{shortcode-17f2052225a2fe61770acf0ae0a075e8eb2fe8f8}
Move over, Hugh Laurie: there’s a new Dr. House in town.
Harvard Medical School associate professor Marinka Zitnik and her lab announced the development of TxGNN, an artificial intelligence model which is able to suggest new treatment applications of existing drugs using neural networks.
The model, which Zitnik and her co-authors published in Nature Medicine last month, is capable of selecting from among nearly 8,000 drugs to treat more than 17,000 diseases — including diseases that are poorly-understood.
She likened the model to Gregory House — the fictional protagonist of the eponymous Fox medical drama — who had an uncanny ability to diagnose patients with unusual conditions.
TxGNN can “transfer knowledge from well-researched diseases to diseases with sparse data, enabling the identification of drug candidates without the need for extensive clinical trials or new drug development,” Zitnik wrote in an email. “This can reduce costs and time-to-market, providing a much-needed avenue for advancing treatments for neglected diseases.”
“Much like Dr. House in the TV series, these models approach medical challenges by combining different medical facts and clues that might not be obvious or directly related,” she added.
Unlike other AI models, whose internal workings are often unknown even to their developers, TxGNN includes an “explainer module” which allows users to see the medical rationale behind its recommendations.
The explainer module “can largely bridge the gap between research and practical applications,” wrote Chao Yan, a researcher at Vanderbilt University whose work focuses on similar topics.
Yan said that models like TxGNN can far surpass humans’ ability to process and analyze medical data to discover new treatments.
They “allow complex knowledge to be easily transferred to unknown domains with minimal model training and only a small amount of labeled data — an outcome that is nearly impossible with traditional methods,” Yan wrote.
Michael W. Mullowney, a scientist at the University of Chicago, wrote in an email that he was “extremely optimistic about the uses of AI for drug discovery and drug repurposing,” adding that TxGNN “would shine where expert human researchers’ intuition or ability fall short.”
Still, he wrote, he maintains “a level of caution that I imagine would be shared among the drug discovery and development community.”
Mullowney wrote that he was concerned about the possibility of AI models producing false positives.
“This danger could be compounded by doctors acting irresponsibly,” Mullowney wrote. “We don’t want AI doing the doctor’s homework.”
Zitnik wrote that as models like hers become more prominent, “regulatory frameworks will need to adapt to this new paradigm.”
“Current regulations may not fully account for AI-generated drug candidates, so regulatory bodies like the FDA and EMA will need to establish clear guidelines for using AI in drug development,” she wrote.
