// NATURE NEWS — SPAZIO & SCIENZA
AI model predicts which breast-cancer drugs work best
Search author on:
PubMed
Google Scholar
Choosing the most effective treatment for people with triple-negative breast cancer (TNBC) is often challenging because the tumours can behave very differently at the molecular level. Now, researchers have developed an artificial-intelligence-based ‘virtual cell model’ that can help in choosing the best treatment for this aggressive form of breast cancer — by predicting how an individual’s tumour cells respond to various drugs.
TNBC cells lack receptors to hormones oestrogen and progesterone, as well as to a protein that controls cell growth. This makes the condition harder to treat because hormone and targeted therapies cannot target the tumour cells.
In experiments using biopsies taken from individuals with TNBC, the new model showed promising results in identifying the same drugs that would prove to be effective when given to people. The authors say this raises the possibility of more personalized care for TNBC, which accounts for 15–20% of breast cancer cases.
“This is the first time that a virtual cell model goes out of the laboratory and is tested in a clinical scenario,” says study co-author Tiannan Guo, a proteomics specialist at Westlake University in Hangzhou, China. Guo adds that this AI model has a “very focused goal” in drug discovery for TNBC, rather than providing a comprehensive simulation of cellular behaviour.
Researchers have been developing computer-based simulations of cellular life for some time. However, most approaches have involved training deep neural networks on single-cell sequencing data, and tended to capture static cell states.
In their study, Guo and his colleagues aimed to develop a virtual cell model of breast cancer cells using proteomics data. They trained an AI model on more than 38 million protein measurements collected from 18 breast cancer cell lines, 16 of which were TNBC cells. The researchers treated the cells with 63 antitumour drugs that have been approved by the US Food and Drug Administration, and 59 drug combinations.
Can AI build a virtual cell? Scientists race to model life’s smallest unit
To create the training data set for the model, the team measured the levels of 5,585 protein groups in the tumour cells before treatment and after 6, 24 and 48 hours of drug therapy.
Systems biologist Hani Goodarzi, at the Arc Institute in Palo Alto, California, says the scale of the proteomics data used to train the model is unique. And this new approach provides researchers with “modalities that we haven't had before” for virtual cell models, he adds.