What’s driving this tumor
Aiming to stymie breast cancer through gene testing and AI
Melissa Pickett didn’t expect her breast cancer diagnosis. As a new mom, the 33-year-old college professor had a lot on her mind, including getting to her six-month postpartum checkup in November 2022.
When her doctor found lumps in her breast, Pickett dismissed them as a side effect of breastfeeding. She wasn’t worried. She had no family history of breast cancer, and she was used to false alarms. “I have a history of benign breast lumps going back to when I was 16,” Pickett said. “So I’ve had a lot of scans and imaging over the years. It was always nothing.”
But this time was different.
When her oncologist at Stanford Medicine, associate professor of medicine Melinda Telli, MD, told Pickett about a newly launched clinical trial for people recently diagnosed with her type of breast cancer, Pickett, who has a PhD in toxicology and is an associate professor of genetics at San Jose State University, didn’t hesitate to enroll.
“I’ve chosen to dedicate my life to the study of basic biology,” Pickett said. “Usually I’m the researcher, rather than the subject. But if my cells can be of use, if my case can increase our understanding of the biological processes underlying this disease, I really wanted to participate.”
Pickett is no stranger to Stanford; prior to her appointment as a faculty member at San Jose State she was a postdoctoral scholar in Stanford University’s biology department studying how epithelial cells (cells that line the inside of organs or glands and that make up the outer surface of the body) orient themselves in three dimensions — designating a region of a cell the top and another as the bottom, each with different functions. Ironically, loss of this orientation, or polarity, is a key hallmark of cancer cells. A few rogue epithelial cells in Pickett’s breast had turned against her.
The trial, which is called Terpsichore after the Greek goddess of lyric poetry and dancing, is a novel effort to gather genetic data from a patient’s tumor immediately after diagnosis and before treatment has started — and to use that data to attack the biological signals predicted to drive the tumor’s growth in its infancy.
It wouldn’t have been possible without artificial intelligence, which not only enabled the researchers to sort breast tumors into subgroups based on a dizzying array of genetic information but also to home in on shared attributes likely to be sensitive to treatment with existing drugs.
The researchers in the Terpsichore trial will also use AI to track the outcomes of the trial’s participants, noting whether and which tumors respond to a two-week window of experimental treatments, as well as helping them follow the patients in the years after their initial diagnosis.
“Most cancer clinical trials focus on metastatic disease, after the standard of care has already failed the patient,” Christina Curtis, PhD, Stanford Medicine’s director of artificial intelligence and cancer genomics, said. “But by that time, the cancer has had time to mutate and accumulate more genetic changes that drive its growth and make it resistant to treatment.”
Focusing on newly diagnosed, untreated tumors will reveal their original “starter pack” of mutations when the cancer is most vulnerable, Curtis and her colleagues reason. They hope this approach will improve breast cancer care for all patients. They also hope it will help answer one of the most pressing questions in breast oncology today: Why do about one-quarter of people with Pickett’s type of cancer, categorized as hormone receptor-positive, HER2-negative breast cancer, face a significant risk of recurrence decades after their diagnosis?
That broad window of risk far exceeds the five-year period cited by many oncologists as a reassuring milestone after which a patient can be considered cured. And the likelihood of recurrence — around 50% — for this subset of patients surpasses even that of triple-negative breast cancer, which has fewer treatment options and higher overall mortality than other breast cancer types.
Until recently, doctors had no way of knowing which of their breast cancer patients were at higher risk. But a body of work by Curtis and her colleagues over the past decade has pinpointed genetic changes that can be used to categorize breast cancer types into 11 clinically important groups and identify which of them are at heightened risk of recurrence.
They did so by training computers to exhaustively analyze tumors’ genomes — the complete set of genetic blueprints encoded by their DNA — and transcriptomes — the genetic messages, or RNA, that hint at the genes and proteins the cancer is using to survive.
The idea is not unique: Cancer researchers are increasingly turning to AI to parse the exponentially growing amount and types of data now gathered from patients. “This technology, and the power of the computational methods used to analyze this kind of information, will completely revolutionize how we think about the disease process,” said Jennifer Caswell-Jin, MD, assistant professor of medicine and principal investigator of Terpsichore. “In the past, we studied sections or samples of tumors; now we can analyze individual cells to identify new drug targets.”
They’re ready to apply their findings to newly diagnosed people like Pickett, with the aim of targeting the Achilles’ heels of tumors earlier in treatment than ever before.
