Cambridge scientists use ChatGPT to find “cheap and safe” cancer drugs

The University of Cambridge researchers used OpenAI’s GPT-4 large language model to identify combinations of inexpensive and safe drugs that could be effective at treating cancer. The results are promising.

Prompted by researchers, the chatbot found that combinations of drugs typically used to treat conditions such as high cholesterol and alcohol dependence could also be effective at treating cancer.

The research team fed GPT-4 piles of scientific literature and asked it to identify hidden patterns that could potentially lead to new cancer drugs, a “promising” new approach to drug discovery, according to the University of Cambridge.

Specifically, the researchers asked the model to find combinations that might work well against a common type of breast cancer used in lab studies. They instructed it to avoid standard treatments and harmful options, and prioritize affordable, approved drugs.

Human scientists then tested the combinations suggested by GPT-4. Three of the 12 in the first lab-based test worked better than current breast cancer drugs. Upon learning the results, the model suggested four more combinations, three of which showed promising results, according to the University of Cambridge.

The results, published in the Journal of the Royal Society Interface, mark the first time experimental data guided an AI model, which in turn directed further experiments.

“New frontier” in scientific research

While AI models are known for “hallucinations” – false or made-up statements – and are typically viewed as flaws, researchers in this case found them useful, as they can lead to new ideas worth testing.

Supervised AI offers “a scalable, imaginative layer of scientific exploration” and “can help us as human scientists explore new paths that we hadn’t thought of before,” said Cambridge Professor Ross King, who led the study.

Dr Hector Zenil from King’s College London, who co-authored the study, sees it as “a new kind of collaboration” rather than as a case of AI replacing scientists.

“Guided by expert prompts and experimental feedback, the AI functioned like a tireless research partner – rapidly navigating an immense hypothesis space and proposing ideas that would take humans alone far longer to reach,” he said.

Among the six combinations identified by GPT-4, simvastatin and disulfiram stood out for their effectiveness against breast cancer cells. The drugs are normally used to lower cholesterol and treat alcohol dependence, respectively.

To ensure they can be used in cancer treatment, the drugs will first have to go through extensive clinical trials. According to Professor King, this marks “a new frontier” in scientific research and shows that AI can be “a collaborator” in the scientific process.