The identification of effective inhibitors targeting β-site amyloid precursor protein cleaving enzyme-1 (BACE-1) is crucial for developing therapeutic strategies for Alzheimer’s disease. This study developed a structure-based computational framework for predicting BACE-1 inhibitory activity using both deep learning and conventional machine learning techniques. A publicly available BACE-1 dataset with chemical structures defined in SMILES (Simplified Molecular Input Line Entry System) format was subjected to feature extraction using the RDKit program. Global molecular characteristics and substructural information were captured using both molecular fingerprint representations and physicochemical descriptors. Circular (Morgan/ECFP4) fingerprints, RDKit fingerprints, and MACCS keys were used to encode molecular substructures into binary vectors. Subsequently, Support Vector Machines (SVM), k-Nearest Neighbors (kNN), deep neural networks (DNN), and enhanced deep neural networks were trained and validated using these features under the same experimental conditions. Confusion-matrix analysis and standard classification metrics (accuracy, precision, recall, and F1-score) were used to assess the model’s performance. Deep learning models outperformed traditional machine learning techniques in capturing intricate nonlinear structure–activity correlations, according to comparison research. The proposed enhanced DNN demonstrated balanced precision and recall across both classes and achieved an accuracy of 0.99 on a test set of 303 molecules, including 138 active inhibitors and 165 inactive non-inhibitors. All things considered, these results imply that deep learning models, in conjunction with molecular fingerprints, offer a robust and reliable approach to BACE-1 inhibitor prediction and could accelerate early-stage virtual screening. All experiments were conducted using a fixed random seed and a held-out random split to ensure reproducibility.
Yılmazcan et al. (Wed,) studied this question.