Natural compounds for Alzheimer's prevention and treatment: Integrating SELFormer-based computational screening with...

This study aimed to develop and apply a novel computational pipeline combining SELFormer, a transformer architecture-based chemical language model, with advanced deep learning techniques to predict natural compounds (NCs) with potential in Alzheimer's disease (AD) treatment. The NCs were identified based on activity related to seven AD-specific genes, including acetylcholinesterase (AChE), amyloid precursor protein (APP), beta-secretase 1 (BACE1), and presenilin-1 (PSEN1).

We implemented a computational pipeline using SELFormer and deep learning techniques, conducted optimal clustering and quantitative structure-activity relationship (QSAR) analyses, and performed a uniform manifold approximation and projection (UMAP) to categorize compounds based on bioactivity levels. Molecular docking analysis was carried out on selected compounds. To validate the computational predictions, we conducted in vitro studies using nerve growth factor (NGF)-differentiated PC12 cells. Finally, we mapped the relationships between food sources containing the identified compounds and their target proteins.

Optimal clustering analysis revealed five distinct groups of NCs, while QSAR analysis highlighted variations in molecular properties across clusters. The UMAP projection identified 17 highly active NCs (pIC₅₀>7). Molecular docking analysis showed that cowanin, β-caryophyllene, and L-citronellol demonstrated decreased binding energy across target proteins. In vitro studies confirmed significant biological activities of these compounds, including increased cell viability, decreased AChE activity, reduced lipid peroxidation and tumor necrosis factor (TNF)-α mRNA expression, and increased brain-derived neurotrophic factor (BDNF) mRNA expression compared to the control. The study also identified natural sources of these compounds, such as anatidae, mangosteen, and celery, providing insights into potential dietary interventions.

This integrated computational and experimental approach offers a promising framework for identifying potential NCs for AD treatment. The results contribute to exploring effective therapeutic strategies against AD.