NEURAL NETWORK-BASED ADVERSE DRUG REACTION PREDICTION USING MOLECULAR SUBSTRUCTURE ANALYSES

Abstract

Objective: This study aims to enhance early detection and prediction by exploiting drug molecular substructures, overcoming challenges posed by limited authentic patient data in the medical domain. Methods: The study implemented a neural network approach to optimize molecular fingerprint algorithms and employed various machine learning algorithms for predictions. Additionally, the study identified and extracted substructures associated with severe Adverse Drug Reactions (ADRs), validating their presence within drug structures through a comparison with a random set of drug structures. Predictions were made for specific molecular structures, and results were validated using clinical evidence from the literature. Results: Optimized molecular fingerprint algorithms and diverse machine-learning models yielded promising outcomes. The Area Under Curve (AUC) value for the fingerprint dataset was obtained at approximately 65%, and integrating it with patient data significantly improved the performance by about 30%. Substructure analysis pinpointed key components linked to severe ADRs, reinforcing the predictive prowess of the model. Predictions for specific molecular structures were corroborated using clinical evidence from the literature, fortifying the credibility of the proposed approach. Conclusion: In conclusion, this research effectively tackles challenges in the early detection and prediction of ADRs by leveraging machine learning algorithms, focusing on drug molecular substructures. The optimized model, incorporating both fingerprint and patient datasets, demonstrated significant improvements in predictive performance. Identifying and validating substructures linked to severe ADRs contribute to the model's reliability. The study's findings are vital for advancing drug safety and laying the groundwork for further strides in predictive modeling within the medical domain.