Protein hide-and-seek: AI-driven exploration of dementia biomarkers

Neurodegenerative diseases, such as Alzheimer’s disease, are the major preliminary cause of dementia. Research shows that physiological changes occur years before the onset of symptoms. Therefore, accurate clinical tests for early and discriminative diagnosis across various types of neurodegenerative diseases are crucial. For this, we need robust biomarkers, which represent measurable traits that can objectively detect and evaluate various biological processes. Early detection with sensitive biomarkers could allow effective treatment options and expand opportunities for monitoring and managing disease progression. Biofluid protein biomarkers are key to precision medicine, providing invaluable insights into underlying disease mechanisms. In body fluids, biomarkers coexist with a plethora of other molecules, potentially interfering with antibody-epitope interactions. Conformational changes or aggregation of proteins may further obstruct or disguise epitopes, preventing their accurate detection and quantification. Therefore, understanding the solvent accessibility of protein biomarkers, or in other words, the game of hide-and-seek is essential for accurate detection and quantification. This study bridges computational and experimental methodologies to delve into protein biomarkers associated with dementia. Due to the diversity of tools and topics involved in this research work, in Chapter 1, we first discuss all the main areas, advances in computational and wet-lab techniques, as well as knowledge gaps leading to the questions explored. In Chapter 2, we explored how hydrophobic (sticky) protein surfaces are and we analyse the implications of exposed hydrophobic surfaces. Chapter 3 attempts to demystify one of the possible mechanisms for hidden (inaccessible) epitopes. Chapter 4 presents a novel progression biomarker identification framework and protein leads for personalised prediction of cognitive decline in dementia. In the following chapter, we zoom in on one of the promising protein biomarkers for brain and spinal cord disorders - Glial fibrillary acidic protein (GFAP). In Chapter 5, we investigate the structural properties of GFAP under different conditions. In Chapter 6 we further discuss the GFAP proteoform puzzle and possible ways to leverage it for developing better clinical assays. In Chapter 7 our focus shifts towards foundation models and their effective fine-tuning for protein property prediction. Finally, in Chapter 8 we discuss research questions and discoveries of this thesis. We explore future directions and touch upon the importance of advances in AI approaches for multi-omics data analysis, and the importance of model transparency.