Skip to main content
Premium Trial:

Request an Annual Quote

Breast Cancer Therapy Response Predicted by Multiomic Machine Learning Model

NEW YORK — By bundling together omic and other data on the breast cancer tumor ecosystem, researchers have developed a tool to predict which patients will likely respond to therapy.

Breast cancer patients are increasingly treated with chemotherapy with or without targeted therapy prior to surgery, but the response to these treatments differs. To see whether response can be predicted from the tumor, a team led by scientists at the University of Cambridge analyzed pre-treatment tumor biopsies from more than 160 breast cancer patients, amassing not only clinical and pathological data, but also genomic and transcriptomic profiles.

Senior author Carlos Caldas, a professor of cancer medicine at Cambridge, said that physicians currently rely on clinical features, like tumor size or grade and IHC markers, to guide treatment. "This is better than flipping a coin but still very imprecise," he wrote in an email.

As he and his colleagues reported in Nature on Tuesday, they combined the data they collected on the tumor ecosystem and pathological endpoints to develop a machine-learning tool that predicts whether tumors are likely to respond to treatment.

They enrolled 180 women with early and locally advanced breast cancer who were undergoing neoadjuvant treatment into their profiling study. For 168 cases, they obtained pre-treatment tumor biopsies for whole-genome, whole-exome, and RNA sequencing analysis. They further collected data on tumor grade, ER receptor status, and lymph node involvement.

Their sequencing analysis uncovered more than 16,000 somatic mutations, including shared ones in driver genes like TP53, PIK3CA, and GATA3. In particular, genomic features like high tumor mutation burden, as well as homologous recombination deficiency and APOBEC signatures, were associated with pathological complete response among patients.

At the same time, the researchers identified 2,071 genes that were underexpressed and 2,439 genes that were overexpressed in tumors that reached pathological complete response. These genes were in particular involved in proliferation and immune activity. The researchers additionally found that there was an enrichment of both innate and adaptive immune cell populations in the tumor microenvironment of both ER+HER2- and HER+ tumors that reached pathological complete response.

Together, these findings suggested that proliferation and immune signatures pointed toward tumors that were more likely to respond to treatment. Also, tumors that are highly proliferative and have an active tumor microenvironment seemed to be more likely to reach pathological complete response.

Based on their findings, the researchers developed a machine-learning framework that combined tumor features into a model to predict pathological complete response. The key factors weighed in the model included age, lymphocyte density, and PGR, ESR1, and ERBB2 expression, though it also included features linked to immune activation and immune evasion. After training the model, they tested it on a separate cohort of 75 patients to find it could predict tumor response with high accuracy.

This suggested to the researchers that their model could be applied in the clinic to predict therapy response and guide treatment selection, as well as to identify patients for enrollment in a clinical trial.

Before then, more study is needed, though. "The next step is to further validate our findings in even larger series," Caldas said, noting that he and his team are already working on that. "Then, if confirmatory studies are positive, you need to implement it into the healthcare system. That will require dedicated comprehensive cancer centers with the required expertise and multi-disciplinary teams," he added.