TL;DR
This paper introduces Smart Leginon, a fully automated system for cryo-electron microscopy (cryoEM) grid screening. By integrating machine learning and advanced image processing, Smart Leginon drastically reduces the time and human effort required for this crucial step, significantly boosting the throughput of cryoEM facilities.
What Problem Does This Paper Address?
Single-particle cryo-electron microscopy (cryoEM) is a powerful technique for determining the 3D structure of biological molecules. However, a major bottleneck in the cryoEM workflow is the screening of prepared grids. This process involves meticulously examining grids under a microscope to identify optimal areas for high-resolution data collection. Factors like grid type, sample concentration, ice thickness, and particle behavior must all be optimized, often requiring the screening of many grids. This manual screening is time-consuming, labor-intensive, and requires significant operator expertise, limiting the overall throughput of cryoEM services. The paper aims to automate this process, making cryoEM more accessible and efficient.
Methods
The core of this work is the development of "Smart Leginon," an extension of the existing Leginon data-collection software. Key methodological advancements include:
- Machine Learning Integration: The paper incorporates machine learning (ML) algorithms, specifically from the "Ptolemy" package, for image analysis and target identification. This allows the system to "learn" and emulate the decision-making process of an expert operator.
- Automated Grid Handling: Smart Leginon Autoscreen is designed to work seamlessly with automated specimen-exchange systems. It can load, screen, and unload multiple grids sequentially without human intervention.
- Multi-Scale Imaging (MSI) Automation: The system automates the entire MSI workflow, from collecting low-magnification "atlas" images of the grid to identifying promising squares, then zooming in on holes within those squares, and finally acquiring exposure magnification images.
- Novel Image Processing Algorithms: Beyond ML, the paper details purpose-built image-processing algorithms within Leginon for tasks like merging partial squares at tile edges and identifying holes even in challenging conditions (e.g., varying contrast, contamination).
- Parameter-less Hole Finding: A significant aspect is the use of a parameter-less hole finder from Ptolemy, which can reliably identify holes across a wide range of grid types and conditions without requiring manual parameter tuning.
Datasets
The paper primarily uses cryoEM grids prepared with mouse apoferritin (mApof) for evaluation.
- mApof Grids: 11 grids were prepared and screened to compare Smart Leginon Autoscreen performance against human operators.
- Real-world User Grids: The utility of Smart Leginon was further demonstrated using three real-world scenarios involving a total of 35 user grids, showcasing its applicability to diverse samples and grid preparation conditions.
The paper does not introduce a new benchmark dataset but rather uses existing cryoEM grid preparation and imaging protocols for evaluation.
Key Insights and Ideas
- Emulating Human Expertise: The most striking idea is how ML can effectively mimic the nuanced decision-making of experienced cryoEM operators during the screening process.
- Automation as a Throughput Multiplier: By automating screening, the system frees up valuable microscope time and operator time, directly addressing a major bottleneck.
- Robustness Across Diverse Conditions: The parameter-less hole finding and automated targeting demonstrate remarkable robustness, working well even on grids with challenging characteristics (e.g., varying ice thickness, contamination, contrast reversal).
- Rapid Feedback Loop: The automated screening and integration with post-processing tools enable a rapid feedback loop for grid and sample optimization, accelerating the entire cryoEM project.
Key Results
- Drastic Time Reduction: Smart Leginon Autoscreen reduced operator time for screening 11 grids from approximately 6 hours to less than 10 minutes for setup.
- Comparable Image Quality: The percentage of "good" holes identified by Autoscreen (95.9%) was comparable to or better than the average of 5 expert operators (90.6%), and comparable to the best operator (95.2%).
- Similar Resolution Estimates: CTF resolution estimation for Autoscreen-targeted holes (7.3 ± 2.6 Å) was comparable to that of operator-targeted holes (7.4 ± 2.9 Å).
- Efficient Grid Utilization: In a real-world scenario, Smart Leginon facilitated the preparation of 12 Krios-ready grids from 35 initial grids over 4 days, significantly improving efficiency.
- Successful Data Collection: The system was successfully used to collect data leading to a 3.1 Å structure, demonstrating its practical utility.
Limitations and Future Work
- Square Selection Limitations: The paper acknowledges limitations in square selection, particularly with highly unbalanced distributions of square areas (e.g., many dry squares vs. few with ice) and when grids have cracks that might be misclassified. Solutions like adding a high-pass filtering step, manual curation of training data, and increasing grid tile overlap are suggested.
- Hole Selection Limitations: Current hole selection struggles with significantly tilted or bent grids and when holes deviate substantially from the training set. The authors suggest stretching images for tilted grids as a potential workaround.
- Generalizability of ML Models: While Ptolemy's parameter-less approach is robust, the paper notes that the success rate of targeting can be influenced by the proportion of relevant grids in the training set.
- Future Work: The authors envision further integration with live-processing software for automated post-processing and using live-processing results as targeting priors to further improve the workflow. They also suggest that the ML and image processing routines could be integrated into other cryoEM software packages.
Summary updated