Old Drugs, New Uses: the Graph Neural Network Approach

Using PyG’s link prediction to find new drug indications

Photo by danilo.alvesd on Unsplash

The road to a new drug is long and costly. According to the National Institutes of Health, it takes a decade and over $1 billion to shepherd a potential drug from initial discovery to final approval. And more than 95% of drugs failed during this process (1). Even more concerning is Eroom’s Law (a play on Moore’s Law but spelled backward), which predicts a future where drug development becomes even more expensive and time-consuming, despite all the advancements in high-throughput screening, biotechnology, and computational tools (1, 2). Simply discovering new chemicals won’t be enough to reverse this trend. Repurposing existing drugs for new uses presents a promising alternative, and a specific type of artificial intelligence called graph neural network (GNN) can be a powerful tool for scientists in this endeavor.

Figure 1. Predicting new indication for existing drugs via graphs. In this figure, a graph neural network is used to investigate the repurposing of chloramphenicol against sinusitis. Different from traditional machine learning, graph neural network aggregates information from the node itself and from its neighbors. This may lead to better performances. Image by author.

GNN belongs to graph machine learning (GML), and GML is in turn a subfield of machine learning. As its name suggests, GML operates on graph data. Graphs use nodes to represent individual entities and connect these nodes with edges (Figure 1). Compared to…