Key points from article :
Traditional drug discovery has long been based on the idea of targeting one protein with one drug. While this has led to major successes, such as kinase inhibitors and HER2 antibodies, it often falls short for complex diseases like cancer, where multiple pathways drive disease progression. Many recent breakthroughs instead focus on reprogramming entire cellular processes rather than a single molecular target.
In a new study published in Nature Biomedical Engineering, Marinka Zitnik of Harvard Medical School and colleagues introduce PDGrapher, an AI model that uses graph neural networks to map how gene networks behave in diseased versus healthy states. Unlike traditional methods that test how drugs change cells, PDGrapher asks the inverse question: which gene targets must be altered to shift a diseased cell back toward health? The model integrates protein–protein and gene regulatory network data to approximate causal links, allowing it to identify effective single targets and drug combinations.
The researchers tested PDGrapher on 38 datasets across 11 cancer types and found it consistently outperformed existing AI approaches. It not only rediscovered targets of known drugs, such as VEGFR2 in lung cancer, but also flagged new ones like TOP2A in lung adenocarcinoma, a finding supported by emerging clinical data. Importantly, the tool also revealed how drugs such as vorinostat and sorafenib rewire cellular networks, suggesting uses in both drug repurposing and mechanism discovery.
Looking ahead, the team aims to expand PDGrapher to neurological diseases such as Parkinson’s and Alzheimer’s. By offering a roadmap of how to reprogram cells, the approach could accelerate the development of precision treatments, potentially enabling tailored therapies for individual patients in the future.
