Deep learning framework reveals drug combinations involving dipeptidyl peptidase-4 inhibitors for treatment of polycystic ovarian syndrome on a heterogeneous network

Polycystic Ovarian Syndrome (PCOS) is associated with infertility, cancer, diabetes and mental disorders. Still, there is no accepted solution for treatment. The crisis pertains to gastrointestinal issues and toxicity of existent Metformin-based drug combinations. Also, there is lack of synergistic and sensitive drug combinations for treatment and solutions for reducing gastrointestinal and mental disorders. Our model proposed here consists of three GCN-based link predictors to predict synergism and sensitivity, PCOS-relevance and feasibility of existent and derived drug combinations. A heterogeneous network consisting of drugs and proteins as vertices and drug-drug and drug-protein interactions as edges serve as input to our model. Derived drug-drug interactions are formed between Dipeptidyl peptidase-4 inhibitors/Gliptins and drugs derived through virtual screening strategies. Gliptins palliate gastrointestinal issues. Drugs involved in Metformin-based drug combinations, gastrointestinal and mental disorders and their PCOS targets form input to virtual screening. Gliptins-based drug combinations with goodness (mean of prediction scores) of 0.5 or more are outputted. These interactions are constituted of drugs having druglikeness and no toxicity, withdrawal and side-effects. These constituent drugs are further categorized into seven classes (depending on use) and compared using ADMET properties. Our model supersedes 13 baselines in accuracy (0.9889), F1-score (0.9890), specificity (0.9830), MCC (0.9780), AUROC (0.9889) and AUPRC (0.9807). It also validates 97.30% of test data and remains triumphant in ablation studies. Further, it can aid in drug repurposing in other diseases along with druggability evaluation of the associated proteins and comorbidity assessment.