While AI has already transformed fields such as recognition and translation of images, researchers now explore its potential in discovery tasks, exploring how different drugs interact with cancer cells.
One of the most exciting applications is the generation of hypotheses, which was once considered as the domain of human curiosity alone.
A recent study, led by researchers from University of Cambridge in partnership King’s College London And Arctoris Ltd tested this idea.
Could he suggest treatments For breast cancer, the use of drugs is not originally to combat cancer? Could these suggestions lead to real breakthroughs testable in the laboratory? The results suggest that the answer could be yes.
AI finds new ideas for cancer medication
Research has been concentrated on GPT-4, a large language model (LLM) which is formed on large quantities of internet text. The team designed prompts that asked GPT-4 to generate pairs of drugs that could operate against the cancer cells in the MCF7 breast but do not harm healthy cells (MCF10A).
They also prevented the model from using known cancer drugs and asked him to prioritize affordable and already approved options for use humans.
“It is not automation that replaces scientists, but a new type of collaboration,” said Dr. Hector Zenil of King’s College London.
“Guided by experts from experts and experimental comments, AI worked as a tireless research partner, quickly sailing in a huge hypothesis space and offering ideas that would take humans alone to reach.”
In his first round, GPT-4 offered 12 combinations of unique drugs. Interestingly, all combinations included drugs not traditionally associated with cancer therapy.
These included drugs for conditions such as high cholesterol, parasitic infections and alcohol dependence.
Despite this, these combinations were not arbitrary. The GPT-4 has provided justifications for each pairing, often linking biological paths unexpectedly.
Some combos work well
The next step was to test the pairs of drugs in the laboratory. Scientists have measured two things: how each combination attacked MCF7 cells and the amount of damage caused to MCF10A cells.
They also assessed whether the pairs of drugs worked better together than separately, a property known as the Synergy.
Three combinations have stood out to have better results than standard cancer therapies. One involved simvastatin and disulfiram.
Another dipyridamole associated the mebendazole. A third involved the itraconazole and atenolol. These pairs of drugs were not only effective against MCF7 cells, but they worked without harming healthy cells too much.
“The supervised LLMs offer an evolutionary and imaginative layer of scientific exploration and can help us while human scientists explore new paths that we had not thought before,” said Professor Ross King of the Cambridge Chemical and Biotechnology Department, who led research.
AI and humans improve ideas together
After the first series of results, the researchers asked GPT-4 to analyze what had worked and to suggest new ideas.
They shared summaries of the laboratory results and prompted the IA Suggest four other combinations of drugs, including some involving cancer drugs known as the fulvestrant.
This time, the AI has returned combinations such as disulfiram with the Quinacrine and the mebendazole with the Quinacrine. Three of the four new suggestions have again shown promising synergy scores.
One of the most effective combinations was disulfiram with simvastatin, which obtained the highest synergy score of the entire study at more than 10 on the HSA scale.
The feedback loop, IA suggesting ideas, humans testing them, then returning the results to AI, represents a new way of conducting science.
The process no longer moves in one direction. Instead, it cycle, with a human machine and adjustment and improve as they learn from each iteration.
AI has found surprising cancer combos
Among the twelve original combinations, six have shown positive synergy scores for MCF7 cancer cells. These included unusual agreements such as furosemide and mebendazole or disulfiram and hydroxychloroquine.
Above all, eight of these twelve combinations have had greater effects on MCF7 cells than on MCF10A cells, indicating a good specificity.
Some of the most toxic drugs of MCF7 cells included disulfiram, quinacrine, niclosamide and dipyridamole. Disulfiram had the lowest IC50 value, which means that it required only a small dose to reduce cellular viability.
GPT-4’s ability to find such effective non-cancer drugs and to associate them significantly, even researchers.
“This study shows how AI can be woven directly in the iterative loop of scientific discovery, allowing the generation and validation of adaptive and informed hypotheses of real -time data,” said Zenil.
Hallucinations as creative jumps
GPT-4 sometimes makes mistakes. These are called hallucinations, declarations not supported by its training data. In most cases, hallucinations are faults. But in the generation of hypotheses, they can be productive.
In this study, one of these hallucinations involved the assertion that itraconazole affects the integrity of the cell membrane in human cells.
Although this is true for fungi, human cells do not use the same biological route. However, this erroneous idea has always led to successful experiences.
“The capacity of supervised LLMs to offer hypotheses through disciplines, to integrate previous results and to collaborate through iterations marks a new border in scientific research,” said the king.
AI can help adapt cancer treatments
The research team thinks that IA And laboratory automation could possibly reduce the cost of personalized medicine.
The treatment of cancer could one day involve a personalized research project for each patient. Instead of general prescriptions, therapies could be tested and adapted in real time.
The cost of execution of a laboratory remains high, but AI as GPT-4 considerably reduces the time and the efforts necessary to generate useful hypotheses. With other progress in robotics, the physical test process could also become cheaper.
“Our empirical results show that the GPT-4 has succeeded in its main task of forming new and useful hypotheses,” concluded the authors.
What it means for future science
This study shows that AI can make more than summarize or analyze. He can participate in the generation of new scientific knowledge.
The GPT-4 has not only criticized the figures. He offered unexpected ideas, drawn from the results and improved his suggestions.
The combinations he proposed still required clinical trials. They are far from becoming approved treatments. But their success in laboratory circles highlights the potential to reuse safe and existing drugs for new uses, which could allow years of development time.
The study is published in the Journal of the Royal Society Interface And arxiv.
––
Do you like what you read? Subscribe to our newsletter To initiate articles, exclusive content and the latest updates.
Check us Eartha free application that has been brought to you by Eric Ralls and Earth.com.
––
