Non-Linear Function Approximation for Predicting Binding Affinity of PPAR-Targeting Antidiabetic Compounds from Molecular Descriptors

Abstract

Diabetes mellitus remains a major global health challenge, necessitating the development of more effective therapeutic agents. The PPAR family plays a crucial role in regulating glucose and lipid metabolism, making it an important target for antidiabetic drug discovery. However, the identification of potent PPAR-targeting compounds is often limited by the high cost and time-consuming nature of experimental approaches. This study aims to develop a non-linear function approximation model to predict docking-derived binding affinity of antidiabetic compounds targeting PPAR using 2D molecular descriptors. A dataset of 3,764 small molecules with IC50 values was curated from the ChEMBL database, followed by data preprocessing to remove duplicates and incomplete entries. Molecular docking simulations were performed using AutoDock Vina to obtain binding affinity scores (kcal/mol), which were used as the target variable. Subsequently, 2D molecular descriptors were calculated from SMILES representations to capture key structural and physicochemical properties of the compounds. These descriptors were used as input features for a Multi-Layer Perceptron (MLP) regression model to approximate the complex non-linear relationship between molecular structure and binding affinity. The model achieved R² values of 0.853 for the training set and 0.632 for the test set, indicating moderate predictive performance and acceptable generalizability. Overall, this approach demonstrates the potential of machine learning as a cost-effective and scalable tool to support early-stage discovery of antidiabetic compounds targeting the PPAR family.