A machine learning assay to predict disease recurrence in hormone receptor-positive breast cancer

Abstract

Background

Hormone receptor (HR)-positive, human epidermal growth factor 2 (HER2)-negative breast cancer associates with a sustained risk of relapse over time. Current multigene assays offer limited validity to identify clinically low-risk tumors at high risk of recurrence, which is particularly relevant in the context of novel adjuvant therapies. In this study, we developed and validated ER-Predict, a machine learning assay leveraging a 14-gene expression signature to classify early stage HR-positive/HER2-negative breast cancer according to the risk of relapse.

Materials and methods

ER-Predict was developed on a cohort of 1413 HR-positive/HER2-negative early breast cancer cases. External validation was carried out across eight publicly available cohorts (n = 1118). Comparative benchmarking was conducted against reproduced prognostic signatures, and particularly, EndoPredict and Oncotype DX. Functional annotation and drug response analysis were carried out using gene expression and pharmacogenomic data from publicly available breast cancer cell lines.

Results

ER-Predict identified high-risk patients with significantly reduced distant metastasis-free survival in the external validation cohort (hazard ratio 2.03, 95% confidence interval 1.57-2.63, P < 0.0001). The assay demonstrated independent prognostic value beyond traditional clinicopathological factors, including tumor grade, tumor size, and nodal status, and consistently outperformed recomputed multigene panels. ER-Predict showed specificity toward luminal-like transcriptomic programs, revealing activation of key cell-cycle regulators governing endocrine resistance among high-risk tumors that had retained expression of retinoblastoma-1, suggesting potential actionability by means of cell-cycle inhibitors.

Conclusions

ER-Predict represents a robust assay with potential utility in early stage HR-positive/HER2-negative breast cancer. Its consistent ability to identify high-risk patients supports further investigation as a decision-support tool to guide treatment intensification in clinically low-risk HR-positive/HER2-negative disease.

Highlights

• •ER-Predict is a leverages 14 genes to predict disease recurrence in HR-positive/HER2-negative breast cancer.
• •ER-Predict showcased consistent clinical validity across independent validation cohorts.
• •ER-Predict high-risk cases demonstrated poor clinical outcomes across tumor stages and clinically relevant subgroups.
• •ER-Predict high-risk tumors demonstrate cell-cycle mediated endocrine resistance.

Introduction

Hormone receptor (HR)-positive, human epidermal growth factor 2 (HER2)-negative breast cancer represents the most common breast cancer subtype, accounting for ∼70% of all breast cancer diagnoses.¹ Despite showing better cancer-related outcomes compared with other molecular subtypes, a substantial fraction of patients experiences disease recurrence, with nearly 50% of events occurring 5 years after diagnosis.¹ The identification of patients with HR-positive/HER2-negative breast cancer who carry a substantial risk of recurrence has been the cornerstone for treatment modulation in the adjuvant therapeutic space. Traditionally, risk of recurrence was used to tailor adjuvant chemotherapy, and more recently to determine novel adjuvant anticancer agents.² This is particularly relevant considering the recent approval of novel compounds, such as cyclin-dependent kinase 4 and 6 inhibitors (CDK4/6i), which have recently expanded the therapeutic options for patients considered at high clinical risk of relapse.2, 3, 4

Multigene panels that assess the expression of specific subsets of genes are used to adapt the clinical management of early HR-positive/HER2-negative breast cancer according to the risk of recurrence. For instance, several prognostication tools have shown clinical utility to predict the benefit of adjuvant chemotherapies, notably Oncotype DX,^5,6 and multigene panels have been used to guide the adjuvant use of CDK4/6i in node-negative early breast cancer.³ Although most assays gathered prospective validation, their ability to capture the full spectrum of risk, particularly in those at low-clinical risk at presentation or at risk of late recurrences, remains debated.7, 8, 9, 10 Moreover, the emergence of novel therapeutic agents have called into question the enduring clinical relevance of existing multigene tools, particularly given the absence of prospective validation studies specifically designed to assess their predictive value in guiding treatment intensification with CDK4/6i. To date, only retrospective translational biomarker analyses have been conducted, which have not demonstrated significant interaction between multigene tools-based risk-estimation and the efficacy of these newer compounds, laying an unmet clinical need to detect low-clinical risk tumors that might benefit from CDK4/6i.¹¹

Finally, although they are routinely excluded from investigational escalated regimens, a subset of tumors at very low clinical risk may still cause distant relapses. This warrants a refined detection of those cases to guide the use of novel therapeutic compounds without resulting in overtreatment on a large scale. For instance, in the MINDACT clinical trial, although patients with good clinicopathological features did not experience higher rates of cancer recurrence according to the intrinsic genomic risk based on MammaPrint, a risk of recurrence of 8%-10% was observed at 5-years, that persisted beyond this timeline.¹² In the same study, a similar significant risk of recurrence was reported in the group of patients with small, early breast cancers.¹³

Transcriptomics is the most informative -omic to date in providing prognostication validity across solid malignancies.^6,14, 15, 16 Advances in the field of machine learning have shown promise to uncover complex features and patterns that outperform traditional statistical model frameworks leveraged by approved breast cancer multigene assays.^17,18 Limited efforts to implement novel machine learning models in breast cancer have been made. However, these models could potentially leverage the informative role of artificial intelligence to better adapt the clinical management of early breast cancer.