“We believe the changes that increase the risk of recurrence years later are already there in the very earliest cancer cells,” Caswell-Jin said.
The researchers hope that the Terpsichore trial will help them reduce the likelihood of recurrence for high-risk patients and that it will reveal the biological underpinnings of how some tumors are able to cool their heels for years, remaining undetected before roaring back to cause a second, devastating round of disease.
Breast cancer lingo can be confusing. Not only are tumors categorized based on where in the breast they occur and which breast tissues are involved, but they are also grouped based on the proteins produced by the cancer cells. The presence or absence of certain proteins, such as human epidermal growth factor or the receptors for the hormones estrogen and progesterone, give clinicians clues about which biological signals are telling the cells to grow.
Blocking those signals with drugs or antibodies can slow or stop a tumor’s growth, and clinicians use this kind of molecular profiling to determine whether a breast cancer patient needs chemotherapy, radiation or hormone therapy like estrogen blockers — and for how long. (Most patients will also have some type of surgery to remove the cancerous tissue, either before or after other treatments.)
Broadly speaking, hormone receptor-positive breast cancer responds well to treatment, as do cancers in which the human epidermal growth factor receptor, HER2, is expressed at high levels. Cancers that don’t have elevated levels of estrogen receptor, progesterone receptor or HER2 — known as triple-negative breast cancers — are more difficult to treat and more deadly.
The concept of separating breast tumors into categories to guide treatment decisions and prognoses isn’t all that new. But at every step, it’s been limited by the technology available at the time. The four subgroups described above (hormone receptor-positive, HER2-negative; hormone receptor-positive, HER2-positive; hormone receptor-negative, HER2-positive; and triple negative) are often determined by a cell-staining technique called immunohistochemistry that has been around since the early 1940s.
Another test, OncotypeDX, was developed in 2004 and is based on the expression levels of just 21 genes. It is used to determine a patient’s five-year risk of recurrence and whether they should receive chemotherapy.
In 2012, Curtis, then at the University of Southern California, led a study that took a more complex approach. The researchers overlaid information about a patient’s genome — the DNA inherited from their parents — with that of the DNA sequences and RNA levels found in their cancer cells. RNA messages, which are selectively copied from DNA in the cell’s nucleus before traveling to its protein-making machinery, provide a snapshot of a cell’s operating instructions: divide now, make more of this gene, fire off a chemical signal to a nearby cell, etc.
The researchers took this approach because cancer cells are a messy bunch. The very act of running off the rails — casting aside any semblance of orderly growth or concern about cellular rule breaking — virtually ensures that they bobble the delicate series of events needed to correctly copy and divide their DNA before each cell division.
Like a wobbly top, every generation tilts a bit more out of control — accumulating an increasing number of mutations and even adding or losing copies of whole genes willy-nilly. As a result, cancer cells often have variable numbers of copies of important genes in their DNA, a genetic outcome known as copy number variation.
Often, this slow-motion molecular car crash results in the cell’s death. But sometimes changes occur that increase the cell’s fitness and allow it to climb to the top of the evolutionary dog pile. Cancer biologists call these changes drivers. In theory, blocking them will deal a significant, perhaps fatal, blow to the growing tumor.
The type of multifaceted analysis Curtis and her colleagues were attempting — comparing DNA sequences from healthy cells with DNA sequences and RNA levels from tumors, and doing so for multiple patients — is complicated.
Too complicated, in fact, for any one person or laboratory team to tackle. Instead, the researchers fed the information into a computer algorithm in an approach called unsupervised machine learning — allowing the computer to sift through millions of comparisons and derive its own conclusions based on the data available. It’s a hands-off approach that avoids bias.
“We wanted to see what kind of groupings would form in the data with minimal supervision,” said Curtis, who is now the RZ Cao Professor and a professor of medicine, of genetics and of biomedical data science at the Stanford School of Medicine. “This allowed us to see these cancers through a whole new lens and identify novel subgroups of disease.”
When Curtis published her 2012 study identifying the subgroups, doctors didn’t know whether or how to use the information to guide treatment. But in 2017, a different group published an eye-opening analysis of 75,000 people diagnosed with hormone receptor-positive, HER2-negative breast cancer that, for the first time, showed that about one-quarter of patients had a 50% chance of their tumors recurring even decades after their initial diagnosis. Unnervingly, even some patients whose cancers had not spread to their lymph nodes (a measure of metastasis) at diagnosis experienced recurrences at much higher rates than had been previously grasped.
