Comparative Deep Learning Models for Predicting AChE Inhibitory Activity from Molecular Structure Data

Abstract

Acetylcholinesterase (AChE) inhibitors are essential in treating neurodegenerative conditions like Alzheimer's disease. Traditional computational methods for screening AChE inhibitory activity, such as QSAR and molecular docking, often struggle with accuracy and efficiency. This study proposes an enhanced deep learning framework incorporating state-of-the-art neural architectures including Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), and Recurrent Neural Networks (RNN) for predicting AChE inhibitory activity based on molecular structure representations derived from SMILES using ECFP4 fingerprints. Using a curated dataset from the ChEMBL database, we compare the predictive performance of these architectures under a unified preprocessing and evaluation pipeline. Our findings reaffirm the superiority of CNNs, achieving 84.34% accuracy and outperforming other models across key metrics. The study advances computational drug discovery efforts by offering a reproducible, scalable, and interpretable approach for virtual screening of potential AChE inhibitors.