Predictive biomarkers for embryotoxicity: a machine learning approach to mitigating multicollinearity in RNA-Seq

Area of Science:

• Transcriptomics
• Gene expression analysis
• Machine learning in biology

Background:

• Multicollinearity in gene expression data can reduce predictive model reliability.
• Identifying reliable biomarkers for embryotoxicity is crucial for developmental toxicology research.
• RNA-Seq data from embryoid bodies (EB) exposed to 5-fluorouracil perturbation was used.

Purpose of the Study:

• To examine multicollinearity among closely related genes in RNA-Seq data.
• To identify potential early embryotoxicity assessment biomarkers.
• To develop a robust methodology for machine learning in transcriptomics data analysis.

Main Methods:

• Correlation studies and variance inflation factor (VIF) to assess multicollinearity.
• Recursive feature elimination with cross-validation (RFECV) for feature selection.
• Quantitative PCR (qPCR) validation of candidate genes (Dppa5a, Gdf3, Zfp42, Meis1, Hoxa2, Hoxb1).

Main Results:

• Zfp42 and Hoxb1 were identified as the top two predictive features by RFECV.
• A two-feature prediction model using Zfp42 and Hoxb1 showed statistical significance (p=0.0044).
• RFECV successfully reduced redundancies and multicollinearity in the dataset.

Conclusions:

• Zfp42 and Hoxb1 are potential biomarkers for early embryotoxicity screening.
• The study presents a systematic machine learning methodology for transcriptomics data analysis.
• This approach enhances predictive model accuracy and feasibility for toxicity screening.