MTDL-GAN: De novo Design of Multi-Target Directed Ligands for Alzheimer’s Disease from Unpaired Sets of Target-Focused Chemical Library

BACKGROUND. Alzheimer’s disease (AD), the most common form of dementia, causes memory loss, cognitive decline, and behavioural changes, affecting over 32 million people globally. Current AD treatments that focus on single-target intervention often fail to significantly slow disease progression and may not be effective for all patients. Given AD's complex nature, a more effective approach may involve targeting multiple pathways simultaneously. METHODS. This study proposes the use of a cycle-consistent adversarial network to design multi-target directed ligands (MTDL-GAN), drugs designed to inhibit two primary AD target enzymes simultaneously. Our targets of interest are acetylcholinesterase (AChE), beta-secretase 1 (BACE1), and glycogen synthase kinase 3 beta (GSK3), which are known to have significant impacts on the progression and development of AD, each with a different mechanism of action. Inhibitor libraries were curated from ChEMBL27 and further characterized to represent each inhibitor domain, resulting in 69 AChE, 572 BACE1, and 246 GSK3 inhibitors. The MTDL-GAN was trained on these unpaired datasets to generate MTDL-like molecules. During the model training, all generated molecular structures were stored to create in silico MTDL libraries. From each MTDL library, 300 molecules were sampled for further analysis. Tanimoto similarity scores were computed to assess the structural similarity of the generated molecules to the target inhibitor domain, and molecular docking simulations were carried out to validate their \textit{in silico} binding affinities to the target enzymes. RESULTS. The proposed method effectively transformed molecules from the original inhibitor domains into MTDL-like molecules while maintaining their structural similarity to the original inhibitors. Less than 0.15% of the generated molecules have a Tanimoto similarity score above 0.85 to the known bioactive molecules in ChEMBL27, highlighting their structural novelty and potential for exploring new chemical space. Further investigation revealed that the sampled MTDLs demonstrate promising in silico dual-binding affinity while having favourable physicochemical properties and synthetic tractability. Among these, several MTDLs surpassed the in silico binding affinities of investigational drugs in phase 2/3 clinical trials. Molecular structures of these hit MTDLs are publicly available for academic research.