AI and Biomarkers Refine Risk in HR+/HER2− Early Breast Cancer

Integrating artificial intelligence (AI) with digital pathology and clinical factors may identify a significant subset of patients with high-risk, hormone receptor (HR)-positive, HER2-negative early breast cancer who have an exceptionally low risk of distant recurrence.

Data presented at the 43rd Annual Miami Breast Cancer Conference suggest that combining independent prognostic signatures—including digital imaging and gene expression—provides a more nuanced risk assessment than current standard-of-care tools alone.¹

“[There have been] incremental but important steps…We should be thinking about what are realistic levels of clinical evidence for clinical utility of prognostic or predictive biomarkers, and how can we feasibly facilitate them so that our evidence remains strong in the future?” said W. Fraser Symmans, MD, in a presentation of the data at the conference.

At the conference, he addressed novel tissue-based prognostic and predictive biomarkers focusing on HR-positive/HER2-negative breast cancer, including prognosis, endocrine prediction, and chemoprediction.

Symmans is the Winterhof Family Chair and a professor in the Departments of Pathology and Translational Molecular Pathology at The University of Texas MD Anderson Cancer Center in Houston.

Beyond Prognosis: The Challenge of Prediction

The clinical challenge in HR-positive breast cancer remains the "prognostic paradox" of chemoprediction, where high-risk cancers are more likely to achieve pathologic complete response (pCR) yet still face a worse overall prognosis due to aggressive residual disease.

Symmans emphasized that while genomic "omics" have simplified prognosis, predicting treatment benefit remains difficult without clear mechanistic associations.

“Prognosis, we learned is quite easy. Prediction was really quite difficult, especially with empirical models, but prediction could be more readily achieved when there was some mechanistic relationship between the biomarker and the mechanism of the therapy,” he explained.

Emerging AI and multimodal models aim to bridge this gap by incorporating histopathologic features—such as mitotic rate and nuclear appearance—that have long been known to hold prognostic value.

Validation of AI and Digital Imaging Models

Recent studies have validated the utility of AI scores that combine clinical factors with digital pathology characteristics.

A prospective-retrospective analysis using the CANTO registry and the UNIRAD trial in France (n=633)² evaluated an AI score that combined 4 clinical factors with 12 digital pathology characteristics. The score identified 20% of clinically high-risk patients who had a distant recurrence-free interval (DRFI) event rate of only 4.6% at 9 years (stringent HR, 0.21; 95% CI, 0.09-0.52; P < .001).

Similar refinements were seen when combining the 21-gene Recurrence Score (RS) with the Sensitivity to Endocrine Therapy (SET2,3 Index). In the SWOG S8814³ and PACS-01⁴ trials, patients with a high SET2,3 index and an RS of 25 or lower, achieved a 5-year distant recurrence-free interval (DRFI) rate of 96.9% (95% CI, 93.6%-98.5%). Conversely, patients with low scores on both indices had significantly worse outcomes, with a 5-year DRFI of 65.0% (95% CI, 56.1%-72.5%). These results indicate that bundling independent prognostic tests can significantly improve risk assessment for node-positive, HR-positive disease.

“[We’re seeing] fully similar potential [in these trial examples],” Symmans said.

Similar performance was observed in models trained on the Recurrence Score (RS) from the TAILORx trial (NCT00310180).⁵ Researchers developed an AI model using digital hematoxylin and eosin (H&E) slides and clinical factors to mimic the RS; the model demonstrated prognostic performance similar to the genomic assay in validation cohorts, though it did not initially demonstrate independent chemoprediction.

The SET Index and Endocrine Prediction

Symmans highlighted the SET index, specifically the SET2,3 and SET ER/PR variants. Data from the NSABP B-42 trial (NCT00382070)⁶ demonstrated that a multimodal-multitask deep learning model (M3T)—trained on digital H&E slides, clinical factors, and bone density T-scores—could predict which patients benefit from extended letrozole (Femara) therapy after the initial 5 years of endocrine treatment. Additionally, the SETER/PR index of endocrine activity has shown potential in identifying patients who derive greater benefit from dose-dense or taxane-based chemotherapy regimens:

• SET ER/PR High (>1.50): Patients showed a 7.1% absolute difference in 10-year risk of breast cancer-free interval (BCFI) when treated with extended letrozole therapy compared to placebo (HR, 0.53; 95% CI, 0.32–0.88; P = .014).
• SET ER/PR Low (≤1.50): Patients showed only a 2.1% difference, which was not statistically significant (HR, 0.82; 95% CI, 0.56–1.20; P = .31).

