Poster L15 - Pathways Analysis of Predicted Diabetes Drugs Reveals Divergent Connections to Approved Diabetes Drug Pathways

Introduction: Protein-Protein Interaction (PPI) network methods, such as PathFX, are increasingly popular for predicting drug downstream effects. However, their use during drug-target selection is limited due to their tendency to overpredict drug effects, yet we lack the gold standard drug effects needed to validate these predictions. Previously, we discovered that classifying drugs by downstream proteins instead of traditional drug classes, such as ATC codes, combined with Electronic Health Record (EHR) analysis, could discover new drug effects. We hypothesized that PPI network classification coupled with large-scale observational studies could validate predicted drug effects and minimize overprediction. To test the feasibility of this approach, we first sought to understand patterns in diabetes-associated drug predictions as anti-diabetics are widely prescribed and metabolic dysregulation is implicated in multiple diseases. We hypothesized that true positives or true negatives could be inferred based on whether they shared connections with approved drugs or not.

Materials and

Methods: We modeled 40 approved antidiabetics in PathFX and separated PathFX predictions into two groups: diabetes-related pathway phenotypes (“approved diabetes pathway”) and additional non-diabetes phenotypes (“predicted non-diabetes pathway”). We also modeled 2,492 approved and experimental, non-antidiabetic drugs using PathFX and selected predictions connected to diabetes pathway phenotypes (“predicted diabetes pathways”). We then quantified the shared genes between the “predicted non-diabetes/diabetes” and “approved diabetes” pathways and used GO enrichment to understand the biological processes associated with pathway genes.

Results, Conclusions, and Discussions: We observed that the approved drugs converged on a shared module of 234 disease-associated genes whereas predicted diabetes drugs either shared approved drug network genes or used distinct pathways. Predicted diabetes pathways split into two groups that either shared or had distinct pathways from approved drugs, suggesting that overprediction could be reduced by eliminating drug predictions that deviate from known pathways through the network (Fig. 1a). Clustering of approved and predicted diabetes drugs revealed distinct groupings based on their diabetes pathways (Fig. 1b). 7,214 predicted pathways connected approved anti-diabetics to non-diabetes phenotypes of which 6,847 pathways overlapped with the diabetes module with a mean of 72.0% overlap. 2,492 non-diabetes drugs were predicted to affect diabetes with a total of 10,205 pathways. 6,801 of these predicted diabetes pathways overlapped with the diabetes module with a mean pathway overlap of 85.0%. Under current classifications, the top three Level 1 ATC classes for predicted drugs were “Nervous system”, “Cardiovascular system”, and “Respiratory system” while the majority of approved anti-diabetics were under “Alimentary tract and metabolism.” GO enrichment similarly found the diabetes disease module to be enriched for metabolism-, cardiovascular-, immune-, and nervous-related processes.

Pathways analysis using PathFX predicted multiple effects for approved anti-diabetic drugs. Our analysis found that “over-predictions” intersected with approved, known-effect pathways suggesting that a subset of the predicted and approved drugs are connected to diabetes genes in a similar manner. Shared pathways hint that drugs might share more effects than their intend-to-treat diseases. Additionally, our results suggest that overprediction of predicted diabetes pathways may be reduced by eliminating paths that deviate from known, approved pathways. Predictions made by PathFX connecting diabetes to other disease areas, such as oncology and brain diseases, can be seen in the literature, providing further motivation to investigate these associations. Future directions include conducting observational studies in the EHR to determine if these associations are reflected in clinical data.

Acknowledgements (Optional):