Fishing for a reelGene: evaluating gene models with evolution and machine learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome assembly and annotation are crucial for understanding gene function.
• New gene models arise from each annotated assembly, leading to inconsistencies.
• Annotation errors stem from pseudogene misclassification, transposon activity, and intron retention.

Purpose of the Study:

• To develop reelGene, a machine learning pipeline for evaluating gene model predictions.
• To assess the accuracy of gene models in Zea mays ssp. mays (maize).
• To provide a resource for investigating genome biology and gene function.

Main Methods:

• reelGene employs a pipeline of machine learning models focusing on transcription boundaries, mRNA integrity, and protein structure.
• Models utilize sequence characteristics and evolutionary conservation across related taxa.
• Machine learning models predict gene function based on conserved evolutionary grammar.

Main Results:

• reelGene evaluated 1.8 million transcript models in maize, classifying 28% as incorrectly annotated or non-functional.
• The tool confirmed 92.2% of maize proteome genes and 99.2% of classical maize genes as functional.
• Analysis revealed 10.3% of dispensable genes are functional and a 30% bias towards M1 subgenome retention in duplicate genes.

Conclusions:

• reelGene is an effective tool for evaluating gene model accuracy in maize and related species like sorghum and miscanthus.
• The pipeline aids in investigating genome biology, identifying functional dispensable genes and biases in gene retention.
• reelGene is accessible via MaizeGDB as a browser track and Shiny App for researchers.