Selection of optimal treatment
The artificial intelligence system selects the most appropriate cancer therapy
Maria Tolmacheva, XX2 century
Only 4% of all cancer drugs under development receive final regulatory approval.
"This is because right now we can't find the right combination of drugs for patients in a reasonable way. And this is especially true for cancer, where we cannot always predict which drugs will work better, given the unique and complex work of tumor cells in humans," explains Trey Ideker, professor at the University of California San Diego School of Medicine and Cancer Center Mursov (Moores Cancer Center).
In an article published in October in Cancer Cell (Kuenzi et al., Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells), the authors describe a new artificial intelligence (AI) system DrugCell predicting possible reactions of cancer cells to treatment based on detailed tumor data.
Drawing from the UC San Diego DrugCell press release: New Experimental AI Platform Matches Tumor to Best Drug Combo – VM.
"Most AI systems," says Ideker, "are black boxes–they can be very predictable, but we don't really know that much about how they work."
He cited the example of searching for images on the Internet for the query "cat": search AI systems train on existing images of cats, but why they actually define a new image for themselves as a "cat" and not as a "rat" or something else is unknown. In order for AI to be useful in healthcare, Ideker believes, we need to understand how the system comes to its conclusions, why a particular decision is made.
The team's work on DrugCell started with yeast. In a previous study, scientists created the DCell AI system, which they called a "visible" neural network. It was an AI simulator of a yeast cell. By training this system on several million genomes, the authors used detailed known information about yeast genes and mutations, and encoded the parameters of about 2500 cellular components. Then the researchers entered the parameters of a specific cell into the trained DCell, and the system predicted its behavior, in particular, growth. The observation showed that the behavior of the simulated DCell cell generally corresponded to the behavior of a real cell. At the same time, the work of the AI system was transparent, understandable to researchers, since the behavior of the cell was mainly determined by known parameters.
DrugCell – next version DCell also works on a similar principle. This AI model has been trained on more than 1,200 types of tumor cells and their responses to nearly 700 FDA–approved and experimental drugs - a total of more than half a million combinations. The team can provide DrugCell with tumor data, and the system in response gives out the most well-known drug used to treat similar types of cancer, demonstrates the biological mechanisms regulating the reaction to it, and predicts the results of combinations with other drugs.
A fairly accurate selection of cancer treatment is already possible at the Mursov Cancer Center. A tumor biopsy of a patient of the center can be checked for potential mutations and evaluated by an interdisciplinary board of experts – Molecular Tumor Board. The Board then recommends individual therapy based on the patient's unique genomic changes and other information. In a sense, DrugCell imitates such expert advice, only instead of doctors and scientists, AI selects the treatment.
The team's ultimate goal is to introduce DrugCell into clinical practice for the benefit of patients, but the authors of the study warn that much more work and clinical research needs to be done before this AI model can be widely used in medicine.
Portal "Eternal youth" http://vechnayamolodost.ru