Interestingly, Symmans noted that these findings both complement and contrast with other assays. While MammaPrint results from the same cohort indicated that ultra-low risk patients did not benefit from extended letrozole therapy, the SET index clarified that the primary driver of extended therapy benefit is the level of endocrine activity within the tumor, not just its prognostic risk category. Symmans emphasized that prognostic risk and endocrine activity are distinct biological axes; a tumor can be high-risk but also highly endocrine-sensitive, making it an ideal candidate for treatment intensification.

Advances in Chemoprediction

Selecting the optimal chemotherapy regimen represents the next frontier. Translational hypotheses suggest that cancers with low endocrine activity (as measured by SET ER/PR < 0.75) may be more vulnerable to dose-dense chemotherapy regimens.

Data from the CALGB 9741 trial, which compared dose-dense versus conventional AC-T (doxorubicin/cyclophosphamide followed by paclitaxel), supported this. In the ER+ subset, patients with low SET ER/PR scores experienced late benefits from dose-density, whereas those with high endocrine activity saw minimal difference between regimens (P_interaction = .03). Similar results were seen in the GEICAM/9906 trial, where the addition of weekly paclitaxel specifically benefited patients with SET ER/PR scores less than 0.75 (HR, 0.48; P = .049).⁷

These findings were further bolstered by an analysis of the FLEX registry, an observational study of nearly 10,000 patients. While the FLEX data is non-randomized and uses propensity-matched cohorts, it confirmed that higher-risk prognostic biology is consistently associated with better outcomes from chemotherapy, particularly taxane-based regimens.⁸

Expert Interpretation: A Cautious Outlook

Symmans noted that "prognosis was quite easy" but "prediction was very difficult using empirical models. The shift toward mechanistic associations—such as linking ER pathway activity directly to endocrine therapy benefit—is essential for moving beyond the "one-size-fits-all" approach. While the AI signatures perform exceptionally well at identifying low-risk subsets within clinically high-risk populations, clinicians must remain cautious about overstating their impact until these tools are commercially available and integrated into routine pathology workflows.

The "take-home" message for the oncology community is the potential for bundling independent prognostic and predictive signatures. By combining digital H&E signatures with genomic indices of endocrine activity, clinicians can gain a multi-dimensional view of a patient’s disease.

“Selected combinations of independent prognostic signatures have in certain examples, demonstrated some degree of improvement. There may be a time where two biomarkers make sense for certain clinically higher risk populations, gaining clearer insight into immigrant prediction,” Symmans concluded.

REFERENCES

1. Symmans WF. The Potential of Artificial Intelligence and Other Novel Biomarkers. Presented at: 43^rd Annual Miami Breast Cancer Conference; March 5-8, 2026; Miami, Florida.

2. Bidard FC, Gessain G, Bachelot T, et al. Identifying Patients With Low Relapse Rate Despite High-Risk Estrogen Receptor-Positive/Human Epidermal Growth Factor Receptor 2-Negative Early Breast Cancer: Development and Validation of a Clinicopathologic Assay. J Clin Oncol. 2025;43(28):3090-3101. doi:10.1200/JCO-25-00742.

3. Speers CW, WF Symmans, Barlow WE, et al. Evaluation of the Sensitivity to Endocrine Therapy Index and 21-Gene Breast Recurrence Score in the SWOG S8814 Trial. J Clin Oncol. 2023;41(10), 1841-1848. doi:10.1200/JCO.22.0149.

4. Penault-Llorcaa F, Lusqueb A, Filleronb T, et al. Combination of predicted sensitivity to endocrine therapy (SET2,3 index) and the recurrence score® in node-positive breast cancer: Independent validation in the PACS-01 trial. Eur J Cancer. 2026;233:116152. doi:10.1016/j.ejca.2025.116152.

5. Boehm KM, El Nahhas OSM, Marra A, et al. Multimodal histopathologic models stratify hormone receptor-positive early breast cancer. Nat Comm. 2025;16(1):2106. doi:10.1038/s41467-025-57283-x.

6. Mamounas E, Wang V, Chen M, et al. A Multimodal-Multitask Deep Learning Model Trained in NSABP B-42 and Validated in TAILORx for Late Distant Recurrence Risk in HR+ Early Breast Cancer. Presented at: 2025 San Antonio Breast Cancer Symposium; December 9-13, 2025; San Antonio, Texas. Abstract RF3-07.

7. Metzger Filho O, Ballman K, Campbell J, et al. Adjuvant Dose-Dense Chemotherapy in Hormone Receptor-Positive Breast Cancer. J Clin Oncol. 2025;43(10):1229-1239. doi:10.1200/JCO-24-01875.

8. Brufsky AM, Hoskins KF, Conter HJ, et al. MammaPrint predicts chemotherapy benefit in HR+HER2- early breast cancer: FLEX Registry real-world data. JNCI Cancer Spectr. 2025;9(5):pkaf079. doi:10.1093/jncics/pkaf079.