A study presented at the ESMO Congress 2026 (Abstract LBA20) demonstrated the clinical validity of an AI-driven version of the MammaPrint assay that predicts the 70-gene recurrence score directly from whole-slide images of tissue from patients with early breast cancer.  

“This digital pathology MammaPrint demonstrates prognostic value for both distant recurrence–free interval and breast cancer–specific survival, is highly concordant with the microarray MammaPrint assay, and shows that it is feasible to do this from standard routine formalin-fixed paraffin-embedded hematoxylin and eosin–stained [FFPE H&E] whole-slide images of early breast cancer,” said presenting author Laura J van ‘t Veer, PhD, Angela and Shu Kai Chan Endowed Chair in Cancer Research at UCSF, Co-Leader of the Breast Oncology Program, and Director of Applied Genomics at the UCSF Helen Diller Family Comprehensive Cancer Center.  

Background 

The 70-gene MammaPrint recurrence signature for early breast cancer was developed and published in 2002. The signature indicated a patient’s risk of recurrence and could be used to guide decisions about adjuvant chemotherapy. The prognostic test was subsequently cleared by the U.S. Food and Drug Administration in 2007.  

Further research on the MammaPrint signature has provided additional information about recurrence risk levels in patients with early breast cancer, with implications for the use of chemotherapy and the need for extended endocrine therapy.  

Study and Model Methods 

Researchers developed a digital assay based on the 70-gene recurrence score to enable the prognostic information captured by MammaPrint to be inferred directly from FFPE H&E–stained whole-slide images.  

An AI-based digital pathology model was trained on 70,000 FFPE slides with available MammaPrint microarray results. The whole-slide images then underwent preprocessing to ensure they contained sufficient invasive tumor tissue. During preprocessing, tissue was detected, patches were extracted, and seven tissue types were classified.  

The pipeline used self-supervised feature recognition followed by multiple-instance learning to identify MammaPrint patterns, as well as “one-shot” microarray MammaPrint-supervised label prediction. The outputs were then combined in a locked-down ensemble module.  

Once a new slide passed preprocessing quality control, the module generated a digital MammaPrint index, a continuous score representing predicted recurrence risk.  

The process was first tested in an independent set of 1,000 whole-slide images. In this cohort, the digital MammaPrint results showed a 95% correlation with the microarray MammaPrint results.  

The  assay then underwent validation for precision, reproducibility, limit of detection, and other performance characteristics as part of the regulatory review process for the in vitro diagnostic test, which is still pending. 

Study Findings 

The researchers used real-world evidence from the RASTER study to independently validate the AI Digital MammaPrint Index. The study included 427 patients diagnosed with cTanyN0M0 breast cancer between 2004 and 2006 in the Netherlands. All patients underwent testing with the 70-gene recurrence signature, and the results were considered in decisions regarding adjuvant systemic therapy.  

Ten-year outcome data are also available for these patients, including a large subset with estrogen receptor (ER)–positive, HER2-negative tumors. Dr. van ‘t Veer noted that previous findings from the study suggested that patients at high clinical risk but with genomically low-risk tumors had excellent outcomes.  

Primary outcome measures were breast cancer–specific survival and distant recurrence–free interval. 

Among 372 evaluable patients from the RASTER study, 51% were classified as low risk by the Digital MammaPrint Index and 49% as high risk. Chemotherapy was received by 74% of patients classified as high risk by the AI score and 24% of those classified as low risk.  

At 10 years, the breast cancer–specific survival rate was 98.1% for patients with Digital MammaPrint low-risk disease and 86.7% for those with high-risk disease (hazard ratio [HR] = 0.12, 95% confidence interval [CI] = 0.04–0.39; P < .0001). In the subset of patients with ER-positive, HER2-negative disease, the HR was 0.10 (95% CI = 0.03–0.36; P < .0001).  

In multivariable analysis, the Digital MammaPrint Index was the strongest predictor of breast cancer–specific survival (HR = 0.084, 95% CI = 0.02–0.31; P < .001), outperforming other clinical factors. Similar results were observed for distant recurrence–free interval.  

“We can integrate this MammaPrint assay now with other AI features, and it will form the foundation for further image-based biomarker developments. So, [the study] also shows us that AI digital pathology broadens the use of molecular-based assays in clinical practice,” Dr. van ‘t Veer said.  

Insights 

Invited discussant Luca Boldrini, MD, PhD, of the Fondazione Policlinico Universitario A. Gemelli IRCCS and Università Cattolica del Sacro Cuore in Rome, Italy, emphasized that a surrogate test must be shown to support the same clinical decisions as the assay it would replace. Referring to the correlation reported in the RASTER data, he explained that correlation is not the same as concordance.  

Although the Digital MammaPrint demonstrated a 95% correlation, clinical decisions depend on the risk classification threshold, where discordant cases become particularly important. “So this is still safe, because nobody here is undertreated, but it may erode precisely what MammaPrint stands for: safely sparing from chemotherapy,” Dr. Boldrini said.  

Regarding the RASTER data, Dr. Boldrini noted several limitations of the real-world evidence, including the small number of breast cancer–specific deaths, which made the prognostic findings less robust.  

He suggested that, to make the Digital MammaPrint suitable for clinical decision-making, the digital assay must demonstrate analytical robustness across scanners, staining methods, and sites; categorical concordance, particularly in patients at high clinical risk but low genomic risk; clinical validity through prospective-retrospective analyses of randomized clinical trials; and independent external validation.  

DISCLOSURES: Dr. van ‘t Veer is a part-time employee and stockholder of Agendia NV.  Dr. Boldrini is a former member of the IBA Victoria Advisory Committee; has been a scientific consultant for Ergea Medipass srl, Roche, Radius srl, J&J, and GE Healthcare; has been a sponsored researcher for Nanovi, Sophia Genetics, View Ray Technologies, and Siemens Varian Healthineers; and is an inventor on a patent.  

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.