2022 Fiscal Year Annual Research Report
Intelligent cryo-electron microscopy of G protein-coupled receptors
| Project/Area Number |
22H02554
|
|---|
| Allocation Type | Single-year Grants |
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
Danev Radostin 東京大学, 大学院医学系研究科(医学部), 教授 (50415931)
|
|---|
| Project Period (FY) |
2022-04-01 – 2025-03-31
|
| Keywords | cryo-EM / automation / sample screening / data acquisition / deep learning |
|---|
| Outline of Annual Research Achievements |
In the first year of the project, we made good progress in this research. We established the main research platform for carrying out the project. This included selection and procurement of a powerful GPU workstation for the machine learning tasks related to automated sample screening and AI-based data acquisition. It was installed in March 2023 at the Kashiwa 2 campus where two new cryo-electron microscopes were also installed in the first half of 2022. The GPU machine is now connected to the microscope network and can communicate directly with the microscope and the data storage server. After installation, we began initial performance and stability testing and installation of the necessary software packages, which are still ongoing.
From the beginning of the project period, we started the development of the software platform for the intelligent cryo-electron microscopy system. It is written in Python and already has capabilities for analyzing sample maps, including identification and quantification of grid squares, accurate localization of support film holes, measurement of various quantitative parameters for each hole and enumeration of all holes and their properties in a global database for each grid. We also began testing the YOLO deep learning framework for applications in the GridNet neural network for classification of grid squares, support film holes, contaminants etc. The results are very encouraging, and we can accurately detect and classify grid squares. We also began testing linear and non-linear image analysis for quantification of the sample condition.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The research is progressing according to our plan. There were a few challenges related to the performance of the new CryoARM 200 electron microscope which will be used for the experimental testing in this project. The contamination rate on the sample was higher than expected and after consultations with the microscope manufacturer, we implemented bakeout procedures that reduced it to acceptable levels. There were also a few issues related to the stability of the camera and the data acquisition computer, but those were also mostly resolved. Consequent performance testing of the microscope system with an apoferritin test sample demonstrated a resolution of 1.54 A, which is a new world record for 200 kV cryo-electron microscopes.
|
| Strategy for Future Research Activity |
In the next year of the project, we plan to continue the development of the intelligent sample screening and automated data acquisition system. Detailed plan:
1. Develop further the grid, square, and hole analysis methods to achieve as detailed as possible quantification of the parameters of each sample. This will allow objective selection of the optimal areas for screening and data acquisition.
2. Continue the development of the GridNet and ScoreNet neural networks and test optimal training strategies based on human-annotated or quantitative image analysis-annotated grid maps. Human annotation is based on extensive experience with cryo-em samples but can be biased depending on the person’s experience. It also requires substantial effort. The advantage of using quantitative image analysis annotation is that it is from its inception “quantitative” and therefore allows objective analysis and tuning. A third approach is to combine the two methods by using quantitative image analysis to guide and augment the human manual labeling. We will experiment with the various options and determine which approach works best in practice for our application.
3. Start experimental trials of the automation system on the JEOL CryoARM 200 microscope. During experiments, the machine annotated sample grids could be corrected by the operator and this can then be used for RLHF (Reinforcement Learning from Human Feedback) tuning of the models.
4. We also plan to experiment with deep learning applications in cryo-tomography, where we recently made substantial performance improvement contributions.
|
Research Products
(6 results)
