Abstract Background: Invasive ductal carcinoma is the most common subtype of breast cancer, with endocrine therapy the standard treatment for ER+ patients. However, certain clinical and genomic risk factors support adding chemotherapy. Among young women under 50, treatment decisions must balance survival benefit with fertility preservation and overtreatment avoidance. Similarly, patients with low Nottingham Prognostic Index (NPI) scores are generally spared chemotherapy, yet a subset may still benefit from combination treatment. In this study, we apply a two-stage machine learning framework to identify (1) a clinically meaningful subgroup of young ER+/HER2− patients likely to benefit from combined chemo-endocrine therapy despite fertility concerns, and (2) genomic subgroups associated with chemotherapy benefit among patients with low-risk NPI profiles. Methods: We analyzed a longitudinal cohort of 423 IDC patients who received adjuvant treatment post-surgery, with 5-year overall survival (OS) and recurrence-free survival (RFS) as outcomes. A two-stage Virtual Twins (VT) ML framework was applied to identify subgroups benefiting from combination therapy (hormonal therapy + chemotherapy ± radiation) versus hormonal therapy alone, focusing on patients 50 years and those with low NPI. In Stage 1, multiple classifiers - random forest, multilayer perceptron (MLP), gradient boosting (XGBoost), Super Learner, and Bayesian additive regression trees (BART)- were compared for overall survival (OS) and recurrence-free survival (RFS) prediction. The best-performing model was used to estimate individual treatment effects (ITEs) via counterfactuals. In Stage 2, regression trees were trained on the ITEs using the top 10 predictive features to identify subgroups with differing treatment sensitivity. Results: OS and RFS were modeled using 19 clinical and multi-omics features (transcriptomics, CNV, mutations, methylation) selected from high-dimensional inputs based on biological clustering in Stage 1 of VT, enabling ITE-based subgroup identification via regression trees in Stage 2. Firstly, in ER+/HER2− patients under age 50, those with the greatest survival benefit from combination therapy had high NPI (5.0) and harbored amplifications at 11q13/14 (CCND1, PAK1, RSF1, EMSY), 8p12, 8q, and 20q—indicative of proliferative and endocrine-resistant biology. These tumors also exhibited TRG/TRA deletions and high CD8+ infiltration. In contrast, younger patients with genomically quiet tumors-defined by isolated 8p12 amplification, 1q gain, 16q loss, or modest 8q/20q gains-and Luminal A or Claudin-low subtypes showed minimal benefit from combination therapy. Secondly, in patients with low NPI (4.0), those with the greatest benefit from combination therapy were young (55) or middle-aged (55-70) and exhibited 17q23 and 20q amplifications, high CD8+ infiltration, and complex genomic instability, including 5q loss, 8q gain, and 10p/12p gains involving mitotic and DNA repair genes such as TTK, AURKB, FOXM1, and RAD51AP1. A second subgroup with modest benefit included patients under 70 years with amplifications at 11q13/14, 8q12, and copy number switches including 16p gain/16q loss and concurrent 8q and 20q gains. Conversely, Older patients (70) with 1q gain and 16q loss showed limited benefit, consistent with quiet Luminal A-like tumors. Conclusion: Virtual twin modeling identified ER+/HER2− IDC subgroups- despite clinical low-risk or fertility -conscious status-- that derive meaningful benefit from adjuvant combination therapy when harboring high-risk genomic features. These findings support genomics-informed escalation in patients typically spared chemotherapy, highlighting the value of precision oncology in tailoring adjuvant care. Citation Format: W. Guo, C. Tatsuoka. Ai-driven chemotherapy decision support via genomic subgroup identification in fertility-conscious and low-risk er+/her2− invasive ductal breast carcinoma abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS1-11-13.
Guo et al. (Tue,) studied this question.