Within this context, we present the development and validation of ER-Predict, an ensemble machine learning assay leveraging a 14-gene expression signature designed to predict relapses in early stage HR-positive/HER2-negative breast cancer. Using a large external cohort, we demonstrated the prognostic validity of ER-Predict to detect tumors at high risk of recurrence, including those in clinically low-risk cancers. We further benchmarked the performance of ER-Predict against established multigene panels, and evaluated the molecular underpinnings of high-risk tumors, demonstrating preliminary evidence about potential therapeutic strategies to tackle resistant luminal-like transcriptomic programs. Altogether, ER-Predict positions itself as a promising tool to guide the use of escalated adjuvant strategies especially among those cancers traditionally considered at low risk of relapse as per current clinical practice.

Methods

Patients and study cohort

To develop ER-Predict, we leveraged a cohort of 2509 early breast cancer samples with available transcriptomic microarray data from cbioportal.org,¹⁹ whose detailed clinical characteristics were previously reported.²⁰ In the development set, long-term outcomes included relapse-free survival (RFS), defined as the time from surgery to the occurrence of any breast cancer-related event, including local, regional, or distant recurrence; and overall survival (OS).

To provide external validation of ER-Predict, we pooled patients from eight different studies, gathering data from 1118 patients with early HR-positive/HER2-negative breast cancer with available distant metastasis-free survival (DMFS) status, defined as the time from surgery to the development of distant metastatic recurrence or death, whichever occurred first. Being DMFS a more accurate measure of long-term outcomes compared with RFS, recapitulating metastatic recurrences usually not amenable to curative treatments,²¹ we hypothesized under confident performances of the model, including low-risk cases affected by local recurrences, would affect the early development of ER-Predict and its internal validation, while retaining an accurate assessment of long-term outcomes in the external validation cohort. In detail, patients included in the cohort reported by Loi et al.²² (accession number GSE6532) were retrieved using the GEOquery R package.²³ The other seven studies24, 25, 26, 27, 28, 29, 30 were retrieved using the MetaGxBreast package,³¹ filtering for HR-positive/HER2-negative cases with available DMFS status. For correcting the batch effect and project ER-Predict classifications to the external test cohort we used a quantile normalization with the development set as reference. To account for batch differences between the studies included in the external test cohort, the custom quantile correction was conducted separately for each study (Supplementary Figure S1, available at https://doi.org/10.1016/j.esmoop.2026.106064). pyComBat³² was used for correcting batch effects for conducting exploratory gene expression analyses.

For assessing ER-Predict in patients affected by metastatic HR-positive/HER2-negative breast cancer, we leveraged the Hartwig Medical Foundation database (License Agreement number HMF-DR-441),³³ filtering for patients treated with CDK4/6i plus endocrine therapy in the metastatic setting and with available progression-free survival status (n = 85), defined as the time from the start of treatment to the first documented disease progression or death from any cause, whichever occurred first.

Definition of a 14-gene signature leveraged by ER-Predict to classify high- and low-risk patients

To gather a subset of genes associated with the risk of relapse, leveraging on the training set (80% of the development set),²⁰ we first computed the median value for duplicated genes resulting from duplicated microarray probes, totaling 20 385 starting genes. Next, we applied a quantile-based filtering to remove low variance genes, followed by a hierarchical clustering step to remove highly correlated features to retain only one gene showing the highest variance within each cluster. Next, we applied a resampling-based feature selection composed of one-hundred bootstrap resamples. Within each iteration, we first optimized a random forest classifier, which was used to compute a five-fold internal averaging of Gini impurity scores. Next, a knee-point detection algorithm was used to select top-ranking genes using the kneedpackage,³⁴ which were then further filtered using Boruta.³⁵ Genes retained by Boruta were normalized and subjected to a univariable ridge-penalized Cox proportional hazards model. The prognostic value of each gene was quantified using a penalized log2 hazard ratio score (log₂^{hazard ratio} × penalized –log₂P value) to emphasize both effect size and statistical confidence. For each iteration, the top 30 genes ranked by the penalized score were retained. Only genes selected in at least one-third of reiterations were kept, resulting in 19 genes from the reiteration process. Finally, we computed SHAP³⁶ scores from a time-dependent non-kernelized support vector machine to identify genes consistently associating with patient predictions in the training set, resulting in a robust 14-gene signature.

Development of ER-Predict and computation of comparative multigene signatures

To develop ER-Predict, we built a composite machine learning model to classify patients at high- versus low-risk of disease recurrence. First, two independent survival models (a linear, non-kernelized support vector machine and a gradient boosting survival forest) were trained using the selected 14-gene signatures. The classification cut-off was established on the training set, wherein an optimal binary cut-off was selected to ensure a minimum 5-year RFS lower confidence interval (CI) above 95% in the predicted low-risk group. In the case where the threshold was not met, the survival requirement was iteratively relaxed. Test samples were classified as low or high risk based on risk score agreement between the two survival models. In cases of discordant predictions, a time-independent classifier, composed of a linear support vector machine and a gradient boosting classifier integrated via a logistic regression meta-learner was employed to result in a final majority voting. Each model was optimized using Bayesian search with five-fold cross-validation using skopt.³⁷ To ensure comparability across models, model outputs were transformed into normalized risk scores using quantile scaling, with each model employing its specific optimal binary cut-off as mentioned above.

To reproduce approved multigene prognostic signatures, we employed the genefu³⁸ package. Oncotype DX (referred as reproduced Oncotype DX - R-Oncotype DX) risk scores were readapted to consider as high-risk cases those showing a score ≥26, as previously defined.³⁹ To compute EndoPredict classifications (referred to as reproduced EndoPredict, R-EndoPredict), starting from microarray normalized data, the endopredict() function was readapted to adhere to the original normalization.⁴⁰ We did not compute the MammaPrint classification using the gene70() function as only a limited number of microarray probes (42/70) were retrieved to calculate confidence scores. The Cohen’s kappa (κ) was used to compare the agreement between ER-Predict and other multigene assays. For developing ER-Predict, we employed the scikit-learn⁴¹ and scikit-survival⁴² Python packages.

