Robust predictions of specialized metabolism genes through machine learning

Area of Science:

• Plant biology
• Genomics
• Biochemistry

Background:

• Plant specialized metabolism (SM) enzymes synthesize unique compounds vital for ecological interactions, evolution, and biotechnology.
• Distinguishing SM genes from general metabolism (GM) genes is crucial for understanding plant biochemical diversity.

Purpose of the Study:

• To identify distinguishing features between SM and GM genes in *Arabidopsis thaliana*.
• To develop a predictive model for identifying SM genes using machine learning.

Main Methods:

• Comparative analysis of gene features: duplication patterns, sequence conservation, transcription levels, protein domain content, and gene network properties.
• Application of machine learning algorithms to integrate multiple features for SM gene prediction.
• Validation of the prediction model on known and unknown SM genes.

Main Results:

• SM genes exhibit distinct characteristics compared to GM genes, including tandem duplication, coexpression with paralogs, narrower and lower expression levels, reduced sequence conservation, and less connectivity in gene networks.
• A machine learning model integrating these features achieved high accuracy (87% true positive, 71% true negative) in predicting SM genes.
• The model successfully predicted 86% of previously uncharacterized SM genes and identified 1,220 *A. thaliana* genes with unknown functions, assigning an SM score for confidence.

Conclusions:

• Integrating multiple gene features via machine learning provides an effective strategy for distinguishing and predicting plant specialized metabolism genes.
• The developed prediction model aids in discovering novel genes involved in plant specialized metabolism, expanding our understanding of plant biochemical capabilities.
• Further improvements to the model are possible by incorporating topological considerations, such as distinguishing between SM, GM, and junction genes.