PregAN-NET: Addressing Class Imbalance with GANs in Interpretable Computational Framework for Predicting Safety Profile of Drugs Considering Adverse Reactions During Pregnancy

Abstract

Adverse Drug Reactions (ADRs) during pregnancy pose significant risks to both the mother and the fetus. Conventional approaches to predict ADR are inadequate due to ethical restrictions that prevent performing medication studies in pregnant women, leading to restricted data samples. Hence, computational techniques have been promising for ADR predictions. However, most of these techniques have focused on the general population and face the challenge of class imbalance and lack of model interpretability. In the present work, an ensemble learning-based PregAN-NET framework has been proposed that addresses the issue of class imbalance by generating synthetic data employing Conditional Tabular Generative Adversarial Network (CTGAN) and integrates neural network and gradient boosting as a Boosted Neural Ensemble (BNE) architecture to predict safe and unsafe drugs considering their adverse reactions during pregnancy. Furthermore, the SHAP method has been employed to enhance the post-hoc interpretability of the BNE architecture by analyzing the contribution of different features towards prediction. The proposed framework has been applied to chemical and biological properties from PubChem and DrugBank, along with class labels from the ADReCS database. CTGAN has been evaluated for data balancing, showing a 2% to 5% performance improvement over SMOTE. The BNE architecture has outperformed six state-of-the-art methods by achieving mean ROC-AUC scores between 77.00% and 90.00% for chemical data, 66.00% and 74.00% for biological data, and 70.00% to 75.00% for combined datasets. Further, the top 20 contributory features in prediction corresponding to the different drug properties have been identified.