Differential gene expression and in vitro drug screening analysis

To investigate biological processes associated with the 14-gene signatures that was retrieved with the abovementioned feature extraction algorithm, we conducted an over-representation analysis using the GSEApy⁴³ package and the Hallmark gene set from the Human Molecular Signatures Database.⁴⁴ We employed limma⁴⁵ for conducting differential gene expression analysis between ER-Predict high- and low-risk cases, whereas the GSEApy⁴³ package was used to compute gene set variation analysis.

For assessing the interaction between risk scores and drug sensitivity, we retrieved breast cancer cell lines with available expression data from DepMap,⁴⁶ to which matched drug sensitivity assays were available from the Genomics of Drug Sensitivity in Cancer database.⁴⁷ Differences in drug sensitivity [inhibitory concentration 50 (IC50)], were compared between predicted high- and low-risk cell lines, whose predictions were computed using the abovementioned quantile-based normalization. The drug sensitivity screening was conducted using approved breast cancer therapeutical compounds. Statistical significance was calculated using the Mann–Whitney U test and effect size determined using the fold-change ratio between the exponentiated median IC50 of high- versus low-risk cell lines. Due to the limited sample size, no family-wise error correction was applied.

Statistical analyses

Continuous variables were reported as median and interquartile range (IQR), and compared between groups using the Mann–Whitney U test. Categorical variables were reported as frequencies and proportions and compared using Fisher’s exact test. Single variable multi-class distributions were compared using the chi-square goodness of fit test.

Median follow-up time was calculated using the reverse Kaplan–Meier procedure using the prodlim⁴⁸ R package. In the time-to-event analysis, the proportional hazard assumption was computed measuring Schoenfeld residuals using the survminer⁴⁹ package. For instances satisfying the proportional hazards assumption, endpoints were calculated using the Kaplan–Meier method with effect size calculated using either unadjusted or adjusted HRs in multivariable models. For cases violating the proportional hazards assumption, we computed the restricted mean survival time lost ratio between high- and low-risk patients [restricted mean time loss ratio (RMTL^{high-risk/low-risk})] using the survRM2⁵⁰ package. For assessing the added multigene panels prognostic information to clinicopathological variables, we computed a log-likelihood ratio difference between the null model (fitted on specified clinicopathological variables) and the nested model including model predictions.

All tests were carried out assuming a two-sided statistical significance with an α level of < 0.05. Family-wise error correction, where specified, was carried out using the Benjamini–Hochberg procedure. Statistical analyses were conducted with Python software version 3.9.21,⁵¹ using the scipy⁵² and statsmodels⁵³ packages. Survival analyses were conducted using the lifelines⁵⁴ Python library and the survival⁴⁹ and survRM2⁵⁰ packages (R software version 4.3.2⁵⁵). Figures were created using matplotlib⁵⁶ and seaborn.⁵⁷

Results

Development of the ER-Predict

A total of 1413 patients affected by HR-positive/HER2-negative cases were retrieved from the development cohort (Figure 1A), including 973 (68.9%) tumors expressing estrogen receptor (ER) and progesterone receptor (PgR) positive samples, 425 (30.0%) ER-positive/PgR-negative, and 15 (1.0%) ER-negative/PgR-positive. Seventy percent (70.8%) and 10.1% of patients had received adjuvant endocrine therapy and chemotherapy, respectively. According to a previously reported prognostic classification,⁵⁸ most patients (n = 809, 57.2%) were classified as moderate- to high-risk cases, with 66.8% (n = 944) of tumors being ≥T2 (AJCC 8th Edition) and with 46.9% (n = 663) patients showing positive lymph nodes at the time of surgery (median positive lymph nodes 5, IQR 1-5).

Figure 1.

ER-Predict development and preliminary internal validation. (A) A large cohort of 1433 patients affected by HR-positive/HER2-negative breast cancer was leveraged to develop ER-Predict. In detail, 80% of the development cohort was included for gathering a prognostic 14-gene signatures, on which an ensemble machine learning model was built to classify patients according to the risk of recurrence. (B) Model performance on the training set, wherein ER-Predict high-risk cases showcased a significantly reduced RFS time. (C) Clinicopathological variables according to ER-Predict classifications on the test set (20% of the development set). In detail, distribution of progesterone receptors (PgR) status, tumor grade (Notthingam Histologic Grade) and the Nottingham prognostic index between high- and low-risk ER-Predict cases (∗denotes statistically significant difference between risk groups). (D) Performance of ER-Predict in the internal validation set, wherein ER-Predict retained robust performance to predict high-risk cases with significantly reduced RFS time. BC, breast cancer; CI, confidence interval; HER2−, human epidermal growth factor 2-negative; HR+, hormone receptor-positive; RFS, relapse-free survival; RMTL, restricted mean time loss ratio.

For developing ER-Predict, we employed 80% of the cohort to result in a training set of 1128 cases, wherein we leveraged gene expression data to retrieve a prognostic 14-gene signature (see methods). Finaly, we built ER-Predict, an ensemble machine learning model designed to classify patients according to the risk of relapse (Supplementary Figure S2, available at https://doi.org/10.1016/j.esmoop.2026.106064). In the training set, 47.8% of cases were classified at high risk of relapse. Although no difference in age was observed between the two risk groups, high-risk tumors showcased a higher proportion of high-grade tumors (52.2% versus 22.1%, chi-square goodness of fit P < 0.001), tumor size [25 mm (IQR 25-32 mm) versus 20 mm (IQR 16-26 mm), Mann–Whitney-U test, P < 0.001], and positive lymph nodes at the time of surgery (51.1% versus 43.8%, Fisher’s exact test P = 0.01).

