TL;DR
This paper introduces "Smart Leginon Autoscreen," a fully automated system for cryo-electron microscopy (cryo-EM) grid screening that significantly reduces operator time and increases throughput. By integrating machine learning and advanced image processing, it automates the tedious process of identifying optimal grids for high-resolution data collection, making cryo-EM more accessible and efficient.
What Problem Does This Paper Address?
Cryo-electron microscopy (cryo-EM) is a powerful technique for determining the structure of biological molecules, but its widespread adoption is hampered by a significant bottleneck: grid screening. Preparing and evaluating cryo-EM grids is a time-consuming and labor-intensive process. Researchers must optimize numerous parameters, such as grid type, ice thickness, sample concentration, and particle behavior, to find grids suitable for high-resolution data collection. This often involves manually screening many grids, consuming valuable microscope time and operator expertise. The paper aims to automate this critical step, making cryo-EM more accessible and efficient, especially for new projects or when dealing with challenging samples.
Methods
The core of this work is the development of Smart Leginon Autoscreen, an extension of the existing Leginon software. Key methodological advancements include:
- Machine Learning (ML) Integration: The system incorporates ML algorithms, specifically from the Ptolemy package, for automated image processing and microscope control. This allows the software to "learn" and emulate the decision-making process of an expert operator.
- Automated Grid Handling: Smart Leginon Autoscreen seamlessly integrates with automated specimen exchange systems, allowing for unattended loading and unloading of multiple grids.
- Multi-Scale Imaging (MSI) Workflow: The system automates a comprehensive MSI workflow, starting from low-magnification "atlas" images to identify promising squares, then moving to higher magnifications to assess holes within those squares, and finally to "exposure" magnification for detailed analysis of particle behavior.
- Ptolemy Hole Finder: A key component is Ptolemy's parameter-less hole finder, which is robust across various grid types (carbon and gold) and challenging imaging conditions (e.g., varying contrast, thick ice, contamination).
- Autoscreen Workflow: A command-line interface and GUI enable users to set up entire multi-grid screening sessions quickly and run them unattended.
Datasets
The paper primarily uses mouse apoferritin (mApof) grids for evaluation. Specifically, they screened:
- Eleven unseen mApof grids for a direct comparison between Smart Leginon Autoscreen and five expert human operators.
- Three real-world multi-grid user screening/collection sessions involving various grid types (Quantifoil R1.2/1.3, Quantifoil R1.2/1.3 with graphene, UltrAuFoil R1.2/1.3) and sample conditions.
- A 35-grid screening session across three samples.
- Eight de novo user grids of an unspecified sample.
- Four freshly-made cryo-EM grids on a Krios.
No new datasets were introduced for training ML models; rather, existing software and data were leveraged.
Key Insights and Ideas
- Automation Emulates Expertise: The most significant insight is that ML-driven automation can effectively replicate and even surpass the performance of experienced human operators in grid screening, a task previously thought to require significant human judgment.
- Bottleneck Reduction: Grid screening is a major bottleneck in cryo-EM, and automating it dramatically frees up operator time and microscope availability for other crucial steps like sample preparation and data analysis.
- Parameter-less Hole Finding is Robust: The parameter-less nature of the Ptolemy hole finder is a critical advantage, making it adaptable to diverse grid types and imaging conditions without manual tuning.
- Idle Time is Valuable: The system can effectively utilize idle microscope time (e.g., overnight, weekends) for screening, significantly boosting overall throughput.
Key Results
- Operator Time Reduction: Smart Leginon Autoscreen reduced operator time for screening 11 grids from an average of 6 hours to less than 10 minutes for setup, with unattended screening taking approximately 5.4 hours for the entire batch.
- Comparable/Improved Hole Quality: Autoscreen achieved a 95.9% rate of "good" holes, outperforming the average operator rate of 90.6% and matching the best operator's performance.
- Similar Resolution and Ice Thickness: CTF resolution estimations for Autoscreen holes (7.3 ± 2.6 Å) were comparable to operators (7.4 ± 2.9 Å), and ice thickness estimates were also similar (Autoscreen: 32.9 ± 7.1 nm vs. operators: 35.0 ± 18.7 nm).
- Robustness Across Grids: The system demonstrated successful hole targeting on various grid types (carbon, gold) and under challenging conditions without parameter adjustments.
Limitations and Future Work
- Square Selection Limitations: The authors acknowledge current limitations in square selection, particularly with grids exhibiting large ice thickness gradients, cracked squares, or slight misalignments in merged atlas tiles.
- Hole Selection Limitations: For hole selection, Ptolemy's performance can be compromised with holes that deviate significantly from the training set, tilted or bent grids, or lacey grids.
- Future Work: The authors are actively working on addressing these square and hole selection limitations to improve generalization. They also envision extending the workflow with live-processing feedback for even more intelligent targeting and collection. The potential for integrating these ML and image processing routines into other collection software packages is also highlighted.
Summary updated