In silico design toward the development of MAO-B selective covalent inhibitors based on rasagiline

Abstract

Rasagiline is a covalent MAO-B inhibitor used in the treatment of Parkinson's disease. Previous studies have shown that modifying rasagiline at the C4 position can extend the ligand toward the entrance cavity through the gating residues Ile199 and Tyr326. This extension may lead to formation additional interactions in the entrance cavity and strengthen binding and reducing side effects of Rasagiline. Reducing side effects is particularly important for neurodegenerative disease therapies because these drugs are typically used long term; improving tolerability can directly enhance patients' quality of life. In this thesis, we designed a series of C4-modified rasagiline derivatives based on the MAO-B active-site. In addition we also used the CReM molecular generator to create chemically meaningful rasagiline derivatives. In total, we built a virtual library of 15,837 C4-modified rasagiline compounds. To prioritize suitable candidates from this large library, we trained an XGBoost classifier. The training set was constructed from MAO-B inhibitors collected from the ChEMBL database with experimental IC₅₀ values. Compounds were labeled as active or inactive using a predefined activity threshold. Before screening the full library, we assembled a validation set from the literature consisting specifically of C4-modified rasagiline derivatives and evaluated the model on this set. We also performed a "spike-in" analysis by embedding the validation molecules into the library and checking whether the model ranked active compounds higher than inactive ones. We then benchmarked the machine-learning ranking against widely used structure-based methods (Glide XP and AutoDock Vina) for the same task. From the top 5% of the machine-learning ranked list, we generated a diverse shortlist of 150 molecules using clustering and diversity selection. Next, we identified consensus candidates that appeared across the ML shortlist and the top-ranked results from Glide XP and Vina, aiming to reduce bias from any single method. These consensus molecules and the top ML candidates were examined with docking to understand their binding modes and were further evaluated using Schrödinger Suite Prime MM-GBSA rescoring. Across the high-priority candidates, we observed consistent interaction patterns. In particular, many compounds formed favorable hydrophobic and aromatic contacts with key residues such as Tyr326 and Ile199. Finally, we assessed ADME-related properties of the leading candidates to evaluate their suitability as drug-like molecules. Overall, this thesis highlights promising C4-modified rasagiline scaffolds and structural patterns that can guide future MAO-B inhibitor development.