Altogether, in the training set, ER-Predict high-risk patients demonstrated a significantly reduced RFS time, resulting in an RMTL^{high-risk/low-risk} of 3.84 (95% CI 2.79-5.26, Wald test P < 0.0001) at 60 months and RMTL^{high-risk/low-risk} of 3.09 (95% CI 2.47-3.87, P < 0.0001) at 120 months, respectively (Schoenfeld residuals P = 0.001) (Figure 1B). This means that, on average, high-risk patients lost approximately three to four times more recurrence-free months over a 5- to 10-year period compared with their low-risk counterparts.

Internal validation of ER-Predict on the development set

For seeking internal validation of ER-Predict, we leveraged the hold-out development set that included 283 patients not previously included in the development of ER-Predict. Among them, the classification of high- and low-risk cases was balanced (n = 137 high-risk cases, 48.3%), closely mirroring the distribution observed in the training set. A higher proportion of high-risk cases received adjuvant chemotherapy (12.4% versus 4.7%, Fisher’s exact P = 0.02), consisted of high-grade tumors (61.0% versus 19.1%, chi-square goodness of fit test P < 0.001), and had a larger tumor size (Mann–Whitney-U test P = 0.005), whereas no difference was observed in the proportion of cases with positive lymph nodes at surgery (Fisher’s exact P = 0.81) (Figure 1C, Supplementary Figure S3, available at https://doi.org/10.1016/j.esmoop.2026.106064).

In the internal validation cohort, ER-Predict showed consistent clinical validity to predict long-term outcomes (Figure 1D), with high-risk cases demonstrating a reduced 60-month RFS time (71.2%, 95% CI 63.7% to 79.9% versus 87.7%, 95% CI 82.4% to 93.4%, RMTL^{high-risk/low-risk} 2.66, 95% CI 1.43-4.97, Wald test P = 0.002), and 120-month RFS (53.0%, 95% CI 48.7% to 66.7% versus 77.2%, 95% CI 70.2% to 85.0%, RMTL^{high-risk/low-risk} 2.17, 95% CI 1.41-3.35, Wald test P < 0.001, Schoenfeld residuals P < 0.001).

Moreover, ER-Predict associated with OS, with high-risk patients showing an inferior median OS (148 months, 95% CI 112-188 months versus 171 months, 95% CI 152-215 months, HR 1.32, 95% CI 0.98-1.79, P = 0.007), resulting in a 120-months RMTL^{high-risk/low-risk} of 1.79, 95% CI 1.20-2.68, Wald test P = 0.004, Schoenfeld residuals P = 0.05) (Supplementary Figure S4, available at https://doi.org/10.1016/j.esmoop.2026.106064).

External validation of ER-Predict to predict cases at high-risk of recurrence irrespective of clinical risk

Following internal validation, for seeking external validation of ER-Predict, we next retrieved a large cohort of patients affected by early HR-positive/HER2-negative breast cancer from eight different studies (n = 1118 patients) (Figure 2A). In the pooled-external cohort, treatment data were available for 92.9% of cases. Only 30.7% of cases received adjuvant endocrine therapy (38.2% with chemotherapy, 11.7% overall), 32.3% received chemotherapy alone, whereas 36.9% did not receive systemic treatment. After a median follow-up of 75.9 months (IQR 42.2-110.9 months), 229 events were recorded (20.4%), resulting in a 60-month DMFS of 79.5% (95% CI 76.9% to 82.1%) and 120-month DMFS of 71.2% (95% CI 67.8% to 74.8%).

Figure 2.

External validation of ER-Predict. (A) Pooled-external cohort including eight different studies (n = 1128 cases) used for validating ER-Predict, along with the treatment status. (B) Performance of ER-Predict in the overall study cohort, wherein ER-Predict high-risk cases showed a significantly reduced DMFS compared with low-risk cases (HR 2.03, 95% CI 1.57-2.63, P < 0.0001). (C) ER-Predict showed consistent clinical validity among cases treated with ET alone (HR of high-risk versus low-risk: 2.96, 95% CI 1.65-5.03, P = 0.0002). (D) Multivariable model including ER-Predict and clinicopathological variables. Notably, ER-Predict showed consistent performance even when adjusted for potential confounding factors (HR 1.66, 95% CI 1.18-2.35, P = 0.003). Values on the plot represent HR (95% CI). (E) Association between ER-Predict classifications and DMFS among patients not receiving chemotherapy, wherein ER-Predict showcased significant clinical validity. (F) In contrast, no statistically significant difference was observed among ER-Predict classes in those treated with adjuvant chemotherapy (HR 1.23, 95% CI 0.76-1.96, P = 0.38), with high- but not low-risk cases showing a statistically significant benefit from adjuvant chemotherapy (HR 0.45, 95% CI 0.23-0.90, P = 0.02). CI, confidence interval; DMFS, distant metastasis-free survival; ET, endocrine therapy; HR, hazard ratio.

Overall, ER-Predict high-risk cases (49.4%) showcased a significantly reduced DMFS compared with low-risk cases (HR 2.03, 95% CI 1.57-2.63, P < 0.0001), with a 60-month DMFS rate of 72.7% (95% CI 68.7% to 77.0%) versus 86.0% (95% CI 82.9% to 89.2%), and a 120-month DMFS of 61.3% (95% CI 56.6% to 67.7%) versus 79.8% (95% CI 75.6% to 84.2%) (Figure 2B), which was consistent among those treated with endocrine therapy (Figure 2C, Supplementary Figure S5, available at https://doi.org/10.1016/j.esmoop.2026.106064). In the multivariable model that included primary tumor size, nodal status, and tumor grading, ER-Predict classifications still showed robust independent association with DMFS (HR 1.79, 95% CI 1.36-2.36, P < 0.001) (Figure 2D), thus suggesting ER-Predict alone could identify high-risk cases irrespective of clinicopathological variables.

