Higher-order partial least squares for predicting gene expression levels from chromatin states

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Gene expression is influenced by combinations of chromatin marks, but these patterns remain unclear.
• Understanding histone modification interactions is crucial for predicting gene expression levels.

Purpose of the Study:

• To introduce a higher-order tensor representation for analyzing chromatin states.
• To explore unknown interactions between histone modifications for improved gene expression prediction.

Main Methods:

• Developed a geometric higher-order representation (tensor) for chromatin states.
• Trained prediction models using five histone modifications (H3K4me1, H3K4me3, H3K27ac, H3K27me3, Pol II) around transcriptional start sites.
• Applied the method to three species: Human, Rhesus Macaque, and Chimpanzee.

Main Results:

• The tensor-based method significantly outperforms popular prediction methods.
• Achieved higher accuracy in predicting gene expression levels, with improvements of 1.7% in R and 11% in RMSE.
• Demonstrated the power of higher-order representations in capturing complex epigenetic interactions.

Conclusions:

• Higher-order tensor representation effectively incorporates unknown interaction information between histone modifications.
• This approach enhances the prediction of gene expression levels across different species.