Evolutionarily informed machine learning enhances the power of predictive gene-to-phenotype relationships.

Area of Science:

• Genomics
• Systems Biology
• Machine Learning

Background:

• Predicting phenotypic outcomes from genomic data is a key challenge in systems biology.
• Gene expression data holds potential for predicting traits, but requires effective feature selection and validation.

Purpose of the Study:

• To develop and validate an evolutionarily informed machine learning approach for predicting phenotypic outcomes from transcriptome data.
• To identify and functionally validate genes and transcription factors associated with nitrogen use efficiency (NUE).

Main Methods:

• Applied machine learning to transcriptome responses conserved across species (Arabidopsis, maize) for phenotype prediction.
• Utilized evolutionarily conserved nitrogen-responsive genes to reduce feature dimensionality and enhance model predictive power.
• Functionally validated candidate genes, specifically transcription factors, for their role in NUE.

Main Results:

• The evolutionarily informed approach significantly improved the predictive accuracy of gene-to-trait models.
• Identified seven candidate transcription factors in Arabidopsis and one in maize with predictive power for NUE.
• Demonstrated the pipeline's applicability to other species like rice and mouse models.

Conclusions:

• Evolutionarily conserved gene expression patterns provide a biologically sound basis for feature selection in machine learning.
• The developed pipeline effectively predicts phenotypic traits and identifies key regulatory genes across diverse species.
• This approach has broad potential for discovering genes influencing physiological or clinical traits in various biological contexts.