Multi-state modeling associates proliferation genes with early estrogen receptor-positive breast cancer recurrence and survival

Technical Summary: Multi-state Modeling of Proliferation Genes in ER+ Breast Cancer Recurrence

Problem Statement
Estrogen receptor-positive (ER+) breast cancer accounts for approximately 70% of breast cancer cases and the majority of related deaths. While the five-year survival rate for localized disease is high, it drops significantly upon metastatic progression, which typically arises from recurrence. A critical clinical challenge is the molecular heterogeneity of ER+ breast cancer, which manifests as distinct risks for early recurrence (within 5 years of surgery) versus late recurrence (after 5 years). Existing commercially available multigene assays (e.g., Oncotype DX, PAM50/Prosigna, EndoPredict) provide composite risk scores but utilize fixed proprietary algorithms that aggregate gene signals. This aggregation limits the interpretability of individual gene contributions and obscures the differential biological mechanisms driving early versus late recurrence. This study aims to bridge the gap between clinically validated prognostic assays and biologically interpretable recurrence mechanisms by identifying baseline gene expression programs specifically associated with the timing of recurrence and post-recurrence survival.

Methodology
The study utilized a retrospective cohort of 79 high-risk ER+ (HER2-negative) breast cancer patients with surgical tumor specimens collected between 2012 and 2016.

• Data Generation: Tumor samples were profiled using a custom NanoString nCounter panel comprising 145 genes (128 endogenous, 17 housekeeping).
• Cohort Definition: Patients were stratified by recurrence status: no recurrence (n=56), early recurrence (n=14, defined as <5 years post-surgery), and late recurrence (n=9, defined as ≥5 years). All breast cancer-specific deaths occurred in the early recurrence group.
• Statistical Framework: The core analytical approach employed a semi-Markov multi-state modeling framework. Unlike standard survival analysis that treats recurrence as a single endpoint, this model explicitly defined states (no recurrence, early recurrence, late recurrence, death) and modeled transitions between them.
  • Baseline Transitions: Modeled the transition from "no recurrence" to "early recurrence," "late recurrence," or "death without recurrence" as a function of time since surgery.
  • Post-Recurrence Transitions: Modeled the transition from recurrence states to breast cancer-specific death or other-cause death as a function of time since recurrence.
• Gene-Level Analysis: Differential gene expression was assessed using negative binomial regression. Transition-specific risks were estimated using semi-Markov multi-state Cox proportional hazards models, adjusted for age, cancer stage, and tumor size. Continuous gene expressions were standardized, and hazard ratios (HR) were calculated per one standard deviation (SD) increase.
• Robustness and Validation: To address sparse events in certain transitions (e.g., other-cause death), ridge-penalized Cox models were applied. Sensitivity analyses were conducted by redefining early/late recurrence thresholds at 3 and 7 years. Pathway coherence was evaluated using Over-Representation Analysis (ORA) and Gene Set Enrichment Analysis (GSEA).

Key Results

• Early Recurrence Drivers: In the transition from no recurrence to early recurrence, higher baseline expression of a specific set of proliferation- and cell-cycle-related genes was consistently associated with increased risk. Key genes included UBE2C, EZH2, CCNB1, PTTG1, and MKI67.
  • The hazard ratios for a 1-SD increase in these genes ranged from 1.91 to 2.14 (all p < 0.05).
  • Conversely, higher expression of SNAI2 was associated with a protective effect (HR: 0.43; 95% CI: 0.18–0.99).
  • Pathway analysis confirmed that these genes clustered in coherent biological processes related to the cell cycle, mitotic spindle assembly, chromosome segregation, and nuclear division.
• Late Recurrence Drivers: In contrast, the transition to late recurrence was associated with a distinct set of genes, primarily related to signaling pathways. Significant genes included IGF1R, LEF1, and WNT7B (HR range: 1.52–3.40).
• Post-Recurrence Survival: Among patients who experienced early recurrence, higher expression of CD24, TMEM45B, and GRB7 was associated with an increased risk of breast cancer-specific death, while STAT1 and CREBBP were protective.
• Consistency: The findings regarding early recurrence genes showed strong concordance across differential expression analyses, transition-specific models, and sensitivity analyses using alternative time cutoffs. The identified gene set overlapped significantly with proliferation modules embedded in established commercial assays (e.g., Oncotype DX, PAM50, EndoPredict).

Significance and Claims
The authors claim that this study successfully identifies a potential proliferative gene assay specifically associated with early recurrence in ER+ breast cancer, while demonstrating that late recurrence is linked to distinct signaling pathways.

• Biological Resolution: By using a multi-state modeling framework, the study disentangles temporal patterns of molecular association that are obscured in composite risk scores. It provides mechanistic insights suggesting that early relapse is driven by heightened proliferative activity, whereas late relapse is less dependent on baseline proliferation and more influenced by signaling pathways and microenvironmental cues.
• Biomarker Prioritization: The study prioritizes a compact set of candidate molecular biomarkers (e.g., UBE2C, EZH2, CCNB1, MKI67) for future validation. These genes represent core, high-information drivers that overlap with established clinical assays, suggesting they reflect well-established proliferative biology.
• Clinical Utility: The findings support the development of time-adaptive surveillance strategies and refined risk stratification tools that can distinguish between patients at risk for early versus late relapse.
• Modesty and Limitations: The authors explicitly acknowledge limitations, including the modest sample size (n=79), which limits statistical power for rare events and generalizability. They note that the study details associations rather than establishing etiologic causality and that the lack of matched commercial assay scores in the cohort precludes direct benchmarking. The authors emphasize that external validation in larger, independent cohorts is essential to confirm the robustness of these findings.