Logo Logo
Hilfe
Kontakt
Switch language to English
Advancing single-molecule imaging analysis via deep learning
Advancing single-molecule imaging analysis via deep learning
Single-molecule experiments have revolutionized our understanding of the physical world, offering unparal- leled insights into dynamic processes. However, a bottleneck persists in the time-consuming and potentially biased nature of data analysis. The main goal of this thesis is to address these issues through the devel- opment of deep learning techniques tailored for the analysis of fluorescence data specifically focusing on surface-based single-molecule measurements of Förster resonance energy transfer (FRET). The culmination of this effort is Deep-LASI (Deep-Learning Assisted Single-molecule Imaging analysis), a software suite lever- aging the predictive capabilities of deep neural networks (DNNs). Designed for rapid analysis of single-color, two-color FRET, and three-color FRET data, Deep-LASI fully automates trajectory sorting and FRET correc- tion factor determination, followed by the automated prediction of observed states and state dwell times for each molecule. The pre-trained ensemble of DNNs are able to analyze previously unseen data sets in approxi- mately 20–100 ms per trajectory. In extensive benchmarking, the DNNs demonstrated their efficacy through ground truth simulations and comparisons with manually analyzed experimental data, validated by expert users. Beyond the development of these deep learning techniques, Deep-LASI has evolved into a compre- hensive software suite that provides robust methods for extracting raw intensity data from single-molecule movies across multiple channels. A key feature was the integration of alternative approaches for user in- tervention, applicable to every step that the DNNs undertake automatically. This user-centric framework of Deep-LASI encompasses human evaluation of single-molecule trajectories, offering flexibility to override DNN classifications, and the option to employ hidden Markov models (HMM) for the analysis of kinetic rates, along with various downstream analysis methods post trajectory sorting. Additionally, the evaluation of various software tools for extracting kinetic rate constants from single-molecule FRET trajectories is pre- sented in this thesis. By analyzing specific data sets with different levels of complexity, the comparison of all employed methods shed light on their limitations and revealed important aspects that need to be considered for consistent analysis results. Lastly, established computational methods were coupled with experimen- tal data to elucidate the conformational dynamics of bacterial adhesin SdrG, which can form an extremely mechanostable complex with its target peptide human fibrinogen β (Fgβ). The SdrG:Fgβ complex can with- stand forces greater than 2 nN, representing the strongest non-covalent bond of known to date. Combining molecular dynamics simulations with single-molecule FRET measurements provided new insights into the behavior of the locking strand and ligand-induced structural changes of the SdrG protein.
Deep Learning, Single-molecule Imaging, Single-molecule FRET
Wanninger, Simon Martin
2024
Englisch
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Wanninger, Simon Martin (2024): Advancing single-molecule imaging analysis via deep learning. Dissertation, LMU München: Fakultät für Chemie und Pharmazie
[thumbnail of Wanninger_Simon.pdf]
Vorschau
PDF
Wanninger_Simon.pdf

67MB

Abstract

Single-molecule experiments have revolutionized our understanding of the physical world, offering unparal- leled insights into dynamic processes. However, a bottleneck persists in the time-consuming and potentially biased nature of data analysis. The main goal of this thesis is to address these issues through the devel- opment of deep learning techniques tailored for the analysis of fluorescence data specifically focusing on surface-based single-molecule measurements of Förster resonance energy transfer (FRET). The culmination of this effort is Deep-LASI (Deep-Learning Assisted Single-molecule Imaging analysis), a software suite lever- aging the predictive capabilities of deep neural networks (DNNs). Designed for rapid analysis of single-color, two-color FRET, and three-color FRET data, Deep-LASI fully automates trajectory sorting and FRET correc- tion factor determination, followed by the automated prediction of observed states and state dwell times for each molecule. The pre-trained ensemble of DNNs are able to analyze previously unseen data sets in approxi- mately 20–100 ms per trajectory. In extensive benchmarking, the DNNs demonstrated their efficacy through ground truth simulations and comparisons with manually analyzed experimental data, validated by expert users. Beyond the development of these deep learning techniques, Deep-LASI has evolved into a compre- hensive software suite that provides robust methods for extracting raw intensity data from single-molecule movies across multiple channels. A key feature was the integration of alternative approaches for user in- tervention, applicable to every step that the DNNs undertake automatically. This user-centric framework of Deep-LASI encompasses human evaluation of single-molecule trajectories, offering flexibility to override DNN classifications, and the option to employ hidden Markov models (HMM) for the analysis of kinetic rates, along with various downstream analysis methods post trajectory sorting. Additionally, the evaluation of various software tools for extracting kinetic rate constants from single-molecule FRET trajectories is pre- sented in this thesis. By analyzing specific data sets with different levels of complexity, the comparison of all employed methods shed light on their limitations and revealed important aspects that need to be considered for consistent analysis results. Lastly, established computational methods were coupled with experimen- tal data to elucidate the conformational dynamics of bacterial adhesin SdrG, which can form an extremely mechanostable complex with its target peptide human fibrinogen β (Fgβ). The SdrG:Fgβ complex can with- stand forces greater than 2 nN, representing the strongest non-covalent bond of known to date. Combining molecular dynamics simulations with single-molecule FRET measurements provided new insights into the behavior of the locking strand and ligand-induced structural changes of the SdrG protein.