Development and application of a virtual screening campaign integrating machine learning and chemoinformatics approaches to identify potential tau aggregation inhibitors

Tau is a microtubule-associated protein (MAP) mainly expressed within neurons in the central nervous system (CNS), which plays a central role in stabilizing microtubule dynamics. Physiologically, Tau’s function is regulated by a series of post-translational modifications including phosphorylation, which modulates its binding to microtubules. Under pathological conditions, Tau can undergo a series of post-translational modifications promoting its detachment from microtubules and aberrant aggregation into insoluble, toxic structures including pair helical filaments (PHFs) and neurofibrillary tangles (NFTs) within the cell cytoplasm. Such protein aggregates are currently recognized as key neuropathological hallmark of tauopathies, which is a heterogeneous group of neurodegenerative diseases including Alzheimer’s disease (AD). Currently approved therapeutic approaches focus on managing symptoms rather than altering the underlying disease process, unfortunately showing low efficacy especially at the advanced stages of neurodegeneration. The lack of effective disease treatments and increasing prevalence of tauopathies underscore the urgency of developing novel therapeutic strategies. In the context of multifactorial diseases such as tauopathies, novel valuable therapeutic opportunities may arise from the design of compounds simultaneously acting on multiple pathways involved tau’s aggregation, in line with the polypharmacology concept. The design of multi-target drug candidates is particularly challenging. However, currently available in silico approaches can offer a rapid, cost-effective means to design drug-like, multi-target compounds. Indeed, these methods can leverage currently available data to enhance compounds’ polypharmacology profile while predicting some of their potential off-target effects. In particular, in recent years, chemoinformatic approaches have assumed a central role in drug discovery; combining molecular descriptors analyses, similarity estimations and predictive modeling demonstrated to enable researchers to prioritize promising sets of candidates for experimental validation, significantly reducing costs and time at the early stages of drug discovery. In addition, machine learning modeling approaches are now routinely employed to detect complex and non-linear relationships between molecular features and biological activity. This thesis frames within a project under development in the Molecular modelling and Drug Design Lab which aims at designing molecules able to inhibit the aggregation of Tau within neurons. In particular, during my internship thesis I performed extensive in silico analyses on tau aggregation inhibitors reported within the ChEMBL database. The study involved the retrieval, filtering and preparation of a high-quality dataset of compounds with a documented activity against tau. Molecular descriptors and fingerprints representations were then computed to identify key chemical and structural properties, to be used for training machine learning models for tau aggregation activity prediction. Model performance was assessed using accuracy and the Matthews Correlation Coefficient (MCC) metrics. The best-performing model was finally applied within an in silico strategy integrating chemoinformatics and machine learning predictions to identify potential tau aggregation inhibitors among of commercially available compounds.