{shortcode-07836984b32515547aa5ddd983b79e3b255e4e7e}

Researchers at Harvard Medical School developed a tool that uses artificial intelligence to accelerate drug discovery and development.

The new technology, called PDGrapher, applies machine learning to identify a list of genes that can reverse disease in cells when targeted by drugs. The study, published last month in Nature, claims the technology works up to 25 times faster than existing methods.

Marinka Zitnik, an associate professor at the Medical School who heads the study’s lab, wrote that researchers were able to reverse the typical process to identify which genes drugs should target to make cells healthy.

Researchers usually use technology to predict the effects of drugs on cells. Zitnik’s lab pivoted to use AI to probe for genetic causes of disease, allowing drugs to be designed for specific genetic mutations rather than a single target.

“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.

Researchers used optimal intervention design, a machine learning device that finds the best practices to achieve a certain outcome. With the tool, researchers can take a large dataset of genes and find which combination of genome changes can restore normal function of a diseased cell and speed the process of designing a drug.

Zitnik explained that this allows PDGrapher to capture previously unobserved nuances by looking at “many genes working together” instead of the “one drug, one target” approach.

The new technology could have multiple implications for the future of disease research. Guadalupe Gonzalez, the first author of the study, said the development could be applied to the research of rare or underresearched diseases given the breadth of data that the technology can analyze.

PDGrapher could also accelerate early-stage drug development by anticipating gene combinations that drugs could target, according to Zitnik.

“PDGrapher can predict target combinations that scientists have not yet tested, pointing the way to entirely new therapeutic strategies,” Zitnik wrote.

While the study provides new insights on drug application, Xiang Lin, a research fellow at HMS and an author of the study, said that the tool still has a number of limitations. Lin said that PDGrapher, like other AI models, currently cannot draw on existing scientific knowledge to better analyze which genes could be related in understanding diseased cells.

PDGrapher can be applied to drug development in the next one to three years, but any new drugs it helps to discover will not be seen for at least 10 years, according to Gonzalez.

“There’s a long way to go,” Gonzalez said. “Big changes in drug discovery require time.”

Still, Zitnik wrote that the technology has the potential to build a new landscape for drug development.

Unlike other methods that examine one gene’s response to a drug at a time, PDGrapher’s ability to compare gene combinations could substantially accelerate the research processes — and discover new treatments.

“This makes it much more relevant for creating new therapies because it connects diseased states to possible interventions in one step,” Zitnik wrote.