Gillings researchers use genomics, machine learning, and population-based research to improve breast cancer prevention, diagnosis, and treatment.
A decades-long population study, recent advances in genomics and machine learning, and a culture of collaboration are helping UNC researchers work together to find ways to improve the prevention, diagnosis and treatment of breast cancer.
Translating Science into Practice: Doing great public health research means putting study results to work. Whether examining molecular-level data or facilitating large-scale organizational changes, we’re Gillings. We’re on it! Gillings faculty go beyond the literature and the lab to make real changes in health-care practice and policies — and real differences for patients.
Melissa Troester, PhD, professor of epidemiology and research professor of pathology and laboratory medicine at the UNC Gillings School of Global Public Health, is the principal investigator on the Carolina Breast Cancer Study (CBCS), a study of breast cancer epidemiology and biology launched more than a quarter-century ago to understand why African-American women disproportionately die from breast cancer.
Since 1993, the study has gathered data on more than 8,000 women from 44 counties in North Carolina. Now in Phase 3 of the study, researchers are conducting a more detailed analysis on how people are interfacing with the health-care system — for example, what kind of therapy they receive and when and whether they have comorbidities, like diabetes or heart disease — and integrating that information with molecular data.
“CBCS has a long history of interdisciplinary science and national and international collaboration,” says Andy Olshan, PhD, Barbara S. Hulka Distinguished Professor of epidemiology at Gillings and CBCS co-principal investigator. “For example, CBCS is collaborating with the National Cancer Institute and an international consortium to explore the genetics of breast cancer. CBCS has also partnered with two other studies to form the world’s largest consortium to study the epidemiology of breast cancer among African-Americans.”
The CBCS seeks to integrate advances in molecular genomics with population-based research. In particular, one recent advance — a National Cancer Institute initiative called The Cancer Genome Atlas project — has led to a clearer picture of tumor genetic variability. “This data is typically not available in large population-based studies and the CBCS seeks to help close this gap,” Troester says.
“We’ve learned that what happens clinically is determined by tumor biology,” she says, “and as public health researchers we want to integrate this information with how people use and access health care.”
Working in collaborative teams across disciplines, Troester and her colleagues use several different methods and technologies to strengthen their understanding of cancer tumors. One area of focus is examining pathology data with “deep learning” — having computers use algorithms to look for unique features of tumor tissues that could help predict how the tumor might progress. Machine learning could identify features that scientists have missed so far, and may be able to standardize assessment of the tumors. For some features, like specific markers indicating tumor aggressiveness, pathologists may agree only 80 percent of the time on whether the tumor is high grade.
Transfer learning is another machine-learning approach that researchers hope to use to predict cancer progression risk so that doctors can select the best treatments and detect recurrence earlier and more rapidly so fewer people experience the worst cancer outcomes. Based on huge databases of images, computers can detect differences between different kinds of animals or objects (dogs and cats, or tables and chairs). Hoping to apply models trained initially on this kind of “computer vision” to tumor tissues, Troester and her collaborators — Marc Niethammer, PhD, professor of computer science; James Stephen Marron, PhD, the Amos Hawley Distinguished Professor of statistics and operations research; and Charles Perou, the May Goldman Shaw Distinguished Professor of molecular oncology — are feeding various images of different types of cancers into a machine learning algorithm.
The scientists aim to use machine learning to find image features that link to genetics to distinguish among different types of cancers, and use these images to predict survival by linking molecular data with the actual outcomes. Building models that have both molecular biomarker data and image data from the tumors might do a better job of predicting risk than models that use only one of these data types.
Microscopic images of tumors taken for biopsies have long been used to predict outcomes, using features like tumor grade or cellular differentiation. “Our hope with this project is that we can learn new image features if we ask computers to distinguish between tumors that are aggressive and those that are not,” Troester says.
Deep learning also has the potential to integrate two different key types of data on breast cancer — mammograms and biopsy histology images, essentially two different photos of cancer, to help distinguish between aggressive and benign tumors, enabling doctors to better assess risk and supporting precision health both for prevention and treatment.
“We can bring together nationally renowned experts in their fields, and we are really lucky in terms of the depth of collaborators we have here at Carolina.”
Melissa Troester, PhD
Professor of Epidemiology and Research
Professor of Pathology and Laboratory Medicine
One million women a year have biopsies. For most of them, results come back as benign, but there is a limit to how much additional information they get back. “We think we can do better and that patients deserve to have better information, especially after a biopsy that’s a relatively invasive procedure,” Troester says. “Extending what we’ve learned about tumors could have benefits on the prevention side as well as the treatment side.”
The CBCS is pursuing many different research directions and takes advantage of advances in a lot of areas — computer science, molecular profiling, social determinants, and health services — so it relies critically on working with investigators across disciplines and across multiple schools within UNC. Troester says, “We can bring together nationally renowned experts in their fields, and we are really lucky in terms of the depth of collaborators we have here at Carolina.”
