An AI model that intuits how cells will respond to drugs could lead to more targeted treatments for triple-negative breast cancer.
An AI-based model that anticipates how cells will respond to drugs will help find more effective, targeted therapies against triple-negative breast cancer, an aggressive form of breast cancer that accounts for about 15% of all new breast cancer diagnoses.
The discovery described above Nature it is of interest because the cells of this type of tumor lack the three main molecular targets (hormonal or protein) for which we have targeted treatments. For this reason, this form of cancer is more likely to develop resistance to therapies.
Training on tumor cells
A group of scientists at Westlake University in Hangzhou, China, worked to develop a virtual model that simulates how breast cancer cells evolve over time. The novelty compared to the past lies precisely in the attempt to capture an ongoing evolution and not just a static situation of malignant cells.
Scientists trained an AI model on data from 18 breast cancer cell lines (16 of which were triple-negative breast cancer). They treated the cells with 63 FDA-approved cancer drugs and evaluated how more than 5,500 groups of proteins expressed by the tumor cells responded to the various therapies 6, 24 or 48 hours after the drugs were administered.
Predictions on unknown drugs
After this training, the model learned to predict the effects of drugs on triple-negative breast cancer cells and understand which proteins contributed to drug resistance. Its skills have extended to unknown drugs: the AI has achieved 88% accuracy in predicting the cellular response to 81 drugs not included in the “training”.
When scientists tested it on tumor biopsies from 501 patients who were about to start chemo, the model correctly predicted the clinical outcome of the drugs they would take.
From theory to practice
At this point it was decided to test the capabilities of AI in a clinically relevant scenario, to find the most promising drug combination to treat three patients with triple-negative breast cancer, based on samples of their tumors taken and kept in the laboratory.
The AI searched among 3,000 drugs already approved or in advanced clinical trials, and correctly defined as most appropriate three drugs that had actually been given to patients, and which had proven effective. It also suggested three compounds that were more effective than standard therapy, which scientists then tested on cell samples and which patients then received.
The conditions are there, but it takes time
The model is still a prototype and, as such, will need to be improved, for example by studying – in perspective – whether the recommendations it provides are valid in the long term, and by also including immunotherapy drugs in the training, which have not yet been considered in the study.
