Assistant Professor University of California Los Angeles, United States
Introduction:: Many drugs fail to reach the market. Preclinical prediction of drug-induced safety events, including in silico tests, is of the utmost importance, though network-based methods, which rely on accurate side effect (SE) pathways, suffer from low prediction accuracy, and tend to over-predict associations. We discovered that downstream proteins, in addition to drug targets, are relevant for predicting drug SEs. We benchmarked PathFX [1], a protein-interaction network model, and discovered moderate sensitivity and specificity. In the PathFX mode, sensitivity and specificity are highly dependent on pathway definition and our goal was to engineer pathways by including identified key genes and omics measurements from various resources and measure changes in specificity and sensitivity, using those engineered pathways.
Materials and Methods:: We used a drug toxicity dataset containing active ingredient-SE pairs extracted from drug labels [2]. We mapped active ingredients to DrugBank identifiers and found PathFX phenotypes relevant to labeled SEs. We developed methods to evaluate PathFX performance per SE and calculated multiple evaluation metrics. We also discovered key genes, including targets and downstream proteins. We defined new custom PathFX phenotype pathways using our new platform, called PathFX_Gen, by including network proteins associated with true positive (TP) drug networks to evaluate TP and false positive (FP) SE predictions. To generalize our approach to cases with insufficient TP example networks, we further generated novel SE pathways using an omics dataset [3] and tested their ability to reduce over-prediction.
Results, Conclusions, and Discussions:: Baseline prediction performance was low and varied per SE. We modeled 890 drugs and matched drug toxicity SEs. For the prediction of 32 SEs, sensitivity and specificity were 0.18 and 0.83. Hypertension had a sensitivity of 0.54 and a specificity of 0.63. However, despite the low performance, we found 224 and 143 genes associated with TP and FP pathways. Including key genes, we eliminated over-prediction for side effects with sufficient TP examples. Using the omics data, we discovered new pathways for Myocardial Infarction, that had no signal previously. Our results indicated that network models could be reliable sources to study drug effects. The discovery of TP hypertension pathways could emphasize the importance of finding core pathway genes. We demonstrated that pathway engineering using omics data can improve the utility of protein interaction network models for preclinical SE prediction. Future work will consider additional sources for improving pathways-based prediction of SEs.
Acknowledgements (Optional): :
References (Optional): : [1] Wilson, Jennifer L., et al. "PathFX provides mechanistic insights into drug efficacy and safety for regulatory review and therapeutic development." PLoS computational biology 14.12 (2018): e1006614. [2] Wilson, Jennifer L., Alessio Gravina, and Kevin Grimes. "From random to predictive: a context-specific interaction framework improves selection of drug protein–protein interactions for unknown drug pathways." Integrative Biology 14.1 (2022): 13-24. [3] Chen, Yen-Wei, et al. "PharmOmics: A species-and tissue-specific drug signature database and gene-network- based drug repositioning tool." Iscience 25.4 (2022): 104052.
