Comparing statistical learning methods for complex trait prediction from gene expression

Area of Science:

• Quantitative genetics
• Statistical genomics
• Systems biology

Background:

• Accurate prediction of complex traits is crucial in quantitative genetics.
• Gene expression levels show potential for higher trait prediction accuracy compared to genotypes.
• A comprehensive assessment of prediction methods using gene expression is needed.

Purpose of the Study:

• To compare the predictive ability of various statistical learning methods using gene expression data.
• To evaluate the impact of functional annotations (Gene Ontology) on prediction accuracy.
• To identify key genes associated with starvation resistance and startle response in Drosophila.

Main Methods:

• Utilized data from the Drosophila Genetic Reference Panel (DGRP).
• Compared multiple statistical learning methods with varying assumptions on gene effects and interactions.
• Incorporated Gene Ontology (GO) functional annotations into prediction models.

Main Results:

• Prediction accuracy varied across methods and traits, with variable selection methods excelling for female starvation resistance.
• Gene Ontology annotations improved prediction accuracy for specific biologically relevant terms.
• Identified key genes: Insulin-like Receptor (InR) for starvation resistance, crumbs (crb) and imaginal disc growth factor 2 (Idgf2) for startle response.

Conclusions:

• Transcriptomic prediction holds significant potential for understanding complex traits.
• The choice of statistical learning method and inclusion of functional annotations are critical for accurate prediction.
• This study provides insights into the genetic architecture of complex traits in Drosophila.