A framework that combines two cutting-edge artificial intelligence technologies is offering a fresh perspective on how pre-existing drugs can be used for other clinical indications.
The path from the discovery of a new drug to its time to market is arduous, taking 12 years on average and costing around three billion U.S. dollars with only a small chance of success.
Advanced machine-learning methods offer the opportunity to accelerate and streamline this, combining vast biological and clinical datasets to predict drug interactions, potential adverse effects, and drug repurposing.
Researchers debuted one such system, called the Large Optimized Vector Embeddings Network—or LOVENet—at the thirty-seventh Conference on Neural Information Processing Systems.
LOVENet integrates large-language models, such as those that power ChatGPT, with structured knowledge graph technology to fuse text with data sources and mathematically unravel the complex relationships between drugs and diseases.
In this way, new clinical indications can be discovered for drugs that were initially developed to treat other diseases.
“The usual path for developing new medicines can take more than a decade,” said Jouhyun Jeon, PhD, the lead scientist and principal investigator at Klick Applied Sciences.
“By using AI to speed up the repurposing process, we hope to shave years off current timelines, identify more uses for existing drugs, and ultimately provide physicians and patients with more treatment options across a wide range of therapeutic areas.”
LOVENet fuses information from pairs of embedding from Llama2—Meta’s open-source large language model—and heterogeneous knowledge graphs to derive complex relations of drugs and diseases.
It leverages the potential of large-language models, which can transform and analyze text data and have shown potential in revolutionizing research in areas such as ligand/drug design and DNA/protein sequencing.
LOVENet harnesses the power of these models to extract comprehensive representations of drugs and diseases, which are integrated with a knowledge graph.
The researchers trained LOVENet models using three established benchmark datasets commonly used for drug repurposing studies and found that it consistently outperformed seven other state-of-the-art methods in cross-validation.
It was further able to highlight drug associations with new indications already flagged in scientific literature. This included quinidine, an anti-arrhythmic agent currently used to treat heart rhythm disturbances, that LOVENet showed could also show benefit with seizures.
Buspirone, which is used for treating anxiety disorders, was revealed to have a novel association with substance withdrawal syndrome that had previously been highlighted in the literature.
Similarly, LOVENet showed that the selective serotonin reuptake inhibitor fluoxetine, which is approved for major depressive disorder, anxiety, and panic disorder, could have benefits for attention deficit disorder with hyperactivity.
“LOVENet is an important first step in a new era of drug discovery,” said Alfred Whitehead Klick’s EVP of data science.
“We think it holds amazing promise to lower development costs, while increasing time efficiency and risk mitigation. It could also greatly assist in streamlining regulatory pathways, expanding market opportunities, while addressing unmet medical needs.”
