AI Finds Novel Cancer Treatments Using Everyday Drugs

Artificial intelligence (AI) has already revolutionized areas such as image recognition and language translation. Now, researchers are delving into its capabilities for discovery-oriented tasks, particularly in understanding the complex interactions between various drugs and cancer cells.

A particularly thrilling application lies in hypothesis generation, a process previously considered exclusive to human intuition and curiosity.

A recent study spearheaded by researchers from the University of Cambridge, in collaboration with King’s College London and Arctoris Ltd, put this concept to the test. The core questions were: Could AI propose effective treatments for breast cancer by utilizing drugs not initially designed for oncology? And could these AI-generated suggestions pave the way for tangible, verifiable breakthroughs in laboratory settings? The study's findings indicate a promising affirmative.

AI Uncovers Novel Drug Candidates for Cancer

The investigation centered on GPT-4, a sophisticated large language model (LLM) trained on extensive internet data. The research team crafted specific prompts instructing GPT-4 to identify drug combinations effective against MCF7 breast cancer cells while sparing healthy MCF10A cells. Furthermore, the model was directed to avoid known cancer drugs and to give preference to combinations that are cost-effective and already approved for use in humans.

Dr. Hector Zenil from King’s College London emphasized, “This is not automation replacing scientists, but a new kind of collaboration. 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.”

In its initial run, GPT-4 suggested 12 distinct drug combinations. Notably, all these pairings involved medications not conventionally used in cancer treatment. These included drugs for managing conditions such as high cholesterol, parasitic infections, and alcohol dependence. Despite their unconventional nature, these combinations were not random. GPT-4 supplied justifications for each proposed pair, often connecting biological pathways in surprising and innovative ways.

Promising Combinations Emerge from Lab Testing

The subsequent phase involved rigorous laboratory testing of the AI-suggested drug pairs. Scientists meticulously measured two key factors: the efficacy of each combination in targeting MCF7 cancer cells and the extent of any damage to healthy MCF10A cells. Additionally, they assessed whether the drug pairs exhibited synergistic effects, meaning they worked more effectively together than as individual agents.

Three combinations demonstrated particularly strong results, outperforming standard cancer therapies. One notable pair was simvastatin and disulfiram. Another effective combination was dipyridamole with mebendazole, and a third involved itraconazole and atenolol. These drug pairings proved potent against MCF7 cells while minimizing harm to healthy cells.

Professor Ross King from Cambridge’s Department of Chemical Engineering and Biotechnology, who led the research, stated, “Supervised LLMs offer a scalable, imaginative layer of scientific exploration, and can help us as human scientists explore new paths that we hadn’t thought of before.”

An Iterative Cycle of AI Suggestion and Human Validation

After the initial round of experiments, the researchers fed the results back to GPT-4, asking it to analyze the successful outcomes and propose new ideas. They provided summaries of the lab findings and prompted the AI to suggest four additional drug combinations, this time including some known cancer drugs like fulvestrant.

In this iteration, the AI proposed combinations like disulfiram with quinacrine and mebendazole with quinacrine. Three of these four new suggestions once again demonstrated encouraging synergy scores. One of the standout combinations, disulfiram with simvastatin, achieved the highest synergy score in the entire study, registering over 10 on the HSA scale.

This feedback loop—where AI suggests hypotheses, humans conduct tests, and the results are fed back to the AI—signifies a novel approach to scientific discovery. The process is no longer linear; it becomes cyclical, with both machine and human intelligence adapting and refining their understanding with each iteration.

Surprising and Specific Cancer-Fighting Combinations

Out of the initial twelve combinations, six exhibited positive synergy scores against MCF7 cancer cells. These included unexpected pairings such as furosemide and mebendazole, or disulfiram and hydroxychloroquine. Crucially, eight of these twelve combinations demonstrated a more significant impact on MCF7 cancer cells compared to healthy MCF10A cells, indicating a desirable level of specificity.

Several drugs proved particularly toxic to MCF7 cells, including disulfiram, quinacrine, niclosamide, and dipyridamole. Disulfiram was especially potent, showing the lowest IC50 value, which means only a small dose was needed to significantly reduce cancer cell viability. The researchers themselves were surprised by GPT-4's ability to identify such effective non-cancer drugs and combine them in meaningful, synergistic ways.

Zenil added, “This study demonstrates how AI can be woven directly into the iterative loop of scientific discovery, enabling adaptive, data-informed hypothesis generation and validation in real time.”

When AI 'Hallucinations' Spark Innovation

Like any LLM, GPT-4 can occasionally produce 'hallucinations'—statements not directly supported by its training data. While often considered errors, in the context of hypothesis generation, these inaccuracies can sometimes lead to productive, unconventional ideas. An example from this study involved a 'hallucination' where the AI claimed that itraconazole affects cell membrane integrity in human cells. This assertion is true for fungal cells, but human cells utilize a different biological pathway. Nevertheless, this technically flawed premise spurred experiments that ultimately yielded successful results.

“The capacity of supervised LLMs to propose hypotheses across disciplines, incorporate prior results, and collaborate across iterations marks a new frontier in scientific research,” Professor King remarked.

Paving the Way for Personalized Cancer Therapies

The research team posits that the synergy between AI and laboratory automation holds the potential to significantly decrease the costs associated with personalized medicine. In the future, cancer treatment could evolve into a bespoke research endeavor for each individual patient. Rather than relying on generalized prescriptions, therapies could be rapidly tested and customized in almost real time. While laboratory operational costs remain substantial, AI tools like GPT-4 can dramatically reduce the time and resources needed to generate viable hypotheses. Coupled with advancements in robotics, the physical testing phase could also become more economical.

The study's authors concluded, “Our empirical results demonstrate that the GPT-4 succeeded in its primary task of forming novel and useful hypotheses.”

The Future of AI in Scientific Discovery

This groundbreaking study illustrates that AI's role extends beyond mere summarization or analysis; it can actively participate in the generation of new scientific knowledge. GPT-4 did not simply process data. It proposed unconventional ideas, learned from experimental outcomes, and refined its subsequent suggestions, showcasing a dynamic learning capability.

It is important to note that the drug combinations proposed by AI still require extensive clinical trials before they can be considered approved treatments. However, their success in laboratory settings underscores the immense potential of repurposing existing, safe drugs for novel applications, a strategy that could shave years off traditional drug development timelines.

The findings of this research are published in the Journal of The Royal Society Interface and on arXiv.

---

If you enjoyed this article, consider subscribing to our newsletter for more engaging content, exclusive insights, and the latest updates.

Also, check out EarthSnap, a free application from Eric Ralls and Earth.com.