To validate our hypothesis, we sought to determine whether ER-Predict could provide clinical validity among meaningful clinical subgroups. Accordingly, we pooled node-negative cases (n = 647, n = 302 high-risk, n = 345 low-risk), wherein ER-Predict high-risk cases still demonstrated a significantly reduced DMFS compared with low-risk tumors (HR 2.38, 95% CI 1.69-3.35, P < 0.001) (Supplementary Figure S6A, available at https://doi.org/10.1016/j.esmoop.2026.106064), which was consistent among those treated with endocrine therapy (ET)-alone (n = 90, HR 4.54, 95% CI 1.44-14.3, P = 0.009) and in patients receiving ET ± chemotherapy (n = 124, HR 2.76, 95% CI 1.14-6.67, P = 0.02) (Supplementary Figures S6B and C, available at https://doi.org/10.1016/j.esmoop.2026.106064). Considering the association between tumor grade and DMFS in the multivariable model, we further investigated whether ER-Predict could identify non-high-grade tumors (G1/G2) at high risk of relapse. Notably, among G1/G2 tumors (n = 634, n = 236 high-risk, n = 395 low-risk), high-risk cases consistently demonstrated a significantly reduced DMFS compared with low-risk cases (HR 2.17, 95% CI 1.48-3.20, P < 0.001) (Supplementary Figure S6D, available at https://doi.org/10.1016/j.esmoop.2026.106064).

Finally, we hypothesized ER-Predict could provide clinical validity particularly among patients not treated with adjuvant chemotherapy, which would rationalize the use of ER-Predict to detect early HR-positive/HER2-negative cancer at high-risk of relapse for escalated treatment regimens. Strikingly, although ER-Predict demonstrated strong clinical validity among cases not receiving adjuvant chemotherapy alone (n = 581, HR 2.59, 95% CI 1.88-3.58, P < 0.0001) (Figure 2E), including when corrected for clinicopathological variables (HR 2.21, 95% CI 1.49-3.28, P < 0.0001), ER-Predict classifications did not associate with DMFS among those treated with adjuvant chemotherapy (n = 458, HR 1.23, 95% CI 0.76-1.96, P = 0.38) (Figure 2F). We observed a statistically significant interaction between ER-Predict risk classification and chemotherapy treatment (P_interaction = 0.01), with high- but not low-risk patients benefiting from the addition of adjuvant chemotherapy to ET (HR 0.45, 95% CI 0.23-0.90, P = 0.02).

Benchmarking ER-Predict to clinically approved prognostication tools

Considering the availability of several prognostication tools for HR-positive/HER2-negative early breast cancer, we compared ER-Predict to clinically approved assays leveraging our large external validation cohort (n = 1118 cases), wherein we recomputed prognostic signatures using a previously published pipeline.³⁸ Most ER-Predict low-risk cases consisted of PAM50 Luminal A tumors (58.7%), while Luminal B tumors represented the most common intrinsic subtype in high-risk cases (52.4%), followed by HER2-enriched (18.6%) and basal-like tumors (16.9%) (Figure 3A). Most cases in the study cohort were defined at high risk of relapse by R-Oncotype DX (70.5%) and R-EndoPredict (73.8%). Notably, only a moderate agreement was observed between ER-Predict and R-Oncotype DX (κ = 40.5) and ER-Predict and R-EndoPredict (κ = 32.9) (Figure 3B and C).

Figure 3.

Benchmarking ER-Predict to approved multigene panels. (A) PAM50 subtypes according to ER-Predict risk classes. In detail, low-risk cases were mostly enriched for Luminal A tumors, while Luminal B represented the most common subtype among high-risk tumors. (B) Distribution of EndoPredict (left) and Oncotype DX (right) scores among ER-Predict high (red) and low-risk (green) cases. (C) Correlation between EndoPredict and Oncotype DX risk scores, with color denoting the respective ER-Predict classification (low-risk in green, high-risk in red). (D) Multivariable model including ER-Predict, EndoPredict, and Oncotype DX, wherein only ER-Predict maintained robust association with DMFS (HR 1.66, 95% CI 1.24-2.22, P = 0.0006) Values on the plot represent HR (95% CI). (E) 10-year DMFS according to ER-Predict, EndoPredict, and Oncotype DX risk classes in the overall study cohort, and in those treated with ET alone (F). (G) Cumulative incidence of late recurrences (after 5-years), where only ER-Predict (HR 1.88, 95% CI 1.09-3.25, P = 0.03) retained statistically significant association with DMFS. CI, confidence interval; DMFS, distant metastasis-free survival; HR, hazard ratio.

