Breathing New Life into Old Drugs with Knowledge-Infused AI

Drug repositioning—finding new therapeutic uses for existing medicines—is a vital shortcut in the slow world of drug discovery. Recently, researchers have turned to Graph Neural Networks (GNNs), a type of AI adept at modelling relationships between complex data points, to predict drug-disease links. However, standard GNNs often begin their learning process with random data, effectively ignoring the vast biological knowledge already available.

To solve this, a new framework called DReKGNN has been developed to bridge the gap between raw data and expert insight. Rather than starting from a blank slate, the system uses Large Language Models (LLMs) to process expert information directly from trusted databases like DrugBank and OMIM. Importantly, it avoids the hallucinations sometimes associated with AI by sticking to verified descriptions of biological mechanisms. These insights are converted into 'node embeddings'—mathematical representations of data—which are then integrated using a mean aggregation strategy to reduce noise. By grounding the AI in established science rather than random initialisation, DReKGNN effectively predicts potential new roles for old drugs.