Graph based link prediction for epilepsy drug discovery

Epilepsy is one of the most prevalent neurological disorders, affecting approximately 23 million people in Asia alone. It is a disorder with severe social impacts and is going to progressively damage the brain. It encompasses a wide range of syndromes and each one of them differs significantly in treatment options. Seizures are the common symptom in all of them. Despite being one of the most researched clinical conditions, the exact mechanism is still unknown, and this poses challenges for coming up with an effective treatment mechanism. Understanding phytochemical-protein interactions, inspired by Ayurveda, offers a natural alternative for treating epilepsy, especially where conventional drugs face resistance and side effects. Traditional lab methods are costly and time-intensive, making graph-based computational approaches a powerful and scalable solution. Inspired by Ayurveda, we propose a computational framework to predict phytochemical-protein interactions for potential epilepsy treatment. We propose that the interaction can be modeled as a bipartite graph, where nodes represent phytochemicals and proteins and edges are represented by interactions. We employ Graph Neural Networks to capture both local and global information about the graph. Initially, the entire graph was trained using Graph Convolutional Networks (GCN), Graph Attention Network (GAT), and GraphSAGE. To enhance predictive performance, we then constructed one-hop enclosing subgraphs for both positive and negative samples and trained the same three models on this refined dataset. Our best-performing model achieved an accuracy of 0.9778, precision of 0.9574, F1-score of 0.9782, and ROC-AUC score of 0.9994.