In the univariable model including ER-Predict, R-Oncotype DX, R-EndoPredict, PAM50 intrinsic subtypes, both EndoPredict (HR 2.16, 95% CI 1.51-3.07, P < 0.0001) and Oncotype DX (HR 1.90, 95% CI 1.39-2.59, P < 0.0001) showed a statistically significant association with DMFS, as observed with PAM50 intrinsic subtypes (Supplementary Figure S7, available at https://doi.org/10.1016/j.esmoop.2026.106064). However, in the multivariable model including ER-Predict, R-Oncotype DX and R-EndoPredict, only ER-Predict strongly associated with DMFS (HR 1.69, 95% CI 1.26-2.26, P < 0.001) (Figure 3D), which was consistent once adjusting for PAM50 intrinsic subtypes (Supplementary Figure S7, available at https://doi.org/10.1016/j.esmoop.2026.106064). In contrast, R-Oncotype DX (HR 1.16, 95% CI 0.79-1.70, P = 0.4) and R-EndoPredict (HR 1.58, 95% CI 1.04-2.40, P = 0.03) retained limited association with DMFS, despite showing robust statistical association with DMFS once correcting for clinicopathological variables (Supplementary Figure S8, available at https://doi.org/10.1016/j.esmoop.2026.106064). Despite ER-Predict classifying a higher proportion of cases at low-risk of relapse, we observed a similar 10-year recurrence rate in low-risk cases between the three assays in the overall study population by ER-Predict (low- versus high-risk: 83.3%, 95% CI 75.3% to 91.9% versus 58.9%, 95% CI 48.1% to 72.2%), R-Oncotype DX (81.0%, 95% CI 70.9% to 92.7% versus 67.3%, 95% CI 58.5% to 77.4%) and R-EndoPredict (86.0%, 95% CI 76.6% to 96.5% versus 65.8%, 95% CI 57.1% to 75.9%) (Figure 3E), which was consistent among those treated with ET alone (Figure 3F).

Next, considering the recent approval of adjuvant CDK4/6i, including for patients with stage T2/N0 tumors at high genomic risk, we investigated the association between DMFS and multigene panels risks among those treated with ET ± chemotherapy (n = 122). Herein, only ER-Predict (ΔLR χ² = 3.42, P = 0.063) and R-Oncotype DX (ΔLR χ² = 3.56, P = 0.059) showed a trend for additional prognostic value in T2/N0 tumors compared with T1/N0 cases. Moreover, among stage I tumors (T1/N0) treated with ET alone (n = 48), which are routinely excluded from escalated treatment regimens, only ER-Predict (HR 4.95, P = 0.19) demonstrated a trend for strong association with DMFS (R-Oncotype DX HR 1.41, P = 0.77; EndoPredict HR 1.36, P = 0.32), thus suggesting a potential role of ER-Predict to identify tumors at very low clinical risk potentially benefiting from escalated adjuvant regimens.

Finally, we investigated whether ER-Predict could be specific to early versus late recurrences. Although several prognostication tools have shown clinical validity to detect early events, previous reports highlighted their limited role to detect late recurrences beyond 5 years.7, 8, 9, 10^,13 Notably, only ER-Predict (ΔLR χ² = 5.68, P = 0.01) showed additional prognostic information to detect late recurrences when added to clinicopathological parameters (tumor grade, nodal status, and tumor size; R-Oncotype DX: ΔLR χ² = 1.69, P = 0.19; R-EndoPredict: ΔLR χ² = 0.47, P = 0.47), which was consistent for patients treated with ET with or without chemotherapy (ER-Predict: ΔLR χ² = 4.70, P = 0.02; R-Oncotype DX: P = 0.7; R-EndoPredict: P = 0.9). Moreover, in the univariable analysis, only ER-Predict consistently demonstrated clinical validity for detecting late recurrences both among the overall study population (n = 546, HR 1.88, 95% CI 1.09-3.25, P = 0.03) and among those treated with ET alone (n = 128, HR 2.53, 95% CI 0.96-6.68, P = 0.056) (Figure 3G).

ER-Predict is specific to predict luminal-like transcriptomic profiles of drug resistance

Following validation, we sought to investigate the molecular underpinnings of ER-Predict classifications. First, to assess ER-Predict specificity, we retrieved the subgroup of patients from the development set affected by triple-negative (TNBC, n = 320) and HER2-positive breast cancer (n = 247). Notably, model predictions negatively projected among TNBC and HER2-positive cases, with most cases classified at high-risk of relapse. In contrast, we hypothesized ER-Predict could retain clinical validity among HR-positive/HER2-negative tumors treated with ET in the metastatic setting, recapitulating its ability to capture the endocrine-dependency and its biological nuances. We leveraged a small cohort (n = 85) of metastatic HR-positive/HER2-negative breast cancer cases treated with CDK4/6i + ET with available baseline transcriptomic data. Of note, ER-Predict high-risk cases still showed a trend for reduced PFS (HR 1.62, 95% CI 0.97-2.73, P = 0.067), suggesting ER-Predict to specifically detect luminal-like transcriptomic profiles of drug resistance (Figure 4A).

Figure 4.

Exploratory gene expression analysis of molecular determinants of ER-Predict classifications. (A) ER-Predict classifications in metastatic HR-positive/HER2-negative breast cancer treated with CDK4/6i, where ER-Predict classifications still retained a numerical association with PFS, suggesting specificity toward luminal-like transcriptomic profiles of endocrine resistance. (B) Gene ontology over-representation analysis of the 14-gene ER-Predict signature, showing an enrichment of genes involved in E2F targets (OR 16.6, q-value = 0.04) and proteins involved in the late cell-cycle transition (G2/M, OR 40.3, q-value <0.0001). (C) Differential gene expression analysis between ER-Predict low- and high-risk cases carried out with limma. Notably, genes showing the strongest effect size and statistical significance were genes involved in the cell-cycle regulation, particularly E2F targets and proteins functioning in the late transition phase from the G2 phase to mitosis. (D) Log₂ fold-change of genes included in the E2F Hallmark gene set, G2/M, in both G/M-E2F or other sets between high- and low-risk tumors, with genes present in the former three pathways showing consistent enrichment among high-risk tumors. (E) Gene set variation analysis between high- and low-risk tumors, with high-risk tumors showing a significant activation of the E2F and G2/M checkpoint pathways. (F) Retinoblastoma-1 (Rb1) protein expression between ER-Predict high- and low-risk cases. Notably, despite high-risk cases demonstrating higher expression of E2F targets, they showed a retained expression of Rb1 in both the development and external validation set. (G) In vitro drug sensitivity analysis among HR-positive/HER2-negative breast cancer cell lines (n = 5) wherein high-risk cell lines did not demonstrate a trend for higher resistance to palbociclib and ribociclib. (H) Correlation between ER-Predict scores (sum of the distance from each cut-off of the three models normalized using a minimum-maximum scaling to range between 0 and 1) and IC50 of palbociclib and ribociclib among breast cancer cell lines irrespective of receptors status (n = 44). CI, confidence interval; HER2−, human epidermal growth factor 2-negative; HR+, hormone receptor-positive; IC50, inhibitory concentration 50; UV, ultraviolet.

