Evaluating Plant Gene Models Using Machine Learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene model prediction is crucial for understanding genome function but often suffers from inaccuracies and false positives.
• Distinguishing true gene models from spurious predictions is a significant challenge in genomic annotation.

Purpose of the Study:

• To develop and validate a machine learning approach, Truegene, for classifying low-confidence gene models.
• To minimize false positives in gene prediction and enhance the reliability of genome annotation.

Main Methods:

• Utilized 14 gene and 41 protein-based characteristics from *Pisum sativum* (pea) genome annotations.
• Trained eXtreme Gradient Boost (XGBoost) classifier models using amino acid and nucleotide sequence features.
• Employed SHapley Additive exPlanations (SHAP) for feature importance analysis.

Main Results:

• Achieved prediction accuracy ranging from 87% to 90% for gene models.
• Obtained an F-1 score between 0.91 and 0.94, indicating high model performance.
• Identified key gene and protein features contributing to accurate gene model classification.

Conclusions:

• Machine learning models effectively use gene and protein features to predict gene models with high accuracy.
• Truegene provides a valuable tool for supporting and improving future gene annotation processes.
• The method successfully reduces false positives, leading to more confident conserved gene models.