Assessment of data-driven and deterministic approaches for predicting bace-1 inhibitor binding affinities

Abstract

BACE1 (beta-site amyloid precursor protein cleaving enzyme 1) is an aspartyl protease with a transmembrane domain that cleaves at the 671st position of amyloid precursor protein (APP) at the beta site. The production and release of amyloid-peptide, a pathological feature of Alzheimer’s disease (AD), is caused by the cleavage of APP by beta-secretase and then by the gamma-secretase complex. BACE1 inhibitors have shown significant promise in reducing amyloid-beta load in the brain and preventing the progression of Alzheimer’s disease. The BACE1 inhibitor binding affinities (IC50) reported in chEMBL for various chemical classes of human-beta secretase-1 inhibitors were computed using data-driven and deterministic approaches. Models for predicting binding affinities were developed using quantitative techniques, along with qualitative models for classifying inhibitors and non-inhibitors. The model was trained using 80 percent of the diverse data set, and the remaining 20 percent was used for external validation. In particular, machine learning models using various molecular descriptors and a deep learning model using molecular graph representation were developed. Based on the metrics of all approaches, the random forest model with 2d and 3d features outperforms other models with (R2 0.81) (RMSE 0.10). When compared to graph-based techniques such as GCN, GAN, and AttentiveFP, the success of the machine-learning approach in predicting the binding affinity of hBACE-1 inhibitors provides a strong thrust for systematically applying such methods for drug screening and developing successful BACE-1 inhibitors. The molecular docking-based virtual screening approach was unsuccessful in ranking the BACE-1 inhibitors, suggesting that such methods cannot be reliably used for identifying the lead compounds from different chemical libraries.