Abstract
Cells consist of many different molecules that all associate together into complex structures required for cellular function. Yet, current scientific techniques face significant limitations in visualizing macromolecular interactions within their native context. Cryogenic electron tomography (cryo-ET) has emerged as a promising method to address these challenges, offering macromolecular imaging within unperturbed, cryogenically frozen cells. Despite its potential, the method is currently hindered by difficulties in automated macromolecular localization and identification.
To address the macromolecular annotation bottleneck, this thesis first focuses on establishing a benchmark for localization that enables systematic comparison of different methods. The benchmark was released in the 2021 SHape REtrieval Contest (SHREC’21) where it challenged researchers worldwide to improve their techniques. By combining multislice simulation with samples composed of macromolecules of a wide range of shapes and sizes, the dataset provided semi-realistic tomograms dataset with ground truth annotations. Competitors were provided a set of training tomograms with annotations, while their technique was evaluated on an unlabeled tomogram. Among the tested methods, deep-learning-based segmentation techniques, particularly employing the U-Net architecture, outperformed template matching (TM) baselines, benefiting from fully labeled ground truth training data and consistent data distribution. However, in practice, these methods rely on annotated experimental data, which remains scarce, whereas TM does not require any training data.
To enhance the utility of TM, which can exhaustively search for known macromolecular structures in tomograms, the algorithm is implemented with graphical processing unit (GPU) acceleration. This implementation provides a significant speed-up and demonstrates the influence of angular sampling on detection sensitivity. The results show that the sensitivity is dictated by the Crowther criterion, which relates the diameter of a macromolecule, resolution, and angular sampling. Increasing the angular sampling, significantly improved the classification performance on ribosomes. Furthermore, this thesis explored an automated method for threshold estimation and filtering for TM. By recognizing that true positives produce steep local maxima in the cross-correlation maps, a tophat transform was applied as an ad hoc solution to filter high signal-to-noise ratio (SNR) artifacts. This approach effectively retrieved ribosome and proteasome instances from both ex situ and in situ datasets. The update method is provided in the software pytom-match-pick.
Finally, this methodology was applied to an ex vivo study of ribosomes on the endoplasmic reticulum (ER) surface. By integrating GPU-accelerated TM in a pipeline with high-resolution subtomogram averaging, this study visualized 8 intermediate states of the ribosomal elongation cycle and a previously unknown intermediate associated with elongation factor eEF1a. Additionally, by correlating the ER-bound ribosome translocon states with their 3D spatial distribution, the study reveals that the multipass and OSTA translocons are uniquely recruited to distinct polysome chains. The ability to analyze the 3D spatial distribution of macromolecules thus enabled insights into the recruitments of translocon complexes.
The findings in this thesis contribute to the growing potential of cryo-ET to analyze the ultrastructure of cells at a molecular level. The developed methodology represents a robust software solution for macromolecular localization that is directly applicable to current data processing workflows.
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 5 Feb 2025 |
Place of Publication | Utrecht |
Publisher | |
Print ISBNs | 978-90-39378-01-4 |
DOIs | |
Publication status | Published - 5 Feb 2025 |
Keywords
- cryo-ET
- tomogram
- marcomolecule
- localization
- identification
- registration
- template matching
- ribosome
- GPU
- deep-learning