AI has already made waves in areas like image recognition and translation, but researchers are now turning their attention to discovery-driven tasks, particularly how various drugs interact with cancer cells. One of the most promising applications being explored is the generation of hypotheses, a process once believed to be purely a human endeavor.
A recent study from the University of Cambridge, in collaboration with King’s College London and Arctoris Ltd, put this concept to the test. The researchers wondered if AI could propose treatments for breast cancer using drugs not initially designed for that purpose. Could these AI-generated suggestions lead to actual, viable breakthroughs? The results indicate a hopeful possibility.
The research utilized GPT-4, a large language model (LLM) trained on extensive internet text. The team crafted prompts for GPT-4 to generate drug pairs that could effectively target MCF7 breast cancer cells while sparing healthy MCF10A cells. They specifically restricted the model from suggesting known cancer medications, focusing instead on affordable and already approved drugs.
Dr. Hector Zenil from King’s College London noted, “This is not about AI taking over scientists’ jobs but about forming a new kind of partnership.” He mentioned that, with expert guidance and experimental feedback, the AI essentially acted as a relentless research assistant, rapidly sifting through a vast hypothesis space and offering ideas that would take humans significantly longer to conceive.
In its initial attempt, GPT-4 generated 12 unique drug combinations, all of which featured medications not typically associated with cancer treatment, including those used for high cholesterol, parasitic infections, and alcohol dependence. What’s intriguing is that these suggestions weren’t random; GPT-4 provided logical explanations for each combination, often linking biological pathways in novel ways.
Moving forward, scientists tested the proposed drug pairs in the lab, measuring their effectiveness against MCF7 cells and any potential damage to MCF10A cells. The researchers also assessed whether the drug pairs demonstrated increased effectiveness when combined, a phenomenon known as synergy.
Three combinations outperformed standard cancer therapies: one composed of simvastatin and disulfiram, another with dipyridamole and mebendazole, and the last involving itraconazole and atenolol. These pairs showed efficacy against MCF7 cells without significantly harming healthy ones.
Professor Ross King from Cambridge remarked, “Supervised LLMs like GPT-4 open up scalable and creative avenues for scientific exploration, helping us discover paths we hadn’t considered before.”
After reviewing the initial findings, the researchers prompted GPT-4 to analyze the successful pairs and suggest new combinations. With updated information from lab results, the AI produced four additional pairs, one of which included known cancer drugs such as fulvestrant. Among these, disulfiram combined with quinacrine and mebendazole also mixed with quinacrine showed considerable synergy, with the disulfiram-simvastatin pairing achieving the highest synergy score in the entire study.
This feedback loop—where AI generates hypotheses, humans test them, and results are fed back—marks a significant evolution in the scientific process, enriching collaboration between machines and humans.
Interestingly, of the twelve original combinations proposed by GPT-4, six demonstrated positive synergy against MCF7 cancer cells, including some unexpected pairings. Notably, eight of these combinations had a more significant impact on the cancer cells than on the healthy ones, indicating a level of specificity that is desirable in treatment.
The most toxic drugs for MCF7 cells identified included disulfiram, quinacrine, niclosamide, and dipyridamole, with disulfiram requiring only a small dose to lessen cell viability. Researchers were indeed surprised by GPT-4’s ability to identify effective non-cancer drugs and pair them in a meaningful way.
“This investigation illustrates how AI can integrate into the continuous cycle of scientific discovery, providing adaptable, data-driven hypotheses in real time,” said Zenil.
There’s a caveat, though. GPT-4 doesn’t always get it right; these inaccuracies, known as hallucinations, can sometimes lead to unexpected revelations. For instance, it incorrectly suggested that itraconazole affects cell membranes in human cells, a statement that, while true for fungi, doesn’t apply to human biology. Even so, this flawed hypothesis led to promising experiments.
Professor King highlighted that the ability of supervised LLMs to propose relevant hypotheses across various fields marks a new era in scientific research.
The research team is optimistic that integrating AI with lab automation could ultimately lower the cost of personalized medicine. Imagine if cancer treatments could be customized for each patient, with therapies tested and tailored almost in real-time rather than relying on general prescriptions.
While the cost associated with laboratory operations remains high, AI can considerably decrease the time and resources needed for generating actionable hypotheses. With advancements in robotics, the physical testing process may also become more affordable.
The study concludes that GPT-4 was successful in generating new and useful hypotheses for drug combinations.
What does this all mean for the future of science? This research indicates that AI can play a role far beyond summarizing data; it can actively participate in producing new scientific insights. Although the combinations suggested still require clinical trials and aren’t close to being FDA-approved treatments, their promising outcomes in lab settings illustrate the potential of repurposing existing, safe drugs for other applications, potentially reducing years in development time.
