Benchmarking Feature Selection Methods and Prediction Models for Flowering Time Prediction in Maize.

Area of Science:

• Plant genomics
• Computational biology
• Crop science

Background:

• Flowering time is crucial for crop adaptation and yield.
• Predicting flowering time requires analyzing complex multi-omics data.
• Machine learning applications in plant genomics, especially feature selection and prediction models, need further investigation for flowering time prediction and gene discovery.

Purpose of the Study:

• To conduct a large-scale benchmarking study of feature selection (FS) and genomic prediction (GP) methods for flowering time prediction in maize.
• To evaluate the impact of different FS-GP combinations on predictive performance and gene discovery.
• To interpret machine learning models for identifying key flowering time regulatory genes.

Main Methods:

• Evaluated 42 combinations of seven feature selection methods and six prediction models.
• Integrated SNP and transcriptomic data for analysis.
• Utilized SHAP (SHapley Additive exPlanations) with a random forest (RF) framework for model interpretability and feature contribution quantification.

Main Results:

• Identified known maize flowering time regulators, including ZmMADS69 and ZmRap2.7.
• Discovered additional candidate genes potentially involved in the flowering regulatory network.
• Demonstrated the effectiveness of integrated multi-omics data and interpretable machine learning for gene discovery.

Conclusions:

• The study provides valuable insights into the genetic regulation of flowering time in maize.
• The developed framework offers an effective approach for discovering candidate genes from multi-omics data for crop improvement.
• Optimizing FS-GP combinations is key for accurate genomic prediction and gene discovery in plant breeding.