Explainable machine learning framework for the molecular classification of triple negative breast cancer

Area of Science:

• Computational biology
• Genomics
• Artificial Intelligence in Medicine

Background:

• Triple-Negative Breast Cancer (TNBC) exhibits distinct molecular profiles across its four main subtypes: basal-like 1 (BL1), basal-like 2 (BL2), mesenchymal (M), and luminal androgen receptor (LAR).
• Accurate subtyping is crucial for targeted therapies and improved patient outcomes.

Purpose of the Study:

• To develop an integrative AI framework combining machine learning and explainable AI for TNBC subtype classification.
• To identify key gene biomarkers driving subtype-specific predictions and enable biomarker prioritization.

Main Methods:

• Analysis of 783 TNBC and non-TNBC cases from public datasets (GEO, GDC).
• Development of a framework with modules for gene signature identification and classification using eight ML algorithms.
• Selection of Random Forest as the best classifier (96% accuracy) and application of Shapley Additive Explanations (SHAP) for interpretability.

Main Results:

• Identification of 47 potential biomarkers for distinguishing the four TNBC subtypes.
• The Random Forest model achieved 96% testing accuracy in classifying TNBC subtypes.
• The explainable AI module successfully prioritized key hub genes driving subtype predictions.

Conclusions:

• The study identified hub genes (e.g., CDC20, CDCA2, PIMREG, KIF2C, CENPW) involved in ubiquitin-proteasome signaling and microtubule dynamics.
• These findings support the development of biomarker-driven therapies and precision medicine approaches for TNBC.
• The proposed framework offers interpretability, crucial for clinical translation and biomarker discovery in TNBC.