The researchers said their revelations could personalise treatment, including picking out those patients likely to benefit from immunotherapy.
“We are at the cusp of a revolution in healthcare”
Their study applied AI and machine learning to gene sequences and molecular data from tumours, to reveal crucial differences among breast cancers that had previously been thought of as one type.
The study, led by a team at London’s Institute of Cancer Research, found two of the types were more likely to respond to immunotherapy than others, while one was more likely to relapse on tamoxifen.
The researchers are now developing tests for the cancer types that will be used to select patients for different drugs in trials, with the aim of making personalised therapy a standard part of treatment.
The same team previously used AI in the same way to uncover five different types of bowel cancer and oncologists are now evaluating their application in clinical trials.
Their aim is to apply the AI algorithm to many types of cancer and to provide information for each about their sensitivity to treatment, likely paths of evolution and how to combat drug resistance.
“We think we will be able to apply this technique across all cancers”
The new research, published last week in the journal NPJ Breast Cancer, has the potential to not only help select treatments for women with breast cancer but also identify new drug targets.
Those behind the work noted that majority of breast cancers develop in the inner cells that line the mammary ducts and are “fed” by the hormones oestrogen or progesterone.
These are classed as luminal A tumours and often have the best cure rates. But patients respond very differently to standard-of-care treatments, like tamoxifen, or new treatments – used if patients relapse – such as immunotherapy.
The study authors applied AI-trained software to available data on the genetics, molecular and cellular make-up of primary luminal A breast tumours, along with data on patient survival.
Once trained, the AI was able to identify five different types of disease with particular patterns of response to treatment.
Women with a cancer type labelled “inflammatory” had immune cells present in their tumours and high levels of a protein called PD-L1, suggesting they were likely to respond to immunotherapies.
“Our study has used AI algorithms to spot patterns within breast cancers that human analysis had up to now missed”
Another group were found to have “triple negative” tumours, which do not respond to standard hormone treatments, but various indicators suggesting they might also respond to immunotherapy.
Those with tumours that contained a specific change in chromosome 8 had worse survival than other groups when treated with tamoxifen and tended to relapse much earlier.
These patients may benefit from an additional or new treatment to delay or prevent late relapse, according to the researchers.
The markers identified in the study do not challenge the overall classification of breast cancer, but they do find additional differences within the current sub-divisions of the disease, said the authors.
Lead study author Dr Anguraj Sadanandam, team leader in systems and precision cancer medicine at the Institute of Cancer Research, said: “We are at the cusp of a revolution in healthcare.
“Our new study has shown that AI is able to recognise patterns in breast cancer that are beyond the limit of the human eye, and to point us to new avenues of treatment among those who have stopped responding to standard hormone therapies.
“AI has the capacity to be used much more widely, and we think we will be able to apply this technique across all cancers, even opening up new possibilities for treatment in cancers that are currently without successful options,” he said.
“The AI used in our study could also be used to discover new drugs for those most at risk of late relapse”
Dr Maggie Cheang, a team leader in the institute’s genomic analysis clinical trials team, said: “Doctors have used the current classification of breast cancers as a guide for treatment for years, but it is quite crude and patients who seemingly have the same type of the disease often respond very differently to drugs.
“Our study has used AI algorithms to spot patterns within breast cancers that human analysis had up to now missed – and found additional types of the disease that respond in very particular ways to treatment,” she said.
“Among the exciting implications of this research is its ability to pick out women who might respond well to immunotherapy, even when the broad classification of their cancer would suggest that these treatments wouldn’t work for them.”
She added: “The AI used in our study could also be used to discover new drugs for those most at risk of late relapse, beyond five years, which is common in oestrogen-linked breast cancers.”
As well as charity funding from the institute itself, the work was supported by the National Institute for Health Research’s Biomedical Research Centre and the Royal Marsden NHS Foundation Trust.