Researchers at Harvard Medical School have developed PDGrapher, a tool that uses Artificial Intelligence to accelerate drug discovery and development by identifying gene combinations that can reverse disease states in cells. Published last month in Nature, the study reports that PDGrapher operates up to 25 times faster than existing approaches. The work is led by associate professor Marinka Zitnik, whose lab aimed to rethink early-stage discovery by starting from the desired healthy outcome rather than the effects of specific drugs.
Unlike traditional methods that predict how drugs will affect cells, PDGrapher uses machine learning to probe the genetic causes of disease and suggest which genes drugs should target to restore normal function. “Instead of asking ‘what happens if we apply this drug?’ we asked, ‘what drug or set of targets would restore the healthy state?’” Zitnik wrote. The system employs optimal intervention design to scan large gene datasets and pinpoint combinations of genome changes most likely to correct diseased cell behavior. By evaluating many genes working together rather than following a one drug, one target paradigm, PDGrapher aims to capture nuances that conventional screening misses.
The researchers say the approach could speed early-stage development by anticipating gene target combinations and guiding scientists toward untested therapeutic strategies. “PDGrapher can predict target combinations that scientists have not yet tested, pointing the way to entirely new therapeutic strategies,” Zitnik wrote. First author Guadalupe Gonzalez added that the technology’s breadth enables applications to rare or underresearched diseases, where data-driven insights could be especially valuable.
Despite its promise, the team acknowledged limitations. Research fellow and co-author Xiang Lin said the tool, like other Artificial Intelligence models, currently cannot draw on existing scientific knowledge to better infer relationships among genes in diseased cells. Gonzalez estimated PDGrapher could be applied to drug development within one to three years, though any new medicines it helps uncover are unlikely to reach patients for at least a decade. Even so, Zitnik said the ability to link diseased states to potential interventions in a single step could reshape the drug discovery landscape over time.
