Development of a machine learning framework for radiation biomarker discovery and absorbed dose prediction

Area of Science:

• Radiation biology
• Bioinformatics
• Machine Learning

Background:

• Molecular radiation biomarkers are crucial for radiotherapy, risk assessment, and space travel.
• Conventional biomarker screening is slow and prone to human bias.
• Machine learning (ML) offers improved sensitivity, specificity, and speed for biomarker identification.

Purpose of the Study:

• Develop a resource-efficient ML framework for discovering radiation biomarkers in normal tissues.
• Identify tissue-specific biomarker panels for predicting radiation dose.

Main Methods:

• Utilized a transcriptomic dataset (GSE44762) of murine kidney tissues.
• Employed ML models within the R caret package for feature selection and analysis.
• Evaluated biomarker performance using Principal Component Analysis (PCA) and dose regression.

Main Results:

• The caret framework significantly reduced processing time compared to traditional methods.
• k-Nearest Neighbor (KNN) demonstrated optimal performance, identifying key genes like *Cdkn1a*, *Gria3*, *Mdm2*, *Plk2*, *Brf2*, *Ccng1*, and *Ddit4l*.
• The identified biomarker panels accurately categorized radiation dose groups and tissues, with high correlation (R²=0.97-0.99) in dose prediction.

Conclusions:

• The caret framework provides a resource-efficient tool for radiation biomarker discovery.
• Novel mRNA biomarkers (*Brf2*, *Ddit4l*, *Gria3*) were identified with potential for dose and tissue-specific radiation response.
• Further validation with larger datasets is recommended for improved accuracy, particularly at lower radiation doses.