BreastSubtypeR: Assumption-aware intrinsic molecular subtyping for breast cancer research

Yang, Qiao; Hartman, Johan; Sifakis, Emmanouil

doi:10.5281/zenodo.17080602

BreastSubtypeR: Assumption-aware intrinsic molecular subtyping for breast cancer research

Yang, Qiao; Hartman, Johan; Sifakis, Emmanouil

Publisher: Zenodo
DOI: 10.5281/zenodo.17080602
Repository: zenodo.org
Language: en

Abstract

Intrinsic molecular subtyping of breast cancer into categories such as Luminal A, Luminal B, HER2-enriched, Basal-like, and Normal-like is fundamental for understanding tumour biology and guiding personalised treatment strategies. In clinical practice, molecular subtyping is standardised and performed with approved diagnostic assays. However, research implementations of these subtyping methods are fragmented across various tools, leading to inconsistencies and reduced reproducibility. We present BreastSubtypeR, an R/Bioconductor package that integrates ten established subtyping approaches into a unified, reproducible framework. Its core innovation, AUTO mode, evaluates cohort characteristics (e.g., ER/HER2 prevalence, subtype purity, subgroup size) and programmatically disables classifiers whose assumptions are violated – reducing misclassification in skewed or small cohorts. Additional features include direct cross-method benchmarking within a single interface, standardised input handling with method-specific normalisation, and Entrez ID–based probe/gene mapping for robust cross-platform fidelity. For accessibility, the companion application iBreastSubtypeR provides a local R Shiny GUI, enabling non-programmers to apply single-method reproducible workflows used by computational researchers. By combining assumption-aware method selection, harmonised preprocessing, and dual accessibility, BreastSubtypeR addresses a critical need for reproducibility and transparency in translational breast cancer research. The package and Shiny app are freely available via Bioconductor and GitHub, with extensive documentation and example datasets. Presented at the European Bioconductor Conference (EuroBioC2025), Barcelona, September 2025.

Full text

Assumption-aware intrinsic molecular subtyping for breast cancer research Qiao Yang1, Johan Hartman1,2, Emmanouil G. Sifakis1 1Department of Oncology-Pathology, Karolinska Institutet, Stockholm, Sweden 2Department of Clinical Pathology and Cancer Diagnostics, Karolinska University Hospital Introduction •Intrinsic molecular subtypes are associated with prognosis and support systemic therapy decisions in breast cancer. •In clinical practice, subtyping is standardised; in research, variable preprocessing pipelines and unmet cohort assumptions can shift calls. •A unified, assumption-aware framework is needed to improve reliability and reproducibility in research. Contact Emmanouil G. Sifakis, PhD Karolinska Institutet [email protected] Objective •Provide a unified, assumption-aware Bioconductor framework for more accurate, reproducible intrinsic molecular subtyping in breast cancer research. Key results •Skewed cohorts: AUTO improves accuracy by 18–19 percentage points and Cohen’s kappa (κ) by +0.20–0.26 vs Excluded (by AUTO) (Fig. 2). •Empirical thresholds: Original NC methods become unreliable outside 39%–69% ER+ or n is small; AUTO excludes accordingly (Fig. 3). •QC: Entropy flags low-confidence calls, supporting cautious interpretation and reproducibility (Fig. 4). Fig. 1 AUTO logic: only run methods when cohort assumptions (ER/HER2 balance, subtype purity, subgroup size) are met. Fig. 2 Skewed cohorts (S10/S90): (A) S10 = 10% ER+; (B) S90 = 90% ER+. AUTO outperforms Excluded (by AUTO) in κ. Blue = AUTOenabled; Orange = Excluded (by AUTO). Fig. 4 Entropy QC: Higher entropy = more uncertainty. QC flags low-confidence intrinsic subtype calls to guide cautious interpretation. Fig. 3 Thresholds (NC): (A) ER+ proportion, unreliable outside 39– 69%. (B) ER+ subgroup size: n < 15. (C) ER–subgroup size: n < 18. AUTO excludes original NC methods under these conditions. A B C A B Resources BreastSubtypeR JohanHartmanGroupBioteam/BreastSubtypeR Scan for package + docs (Bioconductor) Conclusions •Unified toolkit for reproducible intrinsic molecular subtyping in research. •AUTO reduces misuse in skewed/small cohorts and improves accuracy. •Available via Bioconductor and Shiny app. BreastSubtypeR provides a unified, assumption-aware framework that makes intrinsic molecular subtyping in breast cancer research more accurate and reproducible. Methods •Classifiers: Nearest-Centroid (NC) and Single-Sample Predictor (SSP) methods via one interface. •AUTO mode: screens ER/HER2 balance, subtype purity, and subgroup size; enables only methods whose assumptions are met (Fig. 1). •Normalisation and gene mapping: method-specific defaults with Entrez-ID mapping. •Outputs: subtype calls, entropy-based QC, cross-method concordance, and built-in plots.