A robust machine learning approach for breast cancer subtype classification using relative gene expression order...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Breast cancer subtype classification is crucial for personalized treatment and prognosis.
• Machine learning (ML) models analyzing gene expression show promise but struggle with technical data variability.
• Robust and generalizable ML classifiers are needed for reliable breast cancer subtyping.

Purpose of the Study:

• To develop and evaluate a robust machine learning approach for breast cancer subtype classification.
• To leverage relative gene expression order representations to overcome technical variability in gene expression data.
• To assess the performance and clinical relevance of these ML models across diverse datasets.

Main Methods:

• Proposed a machine learning approach using rank- and word2vec embedding-based relative gene expression order.
• Representations were derived from the within-sample relative expression order of PAM50 genes.
• Models were trained and systematically evaluated for robustness and clinical relevance on benchmark datasets (SCAN-B, TCGA-BRCA).

Main Results:

• Rank- and word2vec models achieved high precision and recall (≥91%) on cross-validation and test sets.
• Models demonstrated high accuracy (≥95%) on the external TCGA-BRCA dataset.
• Gene expression order representations preserved biological variation and enabled meaningful survival stratification.

Conclusions:

• Relative gene expression order, via ranks and word2vec embeddings, enables robust breast cancer subtype classification.
• This method reduces dependence on cohort-wide normalization, addressing technical variability.
• The approach supports reliable subtype classification in diverse clinical settings.