We next sought to investigate the molecular underpinnings of the 14-gene signatures incorporated in ER-Predict. We observed our 14-gene signature to be highly enriched for genes involved in the cell-cycle regulation [G2-M checkpoint: odds ratio (OR) 40.3, q-value < 0.0001, E2F targets: OR 16.6, q-value = 0.04] (Figure 4B). To explore the biological basis of ER-Predict risk stratification, we investigated genes differentially expressed between high- and low-risk groups (Figure 4C). Remarkably, E2F targets and proteins involved in the cell-cycle transition from the G2 phase to the mitotic stage were revealed to be highly enriched among ER-Predict high-risk cases (Figure 4D), which translated in a consistent higher activity of the latter pathways both in the external validation cohort and in the development set (Figure 4E). Although cell-cycle dysregulation promotes endocrine resistance, CDK4/6i act by reducing the activity of E2F targets depending on the presence of the retinoblastoma 1 (Rb1) regulator.⁵⁹ Despite presenting an enrichment of E2F targets, ER-Predict high-risk cases still showed a preserved Rb1 expression (Mann–Whitney U test P = 0.7) (Figure 4F), thus suggesting the potential actionability of ER-Predict high-risk transcriptomic programs by means of CDK4/6i.

To further test this hypothesis, we retrieved HR-positive/HER2-negative breast cancer cell lines with available transcriptomic profiles and matched drug sensitivity drug screening data (n = 5 cell lines and 276 therapeutic compounds). As hypothesized, ER-Predict high-risk cell lines associated with numerically higher IC50 for tamoxifen (fold-change 1.82, Mann–Whitney U test P = 0.24), fulvestrant (fold-change 1.95, Mann–Whitney U test P = 0.22), and to relevant breast cancer therapeutic agents, such as olaparib, epirubicin, and cyclophosphamide (Figure 4G). In contrast, we observed no trend for different sensitivity to palbociclib (fold-change 1.06, Mann–Whitney U test P = 0.40) and ribociclib (fold-change 0.77, Mann–Whitney U test P = 0.75), which was consistent once all breast cancer cell lines were pooled together (n = 44) (Spearman correlation between risk scores and (i) palbociclib: rho = −0.26, P = 0.07; (ii) ribociclib: rho = −0.34, P = 0.02) (Figure 4H). Altogether, cognizant of the small sample size, these results suggest ER-Predict high-risk tumors, despite revealing relative resistance to endocrine agents, may not show marked resistance to CDK4/6i, warranting further investigation for the potential role of ER-Predict to detect high-risk cases particularly sensitive to CDK4/6i.

Discussion

In the present work we presented ER-Predict, an ensemble machine learning model leveraging a 14-gene signature to categorize early HR-positive/HER2-negative breast cancer according to the risk of recurrence. Notably, ER-Predict demonstrated robust clinical validity among all clinically relevant subgroups, with high-risk breast cancer cases as per ER-Predict consistently showing poorer prognosis irrespective of clinicopathological variables. In addition, ER-Predict successfully showed robust performance to predict both early and late breast cancer recurrences, consistently outperforming multigene prognostic signatures currently implemented in clinical practice.

The earlier development of breast cancer multigene assays was aimed to detect low-risk patients who could be spared from the use of adjuvant chemotherapies. In this regard, several prognostication tools have been shown to predict the benefit from chemotherapy, with their use currently supported by international guidelines.^12,39,60 Likewise, in the present study, ER-Predict not only demonstrated a robust prognostic value irrespective of clinicopathological variables, but showcased preliminary predictive ability for anticipated benefits resulting from the addition of adjuvant chemotherapy. We observed an interaction between ER-Predict and the receipt of adjuvant chemotherapy, with high-risk cases benefiting the most. In addition, in a high clinical risk population (53% with positive lymph nodes, 33% G3 tumors, 55% with a tumor size between 20 to 49 mm), ER-Predict low-risk cases receiving endocrine treatment alone still showed optimal long-term outcomes, which was not divergent from those of recomputed prognostication tools.

Compared with other subtypes, HR-positive/HER2-negative breast cancer is burdened by high rates of late recurrences, requiring a prognostication tool to consistently show acceptable performances to detect late relapses. In this regard, although all approved multigene panels have shown accuracy in predicting early events, they previously demonstrated inconsistent robustness in detecting late events,^7,8,61 and none has proved clinical utility in this space. Herein, ER-Predict was built leveraging a cohort of patients with long follow-up (median of ∼16 years), which translated in the consistency of ER-Predict to detect late recurrences both in the overall study cohort and in those treated with ET-based regimens. Of note, among those tested, ER-Predict was the only multigene panel showing additional prognostic information to clinicopathological parameters, thus providing preliminary data about its role as a clinical assay to detect patients at risk of late relapses.

