Machine learning models outperform deep learning models, provide interpretation and facilitate feature selection for...

Area of Science:

• Agricultural Science
• Genomics
• Computational Biology

Background:

• Rapid expansion of crop genomic and trait data enables advanced crop improvement strategies.
• Machine learning (ML) and deep learning (DL) are key for predictive data analysis in agriculture.
• Limited studies compare ML and DL for genotype-to-phenotype prediction and model interpretability.

Purpose of the Study:

• To develop accurate genotype-to-phenotype prediction models using genome-wide molecular markers and traits in soybean.
• To compare the performance of ML (XGBoost, random forest) and DL models.
• To interpret prediction models and assess the impact of feature selection on performance.

Main Methods:

• Utilized genome-wide molecular markers and phenotypic data from 1110 soybean individuals.
• Developed and compared prediction models using XGBoost, random forest, and deep learning algorithms.
• Employed F-score and feature importance for ranking single nucleotide polymorphisms (SNPs) and reducing marker input.

Main Results:

• XGBoost and random forest models outperformed deep learning models in 13 out of 14 prediction tasks.
• Identified top-ranked SNPs via F-score from XGBoost, showing overlap with significant loci from Genome-Wide Association Studies (GWAS) and prior research.
• Reduced marker input by up to 90% using feature importance, with models maintaining or improving prediction accuracy.

Conclusions:

• Interpretable machine learning approaches, such as XGBoost, are effective for genomic-based trait prediction in soybean.
• Feature selection based on model interpretability can enhance prediction efficiency without sacrificing accuracy.
• Findings support the utility of interpretable ML for accelerating crop improvement in soybean and other crops.