“Doctors had seen unusual, late recurrences before in individual patients, but the patterns had not been systematically analyzed,” Curtis said. “When the data from thousands of patients was compiled, it suddenly became very clear that there is a subset of people at significant risk.”
In 2019, Curtis, Caswell-Jin and other researchers at Stanford Medicine and the University of Cambridge published a paper in Nature describing how it’s possible to combine information from immunohistochemistry and their new subgroups — termed integrative clusters — to not only predict which people were at increased risk of late recurrence but also to identify a subset of people with triple-negative tumors who were unlikely to see their cancers return after five years.
The researchers found that four of the 11 integrative subgroups were significantly more likely to return even 10 to 20 years after diagnosis. An analysis of each of their DNA and RNA profiles hints at possible reasons, but the driving factors aren’t the same for each group.
For example, although each of the four subgroups has tumors with increases in the numbers of copies of several cancer-associated genes called oncogenes, they differ in the number of copies of other genes involved in cancer cell survival and cell proliferation, including a notorious cancer driver called Myc. Many of the genes they appear to rely on for growth are known, and there are already approved drugs that block their actions.
“This molecular profiling gave us the different categories of tumors and helped us understand how these groups fare over time,” Caswell-Jin said. “Now we have the information to begin to figure out the right drugs and treatments to interrupt that path to poor outcome.”
Terpsichore is an apt name for an effort that requires an intricate dance to delicately balance patient care and a multi-armed experiment with many moving parts, each of which needs to mesh seamlessly over a period of about three weeks.
“We need to do this fast enough that we don’t delay the standard of care for these patients,” Curtis said. Currently, an in-depth genetic analysis of breast cancers is usually conducted only as a last-ditch effort to fight advanced metastatic disease, and the time pressure is less because the patients are already undergoing treatment. For Terpsichore, the researchers have set a goal of nine days in which to gather and analyze the genetic information needed to categorize each patient’s tumor into high or typical risks of recurrence using new and improved approaches optimized for clinical samples.
Because only one-quarter of people with hormone receptor-positive breast cancers will fall into the high-risk categories, Curtis and Caswell-Jin expect they’ll need to screen hundreds of people to find the 150 they’d like to include in the trial. Of those enrolled, about one-third will be in categories predicted to have a typical — that is, low — risk of recurrence after five years, and the other two-thirds will fall into groups with a higher risk.
Once patients enroll, they will be designated at random to receive either standard treatment for their cancers — the control arm of the study — or a 14-day treatment with drugs predicted to block the biological pathways that drive the growth of their tumors — the experimental arm. After two weeks, the researchers will assess the effect, if any, of the treatment on the growth of the tumor. All patients will then undergo a conventional course of treatment, including surgery and a yearslong course of hormone therapy. As the years tick by, Caswell-Jin and Curtis will monitor the participants’ health and disease status.
Since Pickett’s diagnosis, she’s had a lumpectomy to remove the cancerous tissue, followed by radiation and hormone therapy to stop the growth of her estrogen receptor-positive cancer cells. She is matter-of-fact about having been randomized to receive standard care, rather than the experimental intervention.
“I knew my participation in the trial would not be likely to have any direct benefit to my health,” she said. “But without these types of studies, researchers have no way of knowing if they are targeting the right pathways. My prognosis is good, but I know my cancer might return in 20 or even 30 years.”
Pickett, who would like to have another child, is particularly concerned about the reproductive effect of cancer drugs. “Cancer drugs have improved so much during the past 50 years,” she said. “They used to be incredibly toxic, but now we can target some specific pathways in some cells. Maybe one day we won’t have to shut down estrogen production entirely for a person with estrogen receptor-positive cancer. We started with a wrecking ball approach; now we’re down to a hammer approach; maybe one day, with studies like these, we can get down to a needle.”
Curtis and Caswell-Jin envision a future where even more layers of information can be integrated into ever more sophisticated models of breast cancer biology to help realize Pickett’s vision.
“We could integrate the genome and transcriptome data from a tumor with information from pathology, radiology and even spatial data showing where proteins are in a cancer cell or identifying neighborhoods of cell types,” Curtis said. “The more data we have, the more powerful these AI approaches can be. We should be leveraging it all to make advances much faster. We have the tools; it’s on us. Let’s move the dial so patients can benefit as soon as possible.”