International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Machine Learning-Driven Prediction of Anti-Alzheimer Drug Efficacy Using PubChem Molecular Fingerprints

Authors :

Mohammad Javad Sadeghi¹ Mohammad Javad Nemati² AliAsghar Zare³ Mohammadreza Shams⁴

1- Department of Computer Engineering, Shahreza Campus, University of Isfahan, Iran 2- Department of Computer Engineering, Shahreza Campus, University of Isfahan, Iran 3- Department of Computer Engineering, Shahreza Campus, University of Isfahan, Iran 4- Department of Computer Engineering, Shahreza Campus, University of Isfahan, Iran

Keywords :

Alzheimer’s Disease،Drug Structure Analysis،Machine Learning Models،XGBoost Regression،Artificial Neural Networks،Canonical SMILES Representation،Molecular Fingerprint Analysis

Abstract :

Alzheimer’s disease is a chronic and progressive neurodegenerative condition that predominantly affects the elderly, manifesting through memory impairment, cognitive dysfunction, and behavioral abnormalities. Despite decades of research, a definitive cure remains elusive, and current therapeutic options only offer limited symptomatic relief. Traditional drug discovery approaches for Alzheimer’s are often time-consuming, resource-intensive, and associated with high attrition rates during clinical trials. In this study, we propose a robust machine learning (ML) framework that applies a logarithmic transformation to IC₅₀ (half-maximal inhibitory concentration) values, aiming to improve prediction accuracy and model stability. Molecular structures of anti-Alzheimer compounds are retrieved in Canonical SMILES format from the PubChem and NCBI databases. Drug efficacy is evaluated using both raw and log-transformed IC₅₀ values, with the latter demonstrating superior modeling characteristics by normalizing the dynamic range of pharmacological activity. Molecular feature extraction is performed using the PaDEL-Descriptor tool, yielding 883 descriptors. Through rigorous correlation analysis and the Interquartile Range (IQR) method for outlier removal, we identify seven PubChem fingerprints with the strongest association to therapeutic efficacy. Three predictive models—Support Vector Machine (SVM), XGBoost Regression, and Artificial Neural Networks (ANN)—are systematically evaluated. The XGBoost model achieves remarkable performance with a Mean Squared Error (MSE) of 0.26 for log-IC₅₀ predictions. This optimized pipeline not only demonstrates the effectiveness of AI-driven drug discovery but also introduces a novel data transformation protocol that better reflects the logarithmic nature of dose-response relationships in pharmacology.