Abstract
OBJECTIVE: Alzheimer's disease poses a significant threat to human health. Currenttherapeutic medicines, while alleviate symptoms, fail to reverse the disease progression or reduce its harmful effects, and exhibit toxicity and side effects such as gastrointestinal discomfort and cardiovascular disorders. The major challenge in developing machine learning models for anti-acetylcholinesterase peptides discovery is the limited availability of active peptide data in public databases. This study primarily aims to address this challenge and secondarily to discover novel, safer, and less toxic anti-acetylcholinesterase peptides for better Alzheimer's disease treatment. METHODS: A Random Forest Classifier model was constructed from a hybrid dataset of non-peptide small molecules and peptides. It was applied to screen a custom peptide library. The binding affinities of the predicted peptides to acetylcholinesterase were assessed via molecular docking, and top ranked peptides were selected for experimental assay. RESULTS: The top six peptides (IFLSMC, WCWIYN, WIGCWD, LHTMELL, WHLCVLF, and VWIIGFEHM) were selected for experimental validation. Their inhibitiory effects on acetylcholinesterase were determined to be 0.007, 3.4, 1.9, 10.6, 1.5, and 3.9 μmol/L, respectively. DISCUSSION: Predicting anti-acetylcholinesterase peptides is challenging due to the absence of a comprehensive, publicly accessible peptide database. Traditional approaches using only non-peptide small molecules for model construction often have poor performance on predicting active peptides. Here, we developed a machine-learning model from a hybrid dataset of non-peptide small molecules and peptides, which find six potent peptides. This model was as/superior accuracy compared to small-molecule-only models reported before, but has a significant higher capability of discriminating active peptides. Our work shows that hybrid datasets can boost machine-learning model prediction in peptide drug discovery.