Notably, the evolving landscape of therapeutic options in early HR-positive/HER2-negative breast cancer is challenging the informative role of multigene panels to predict long-term outcomes. It is currently unknown how genomic predictors could interact with the use of newly approved compounds, such as CDK4/6i. At the same time, there is an urgent need for additional biomarkers to maximize the clinical benefits resulting from escalated adjuvant regimens. De facto, such a need with CDK4/6i is pervasive, regarding all settings of current clinical use, including in the metastatic disease. In fact, although the use of CDK4/6i could yield higher curative rates for a subgroup of patients, they result in additional health-related and financial toxicities on the large scale.⁶² Notably, in the phase III NATALEE study, investigating the role of 3-years of adjuvant ribociclib in HR-positive/HER2-negative breast cancer, multigene panels have guided patients’ eligibility for stage IIA tumors (T2/N0) besides tumor grade and the proliferative index. Regardless, the addition of 3-years of adjuvant ribociclib was associated with uncertain benefits among node-negative tumors,³ warranting a better identification of biomarkers capable to detect low-clinical risk tumors at high risk of relapse. Of note, the gene signatures used to identify patients at high risk were not per se predictive of the benefit with the CDK4/6i, as reported in a biomarker analysis.¹¹

In this regard, ER-Predict showcased robust performance to identify node-negative tumors at high risk of relapse, both in the overall study cohort and among those treated with ET with or without chemotherapy. At the same time, ER-Predict was demonstrated to predict very low clinical risk tumors still at high-risk of recurrence, which is particularly relevant considering their routine exclusion from escalated adjuvant regimens and from multigene prognostic assay testing. Indeed, ER-Predict, among those tested, was revealed to be the only multigene predictor that strongly associated with long-term outcomes once clinicopathological variables were adjusted for, and showcased prognostic ability among stage I tumors both in the overall cohort and in those treated with ET alone.

Lastly, we investigated the molecular mechanisms underpinning ER-Predict high-risk cases, wherein we showed a strong specificity toward genes involved in the cell-cycle regulation. We observed an enrichment in our multigene signature for E2F targets and proteins involved in the late cell-cycle transition. Importantly, although cell-cycle dysregulations have been shown to promote endocrine resistance, their actionability by means of CDK4/6i has yielded unprecedented positive outcomes in metastatic, and, recently, in early HR-positive/HER2-negative breast cancer.^3,63, 64, 65 Although loss of RB1 represents a mechanism promoting resistance to CDK4/6i,^59,66 its preservation together with high expression of E2F targets has shown to be predictive of CDK4/6i benefits,67, 68, 69 which molecularly act to prevent the phosphorylation of Rb1 to retain E2F factors from promoting the cell-cycle transition. Notably, we showed ER-Predict high-risk tumors to retain the expression of RB1, positioning ER-Predict as a potential tool to detect high-risk cases particularly susceptible to CDK4/6i, thus warranting further validation.

It must be noted our study presents relevant limitations that should be considered. Firstly, our external validation cohort comprised variegated treatment regimens, with a relevant proportion of cases not receiving any systemic treatment. Moreover, detailed annotation about specific endocrine agents and chemotherapy regimens was not available for the employed datasets. Although we showed the consistency of ER-Predict across relevant subgroups, the relatively small sample size warrants further validation. Secondly, we could not assess whether ER-Predict classifications could statistically interact with the menopausal status, since it was not available for most cases in the external validation cohort. In this regard, being cognizant that menopausal status is a key modifier of ET outcomes, future validation studies should stratify the performance of ER-Predict according to menopausal status. Thirdly, although we provided preliminary evidence regarding the benefit from chemotherapy among high-risk cases, these findings are exploratory and hypothesis-generating, and thus should be equally validated among a larger cohort of patients treated with ET + chemotherapy. Fourthly, although we investigated the molecular determinants of ER-Predict classifications, our study findings should warrant further validation in preclinical models to ascertain a possible informative role of ER-Predict in detecting tumors at high-risk of recurrence yet showing therapeutic vulnerability to CDK4/6i. Fifthly, although we committed on benchmarking ER-Predict to approved multigene panels, these were recomputed using a published pipeline that has not been formally validated, and thus demands further comparisons between ER-Predict and the original approved multigene panels. Sixthly, we could not compare ER-Predict with established clinical risk estimators. The number of lymph nodes required for calculating clinical risk estimators,70, 71, 72 was not available in the external validation cohorts. Regardless, we confirmed the clinical validity of ER-Predict both in multivariable models including clinical covariates, and in models leveraging clinical parameters that demonstrated statistically improved performances once considering ER-Predict predictions. Finally, although ER-Predict showed robust predictivity for late recurrences, our external validation cohort did not comprise patients treated with prolonged ET, and thus we could not assess how prolonged ET regimens may interact with the performance of ER-Predict to predict later recurrences.

Altogether, despite the abovementioned limitations, we developed and conducted the first retrospective validation of ER-Predict, a machine learning assay designed to stratify early stage HR-positive/HER2-negative breast cancer according to the risk of recurrence. Our findings support the need for further validation to define the clinical utility of ER-Predict and to assess its potential as a decision-support tool for guiding treatment escalation. Additionally, we characterized the molecular underpinnings of ER-Predict high-risk cases, revealing luminal-like transcriptomic programs associated with disease recurrence, which may inform future therapeutic strategies targeting ER-Predict high-risk cases. Finally, although ER-Predict classifications demonstrated prognostic value independent of clinical risk, further studies are warranted to evaluate whether integrating ER-Predict with traditional clinicopathological parameters could enhance its sensitivity and specificity, ultimately enabling more precise tailoring of adjuvant treatment in early stage HR-positive/HER2-negative breast cancer.