Prediction of adverse drug reactions using gene expression data and neural networks with synthetic data augmentation

Adverse drug reactions (ADRs) remain a major clinical and public health challenge, contributing significantly to global morbidity and mortality. Advances in pharmacogenomics and large‑scale transcriptomic profiling offer new opportunities for predicting ADR risk; however, accurate modeling remains difficult due to high‑dimensional gene‑expression data and pronounced class imbalance in ADR labels. In this study, we developed machine‑learning models that predict ADRs using gene‑expression signatures from the LINCS L1000 platform, comprising 978 landmark genes across diverse drug‑perturbed human cancer cell lines. We evaluated traditional classifiers and deep learning models, and we addressed class imbalance using Synthetic Data Vault (SDV)–based augmentation. The optimized neural‑network model with SDV augmentation achieved the strongest performance across 232 ADR categories, while an optimized XGBoost model yielded comparable performance with enhanced interpretability. SDV augmentation improved generalization and mitigated imbalance‑related degradation observed in baseline models. Together, these findings demonstrate that integrating neural‑network architectures with synthetic data augmentation supports robust prediction of ADR‑associated transcriptional signatures in cancer cell‑line models. Because these signatures derive exclusively from in vitro cancer cell lines, the framework should be regarded as a foundation for future, clinically validated personalized‑medicine applications in drug safety rather than a ready‑to‑deploy clinical tool.