Neurodegenerative diseases, such as Alzheimer's, Parkinson's, and Huntington's, afflict tens of millions of patients worldwide. They are characterized by protein aggregation and progressive neuronal loss, leading to cognitive and motor impairments, and ultimately, death. Despite decades of research, effective treatments remain elusive, largely due to the complex and poorly understood mechanisms underlying these conditions.
The limitations associated with fluorescent labeling underscore the need for alternative label-free techniques to provide a more comprehensive understanding of the dynamics of protein aggregation, without introducing unwanted perturbations. This thesis introduces LINA, a label-free deep learning method with 96% accuracy for detecting and analyzing aggregates in living cells. Leveraging convolutional neural networks, LINA circumvents the challenges of fluorescent labeling and enables quantitative analysis of aggregate morphology, area, and dry mass. Studying different kinds of label-free Huntingtin protein aggregates using LINA reveals correlations between their ultrastructures and aggregation mechanisms. LINA allows for precise, non-invasive measurements of aggregate growth dynamics in live-cell imaging. Generalizable across imaging conditions, aggregate types, and cell lines, LINA promises a streamlined, highly-specific, automated, gentle, and high-fidelity approach for neurodegenerative disease research.
To address the challenges posed by the rapid and transient nature of protein aggregation, we present a vision transformer model that is 91% accurate at predicting aggregation onset from images of soluble protein. Integrating it into a modular, self-driving microscope pipeline enables the autonomous detection and capturing of aggregation events in real time. The autonomous microscope supports multiple imaging modalities and is fully customizable by the user, allowing optimized scans, tailored to each experiment. Our approach offers a versatile and easy-to-use tool for monitoring aggregation kinetics and dynamics across diverse imaging setups. For the first time in the field, we enable intelligent Brillouin imaging, which can autonomously detect the onset of aggregation and thereby retrieve the viscoelastic properties throughout the process. Furthermore, by integrating an image-classification version of LINA into our self-driving Brillouin microscope, we can detect aggregates from brightfield images in real time to avoid the typical reliance on a fluorescence marker, preserving sample health.
LINA's extensibility offers promising avenues for future applications, including characterizing aggregates in other neurodegenerative diseases and correlating with other imaging techniques for comprehensive molecular interaction studies. Additionally, LINA's compatibility with high-throughput screening processes provides a label-free alternative for drug discovery, allowing detailed assessments of drug effects on aggregates' size, morphology, propensity, and biomechanical properties. Autonomous microscopy can deepen our understanding of protein aggregation kinetics and aid in studying the dynamic interactions between the aggregating protein and drug candidates or other proteins, organelles or subcellular structures. The advancements presented in this thesis mark a significant step toward fully automated and label-free methodologies that enhance the throughput, accuracy, efficacy, and scope of neurodegenerative disease research.