Metabolomics-Based Machine Learning Models Accurately Predict Breast Cancer Estrogen Receptor Status.

Area of Science:

• Metabolomics
• Biomarker Discovery
• Machine Learning in Oncology

Background:

• Breast cancer is a leading cause of female mortality globally.
• Accurate estrogen receptor (ER) status determination is critical for prognosis and treatment.
• Current ER status detection methods can be invasive and challenging.

Purpose of the Study:

• To identify non-invasive plasma metabolite biomarkers for distinguishing ER-positive from ER-negative breast cancers.
• To develop and validate a machine learning model for breast cancer subtyping based on metabolomics data.

Main Methods:

• Utilized metabolomics data from plasma samples of breast cancer patients and healthy controls.
• Employed Recursive Feature Elimination (RFE) with Random Forest (RF) for feature selection.
• Applied data augmentation techniques (Gaussian noise, ADASYN) to handle class imbalance.
• Evaluated four machine learning algorithms (RF, SVC, XGBoost, LR) using grid search.

Main Results:

• An optimal subset of 30 features (29 biomarkers and age) was identified using RFE-RF.
• The Random Forest classifier achieved the highest performance with an Area Under the Curve (AUC) of 0.95 and 93% accuracy.
• The developed model demonstrated high efficacy in distinguishing ER-positive from ER-negative breast cancers.

Conclusions:

• Machine learning analysis of plasma metabolomics holds significant promise for non-invasive breast cancer subtyping.
• This approach can potentially lead to a novel analytical tool to overcome current challenges in ER status determination.
• Improved precision in breast cancer subtyping can enhance patient prognosis and treatment strategies.