Accurate prediction of drug side effects remains a significant challenge in pharmaceutical development, as drug programs often fail due to unforeseen adverse reactions. Traditional preclinical approaches, such as animal testing, face limitations, including high costs, ethical challenges, and limited translatability to human biology. To address these limitations, we enhanced side effect prediction using computational models, specifically protein-protein interaction networks. We utilized the PathFX algorithm, a protein-protein interaction network tool, to predict adverse drug effects, focusing on improving prediction performance by addressing issues of overprediction and underprediction. Two main strategies were employed in this study: First, we refined pathway phenotypes by integrating key network genes and omics data, enhancing the biological context for side effect prediction. Second, we incorporated new drug-binding targets from multiple databases to broaden the view of potential drug-target interactions. Using a drug toxicity dataset, we assessed PathFX’s ability to predict side effects by generating drug networks and evaluating their pathway phenotypes. The baseline performance was low and variable across drugs and side effects. However, by refining pathway phenotypes, we reduced overprediction and false positive results, while observing a trade-off between specificity and sensitivity. Specifically, incorporating omics data increased sensitivity but led to reduced specificity, highlighting the balance between limiting false positives and increasing true positive predictions. When compared to animal studies, our computational predictions demonstrated similar performance metrics, suggesting that protein-protein interaction network models, despite limitations, hold value in drug safety evaluation. Furthermore, integrating drug-target interactions from diverse sources enabled the prediction of previously unrecognized side effects, effectively addressing underprediction. We also observed a trade-off between specificity and sensitivity in this approach. Notably, incorporating multiple sources of drug-target interactions proved more impactful in improving side effect predictions than refining pathway phenotypes alone. Our results suggest that databases with broad target coverage are particularly advantageous during the early stages of drug development, providing a wide array of potential interactions. Conversely, databases with a higher percentage of predictive targets are more valuable in later stages, offering refined and specific predictions to enhance phenotype-target associations. Taken together, by tuning pathway definitions and drug target inputs, this study suggests a pathway towards improved prediction performance and rational application of protein-protein interaction models to anticipate drug-induced side effects.
This research also emphasizes shifting focus from conventional performance metrics, such as sensitivity and specificity, to model utility, the practical impact of predictions on clinical decision-making. This approach has broader applications. For instance, we found that the utility of a model, in deep learning-based brain vessel segmentation, often supersedes traditional performance metrics like the Dice score, which measures overlap between predicted and actual structures but may not fully capture clinical relevance. By prioritizing the model’s practical application in accurately identifying anatomically significant structures, we demonstrated how focusing on utility can lead to more meaningful outcomes in both drug safety predictions and medical imaging.
