Revolutionizing Treatment: Research Harnesses Machine Learning to Predict Chemotherapy Response in Triple Negative Breast Cancer Patients

Triple negative breast cancer is a highly aggressive form of cancer and is one that grows and spreads fast. According to the American Cancer Society, it accounts for 10 percent to 15 percent of breast cancers and is one that has fewer treatment opportunities for patients.

Chemotherapy is a common treatment for TNBC, but studies show that only 40 percent of tumors respond to the treatment — meaning the tumor reduced in size partially or completely. Chemotherapy has several side effects, including nausea, hair loss and fatigue, which can be severe for some patients and decrease their quality of life.

With a 60 percent chance that a tumor will not respond to chemotherapy, researchers from the University of Alabama at Birmingham, Georgia State University and the University of Galway in Ireland sought to determine whether responses to the treatment could be predicted before a patient’s first infusion.

“We created an algorithm using machine learning that can predict in advance, with high accuracy, which woman’s tumor will respond to treatment versus those with tumors that will not respond,” said Ritu Aneja, Ph.D., associate dean for Research and Innovation in the School of Health Professions. “This will allow those with non-responding tumors to be spared the unnecessary side effects that come with chemotherapy.”

To distinguish between responders and non-responders, Aneja and the team of researchers examined biopsy tumor samples from women with TNBC before treatment administration. They inspected the tumor’s microenvironment to see whether there was any difference between the tumors that respond and those that do not respond to treatment. The tumor microenvironment is the complex ecosystem that surrounds a tumor inside the body. It includes the tumor cells, immune cells, the stroma, blood vessels, and other cells that surround and feed the tumor cells.

“A tumor and its microenvironment constantly interact and influence each other, either positively or negatively,” Aneja said. “Our rationale for studying the TME was its profound influence on a tumor’s chemo-responsiveness. We used machine learning algorithms to successfully extract features and uncover patterns from the TME that distinguished responders from the non-responders.”

Using the power of TME histological components, the machine learning model was able to correctly predict 42 of 51 cases that would positively respond to chemotherapy, and 29 out of 34 cases were correctly identified as patients who would be non-responders. Aneja hopes that they can continue to improve the prediction algorithms by using larger patient datasets that incorporate patients on various treatment regimens. This will help fine-tune the algorithm and make it more generalizable for use on patients.

“I believe that team science is the way forward,” Aneja said. “The problem of over- and under- treatment is huge in cancer clinics, and we cannot solve such complex at the level of an individual lab or even a single discipline. Addressing such intractable problems requires coming together as a team and collectively changing the way we think.”