing the boundaries of what is achievable for difficult datasets.
• Quality Control: The reliance on the 4.75 Å signal acts as a built-in quality control, automatically excluding disordered filaments that would otherwise limit resolution.

In summary, this work provides a comprehensive, automated, and optimized software suite that significantly lowers the barrier to high-resolution amyloid structure determination, making the process more objective, efficient, and capable of resolving complex structural heterogeneity.