Breast cancer prediction with transcriptome profiling using feature selection and machine learning methods

Area of Science:

• Biomedical Informatics
• Computational Biology
• Oncology

Background:

• Breast cancer diagnosis relies on accurate and early detection methods.
• Hybrid Machine Learning Systems (HMLS) offer a promising approach for complex disease classification.
• Transcriptome profiling is crucial for identifying molecular signatures associated with cancer.

Purpose of the Study:

• To discover optimal Hybrid Machine Learning Systems (HMLS) for breast cancer diagnosis.
• To identify a high-importance transcriptome profile for early breast cancer detection.
• To enhance classification procedures using advanced machine learning techniques.

Main Methods:

• Utilized a dataset of 762 breast cancer patients and 138 normal subjects.
• Employed four feature selection algorithms (ANOVA, Mutual Information, Extra Trees, Logistic Regression) and Principal Component Analysis for feature extraction.
• Evaluated 13 classification algorithms with automated hyperparameter tuning, including Logistic Regression (LGR) and Multilayer Perceptron (MLP), using balanced accuracy and Area Under the Curve (AUC).

Main Results:

• The LGR feature selection combined with the MLP classifier achieved the highest balanced accuracy (0.86) and AUC (0.94).
• The LGR + LGR classifier also demonstrated strong performance with a balanced accuracy of 0.84 and AUC of 0.94.
• Identified 20 specific biomarkers (e.g., TMEM212, SNORD115-13, ATP1A4) with high ranking for breast cancer detection.

Conclusions:

• The combination of Logistic Regression (LGR) for feature selection and Multilayer Perceptron (MLP) as a classifier yielded the best performance.
• The identified 20 biomarkers are highly ranked and significant for breast cancer detection.
• This HMLS approach facilitates early breast cancer detection through accurate transcriptome profiling and classification.