Healthcare

Deciphering Breast Cancer Heterogeneity Using Machine Learning

Share this post:

As per the latest report from the World Health Organization, the global cancer burden is estimated to have risen to 18.1 million new cases and 9.6 million deaths in 20181. Of this total, female breast cancer contributes 11.6 percent and is one of the top three cancer types worldwide, followed by lung and colorectal cancer1. Breast cancer is the most commonly diagnosed cancer in women with more than 1.7 million new diagnoses each year and half a million cancer-related deaths1.

The Systems Biology Group led by Maria Rodriguez Martinez at IBM Research-Zurich has been working in collaboration with Professor Bernd Bodenmiller’s Group at the Institute of Molecular Life Science, University of Zurich (UZH) to elucidate cellular and phenotypic diversity in breast cancer ecosystems. We used mass cytometry, which enables the measurement of over 40 parameters in millions of cells simultaneously at the single-cell level, and machine learning to identify and classify tumor and immune cell types as well as their relationships.

The goal is for this work to lay the foundation for future precision medicine approaches that could potentially help patients win the fight against breast cancer. Our work is featured in the current issue of Cell2; the print version was released today.

A need for personalized treatment

While researchers have been working hard to develop novel therapeutic approaches to fight against breast cancer, the main reasons for cancer-associated deaths are still therapy resistance, relapse, and metastasis. Methods to better target cancer cells and treat the tumor-associated immune system have indeed progressed over the last years. Still, very little is known about the cellular composition of tumors, which can make it challenging for physicians to provide individual breast cancer patients with the unique medical treatment they need. We therefore have a strong need to decipher the diversity of cancer cells and immune cells and their relationships within the breast tumor ecosystems of patients.

Understanding cancer ecosystems

My colleague Johanna Wagner (PhD student at UZH) and I set out to fill this knowledge gap. Our premise was that breast cancer is a heterogenous disease. We knew that breast cancer ecosystems are comprised of tumor cells that communicate and interact with surrounding non-cancer cell types, including immune cells, stromal cells, and cells of the vasculature. Cancer cells and tumor-associated cells are phenotypically and functionally heterogeneous with characteristics determined by both genetic make-up and environmental influences. Taking this into consideration, we hypothesized that patterns within the heterogeneous nature of breast cancer ecosystems might be linked to disease progression and response to therapy.

Forming a team of IBM and UZH researchers, we began investigating breast tumor and non-tumor samples from 144 patients using mass cytometry. In contrast to bulk approaches, this novel technology enables the comprehensive analysis of tumors at a single-cell level. With mass cytometry, we simultaneously quantified more than 70 proteins in over 26 million cancer and immune cells. The resulting data was very challenging to analyze, which is why we turned to advanced machine learning models.

Breast tumors are complex ecosystems. By combining volumes of tumor data uncovered with mass cytometry with the analysis capabilities of machine learning, researchers were able to identify and distinguish between the many different types of cells interwoven within tumors. Credit: Cell.

Using a combined experimental and computational approach with machine learning, we were able to identify various different populations of tumor and immune cells and create a detailed atlas of breast cancer ecosystems. Moreover, we successfully examined and precisely defined the heterogeneity of individual tumors and quantified their abnormality in comparison to matched non-tumor tissue. But that’s not all: we also made a thorough analysis of tumor-associated macrophage and T cell populations, which can exert both tumor-suppressing and tumor-supporting functions. Last, our findings were associated with clinical information, such as disease grade or tumor aggressiveness.

The discovery

Our “aha” moment was discovering that a previous belief of diversity being increased in more aggressive tumors does not hold up. Actually, the more aggressive tumors were often found to be dominated by a single tumor cell phenotype, which was highly abnormal and distinct between patients. And as we suspected, each of the studied tumors was unique in its cellular composition, with the more aggressive tumors differing most from the rest of the cohort. This could be a reason why a one-size-fits-all approach to cancer treatment is not always effective3.

Immunotherapy for breast cancer

In addition to characterizing tumor cell phenotypes and tumor abnormality and individuality, we also found similarities in the tumor-associated immune system among more aggressive tumors. In one subgroup of breast cancer patients, we discovered an accumulation of exhausted T cells and tumor-associated macrophages that have been associated with immune inactivation, which can in turn support cancer growth. Such immune cell types have previously been identified in lung cancer4 and melanoma5 patients and, in some instances, have been successfully treated by immunotherapy.

Based on our findings, we believe that a specific group of breast cancer patients could potentially benefit from immunotherapy as well. Moving forward, we will investigate the possibilities of immunotherapy in additional studies, which could potentially lead to a clinical study. Our comprehensive tumor and immune atlas of breast cancer ecosystems lays the foundation for further research into the design of precision medicine approaches and indicates a potential success of immunotherapy in a group of breast cancer patients2.

Cellular relationships

Much of our success was due to our interdisciplinary approach and a highly motivated team. The cooperation of researchers from various backgrounds was paramount to recognizing and understanding the complexity and diversity of a disease such as breast cancer. Our investigation has shown that tumor ecosystems are diverse in their cell composition as well as shaped by the relationships between their cell subpopulations. Naturally, strategies to target these cellular relationships within cancer ecosystems, particularly those that facilitate tumor growth, hold the potential to advance breast cancer treatment. Just imagine if women could one day fight the disease at fair odds with therapies tailor-made to their medical needs. Our single-cell atlas of breast cancer ecosystems is therefore an important step in research that may lead to precision medicine approaches that target both the tumor and its immune environment, which could ameliorate disease outcome for patients in the future.

This project was made possible through a grant from MetastasiX.


  1. https://www.who.int/cancer/PRGlobocanFinal.pdf
  2. Wagner, J. et al. A single-cell atlas of the tumor and immune ecosystem of human breast cancer. Cell. April 11, 2019. DOI: 1016/j.cell.2019.03.005
  3. Ramos, P. and Bentires-Alj, M. Mechanism-based cancer therapy: resistance to therapy, therapy for resistance. Oncogene. 2015; 34: 3617-3626
  4. Lavin, Y. et al. Innate Immune Landscape in Early Lung Adenocarcinoma by Paired Single-Cell Analyses. Cell. 2017; 169: 750-765.e1
  5. Sade-Feldman, M. et al. Defining T Cell States Associated with Response to Checkpoint Immunotherapy in Melanoma. Cell. 2018; 175: 998-1013.e20

IBM Research

More Healthcare stories

AI Models Predict Breast Cancer with Radiologist-Level Accuracy

Our team of IBM researchers published research in Radiology around a new AI model that can predict the development of malignant breast cancer in patients within the year, at rates comparable to human radiologists.

Continue reading

AI Offers Hope for Earlier Screening for Type 1 Diabetes

IBM researchers launch the first in-depth, large scale study of time courses when different antibodies appear, and their correlations to type 1 diabetes.

Continue reading

IBM and Johns Hopkins University School of Medicine Discover Unique, Pathogenic Autoimmune Cells in Type 1 Diabetes

Researchers at JHU and IBM identify new cells involved in the autoimmune attack on insulin-producing cells in type 1 diabetes.

Continue reading