AI-Powered Biomarkers for Predicting Chemotherapy Response in Breast Cancer

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

Breast cancer is one of the most common and deadly cancers worldwide. Chemotherapy is often used to shrink the tumor before surgery, but not all patients respond equally. How can doctors predict which patients will benefit from chemotherapy and which will not?

In a new study, researchers at the Radboud University Medical Center (RUMC) and the Netherlands Cancer Institute (NKI) proposed computational biomarkers based on deep learning to predict pathological complete response (pCR) to neoadjuvant chemotherapy in breast cancer, using only routine H&E-stained biopsy slides.1

Although further evaluation and fine-tuning are needed before clinical application, the study demonstrated the potential of artificial intelligence (AI) to predict which patients with breast cancer will respond to commonly used pre-surgery chemotherapy from routine biopsy slides, allowing oncologists to better personalize treatment.

“We found that our approach to encode computational tumor-infiltrating lymphocytes had predictive value across multiple cohorts,” said Associate Professor Francesco Ciompi, PhD, who was, together with Dr. Esther Lips from the NKI, the senior investigator and corresponding author of this study. “Furthermore, for the first time to the best of our knowledge, we showed that a biomarker derived from mitotic count in H&E has predictive value in a neoadjuvant setting.”

The work was developed by Witali Aswolinskiy, PhD, and funded by the Dutch Cancer Society (KWF; PROACTING project number 11917) and the European Union’s Horizon 2020 research and innovation program (ExaMode project, grant agreement No 82529, htttp://www.examode.eu/). The results of this study were published in Breast Cancer Research.

 

Unmet Need: Predicting Response to Neoadjuvant Chemotherapy

Neoadjuvant chemotherapy is increasingly being used for the treatment of breast cancer. The use of chemotherapy before surgery aims to shrink tumors to enable breast-conserving surgery.1 However, only a fraction of women with breast cancer respond. This exposes non-responders to serious side effects and delays surgery while their disease continues to progress.

According to Dr Ciompi, there is evidence that the rates of pathological complete response (pCR) are particularly low for patients with triple-negative breast cancer (TNBC) and those with luminal B breast cancer.

The use of biomarkers to predict treatment response would help spare non-responders from the toxic effects of chemotherapy, reduce treatment costs, and provide the opportunity to offer a better and more personalized treatment option.

“We only want to treat the patients who we know will gain clinical benefits from this aggressive and toxic treatment, and for the other ones perform surgery right away or consider other treatment options,” explained Dr Ciompi.

 

A Computational Approach for Quantifying Immune Cells and Cell Division in Tumors

The researchers developed a computational method to analyze the tumor microenvironment — the mix of cancer and immune cells in the tumor — from which biomarkers can be derived.

Their approach uses deep learning to automatically detect two robust biomarkers from routine biopsy slides: immune cells called tumor-infiltrating lymphocytes (TILs) and cell division. TILs reflect the immune system attacking tumors, while proliferation indicates quickly dividing cancer cells.1 More TILs and less proliferation suggest that chemotherapy is more likely to work.

“Based on evidence from the literature, we focused on some aspects of the tumor microenvironment that may correlate with treatment response, such as the interaction of immune cells with the tumor, the tumor-associated stroma, and the growth of tumor cells,” noted Dr Ciompi.

The proposed approach has two steps: 1) Use deep learning (U-Net and a pre-trained model) to segment slides into tumor, lymphocytes, stroma, and necrotic tissue and detect mitoses; 2) Derive computational biomarkers encoding relationships between components like tumor, lymphocytes, stroma, and mitoses.1

From this, the scientists computed four morphologically interpretable biomarkers encoding relationships between components like tumor cells, lymphocytes, and mitotic figures in H&E-stained slides of routinely available diagnostic (core needle) biopsies of breast cancer tissues.

The group assessed their biomarkers in TNBC and luminal B breast cancer cases from 721 patients in the Netherlands and Italy. They also validated the biomarkers in 126 cases from a public dataset.

“Using different cohorts, we correlated each computational biomarker with patient outcomes in terms of pCR. We performed internal validation on data from the NKI and RUMC. We also applied the biomarkers to external independent data from an Italian hospital, Sacro Cuore Don Calabria in Verona, as well as to data from the recent public study IMPRESS,” said Dr Ciompi.

 

Computational Biomarkers Predict Response to Adjuvant Chemotherapy

All computational biomarkers achieved promising predictive performance, with areas under receiver operating characteristic curves (AUCs) ranging from 0.66 to 0.88 for predicting complete response across subtypes and centers. TILs achieved 100% sensitivity, but false positive rates were high.1

“For TNBC, compared to clinical practice where almost all patients are treated, using computational biomarkers to predict response would allow identifying all responders for treatment, and about 40%–50% of the non-responders, but would also spare the rest of non-responders from the therapy, which is already an improvement compared to current practice. Of course, this needs to be improved, and false positive rates need to drop while keeping high sensitivity.”

TILs were the most robust computational biomarkers across cohorts, with the highest AUC score in the Dutch cohort with luminal B breast cancer. The AI-generated TIL scores also matched or slightly improved pathologists’ visual TIL estimates.

Counting dividing cells computationally demonstrated some potential in predicting pCR in women with breast cancer but was less reliable than TILs, failing to significantly predict outcomes in several breast cancer subtypes.

 

Larger Studies Needed Before Clinical Use

The scientists cautioned that larger studies are warranted to validate the AI biomarkers before they can be adopted in clinical practice. The number of patients in the current sub-analyses was relatively small, limiting statistical power. In addition, the team focused on analyzing individual biomarkers; thus, the predictive power of the biomarkers when used in combination remains to be determined.

Although further validation is needed, the interpretable algorithms can quantify key aspects of the tumor environment known to influence chemotherapy success from standard biopsy slides. With additional validation, they could form the building blocks of sensitive predictive tests to determine the optimal treatment for patients with breast cancer based on the unique characteristics of each person’s disease.

Commenting on the generalizability of their findings, Dr Ciompi said, “We derived explainable biomarkers based on measurable features, such as the amount of tissue or number of cells in the slide, which can also be assessed using other algorithms or visually by pathologists, not constraining these results to models developed in our group, opening to broad validation studies.”

To facilitate progress in this field, the team made part of the methodology publicly available via the grand-challenge.org web platform, where researchers can upload their digitized tissue slides to obtain quantification scores for computational biomarkers.

The study was funded by the Dutch Cancer Society and the European Union’s Horizon 2020 research and innovation program.

 

References

  1. Aswolinskiy W, Munari E, Horlings HM, et al. PROACTING: predicting pathological complete response to neoadjuvant chemotherapy in breast cancer from routine diagnostic histopathology biopsies with deep learning. Breast Cancer Res. 2023;25(1):142. Published 2023 Nov 13. doi:10.1186/s13058-023-01726-0

 

Share This Post

Leave a Reply