Preview

PDF, English - main document Download (80MB) | Lizenz: Self-supervised deep learning for advancing macromolecular analysis in cryo-electron tomography by Stojanovska, Frosina underlies the terms of Creative Commons Attribution-NonCommercial-NoDerivatives 4.0

Abstract

Cryo-electron tomography (cryo-ET) provides unprecedented insights into cellular structures and macromolecular complexes in their native states. However, its interpretation remains challenging due to high noise levels, low contrast, and structural complexity. This thesis presents CryoSiam, a novel self-supervised deep learning framework that addresses these challenges through denoising, semantic segmentation, and particle identification tasks.

Trained entirely on simulated tomograms, CryoSiam leverages self-supervised learning to generate robust voxel- and subtomogram-level embeddings, circumventing the need for annotated real data. Comprehensive ablation studies identified key design choices that optimize performance across tasks. When applied to publicly available real datasets from EMPIAR and the CryoET Data Portal, CryoSiam demonstrated effective generalization to real-world conditions, achieving performance comparable to, and at times exceeding, state-of-the-art supervised methods.

The framework showed versatility in delivering high-quality noise suppression, accurate membrane segmentation without real-data training, and reliable particle identification based on learned embeddings. These results highlight the potential of self-supervised learning to bridge the gap between simulated training environments and real cryo-ET applications.

This thesis also addresses existing limitations, such as reliance on simulated data and challenges in representing structural diversity. Future directions include expanding simulation diversity, exploring semi-supervised approaches, and enhancing computational efficiency. CryoSiam establishes a foundation for advancing cryo-ET analysis, promoting open science, collaboration, and a deeper understanding of cellular architecture at molecular resolution.