Predicting ionizing radiation exposure using biochemically-inspired genomic machine learning.

Area of Science:

• Genomics
• Bioinformatics
• Radiation Biology

Background:

• Transcriptomic data and machine learning (ML) show promise for biodosimetry.
• Existing gene signatures lack robust validation on independent datasets for large-scale use.
• This study focuses on developing and rigorously validating ML-based gene signatures for radiation exposure.

Purpose of the Study:

• To develop and validate robust human and murine gene signatures for biodosimetry using ML.
• To assess the accuracy and specificity of these signatures in dose estimation.
• To explore the potential of gene signatures in differentiating radiation exposure from other treatments.

Main Methods:

• Utilized Gene Expression Omnibus (GEO) datasets of exposed human and murine lymphocytes.
• Preprocessed data using nearest neighbor imputation and selected radiation-responsive genes.
• Employed Minimum Redundancy Maximum Relevance (mRMR) for gene ranking and Support Vector Machines (SVM) with sequential feature selection for signature derivation.
• Validated signatures using k-fold and traditional methods on independent datasets.

Main Results:

• Achieved high k-fold validation accuracies up to 98% for human signatures (e.g., DDB2, PRKDC, TPP2, PTPRE, GADD45A).
• Attained traditional validation accuracies up to 92% for human signatures (e.g., DDB2, CD8A, TALDO1, PCNA, EIF4G2, LCN2, CDKN1A, PRKCH, ENO1, PPM1D).
• Demonstrated signature specificity to differentiate between chemotherapy and radiotherapy, and granularity for murine dose estimation relevant to cytokine therapy.

Conclusions:

• ML-derived gene signatures for ionizing radiation exposure exhibit low error rates on external, independent datasets.
• These validated signatures demonstrate high specificity and granularity for accurate biodosimetric dose estimation.
• The findings support the utility of robustly validated gene signatures in radiation biodosimetry and treatment response assessment.