기계학습 기반 국내 벼 단백질 함량, 등숙률, 초장, 수장 예측을 위한 주요 SNP 마커 선발
Machine Learning Method to Select Single Nucleotide Polymorphism Markers for Protein Content, Grain Filling Rate, Height, and Panicle Length in Korean Rice
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Early selection of grain quality traits in rice (Oryza sativa L.), is essential to improve the yield and quality of this staple crop. We analyzed four key traits—protein content, grain filling rate, height, and panicle length—in 85 Korean cultivars. Through whole- genome resequencing we identified 12,718,879 raw single nucleotide polymorphisms (SNPs); after PLINK-based quality control (bi-allelic selection, call rate≥0.90, MAF≥0.03), ~2.20 million high-quality SNPs remained for machine-learning (ML) pre-screening. To rank the features (without marker-level inference), we applied a liberal univariate PLINK case-control scan using the top and bottom 30% per trait. We also analyzed associations with a linear mixed model (GCTA v1.93.2, MLMA; fixed covariates: ecotype, PC1, PC2; random effect: GRM) to verify calibration under population structure; with n=85, no genome-wide significant hits were detected, and QQ-plots indicated adequate calibration (per-trait effective tests m≈1.54-1.57 million under stricter filters). The random forest feature importance prioritized 26, 51, 19, and 20 core SNPs for the four traits, respectively. Across the algorithms, the best models achieved mean accuracies of 81.8% (protein content), 81.0% (grain filling rate), 73.1% (height), and 94.0% (panicle length). All selected SNPs met the Fluidigm array design requirements, supporting its deployment as a compact, genotype- based panel for early selection and a practical step toward digital breeding in rice.
Abstract
Early selection of grain quality traits in rice (Oryza sativa L.), is essential to improve the yield and quality of this staple crop. We analyzed four key traits—protein content, grain filling rate, height, and panicle length—in 85 Korean cultivars. Through whole- genome resequencing we identified 12,718,879 raw single nucleotide polymorphisms (SNPs); after PLINK-based quality control (bi-allelic selection, call rate≥0.90, MAF≥0.03), ~2.20 million high-quality SNPs remained for machine-learning (ML) pre-screening. To rank the features (without marker-level inference), we applied a liberal univariate PLINK case-control scan using the top and bottom 30% per trait. We also analyzed associations with a linear mixed model (GCTA v1.93.2, MLMA; fixed covariates: ecotype, PC1, PC2; random effect: GRM) to verify calibration under population structure; with n=85, no genome-wide significant hits were detected, and QQ-plots indicated adequate calibration (per-trait effective tests m≈1.54-1.57 million under stricter filters). The random forest feature importance prioritized 26, 51, 19, and 20 core SNPs for the four traits, respectively. Across the algorithms, the best models achieved mean accuracies of 81.8% (protein content), 81.0% (grain filling rate), 73.1% (height), and 94.0% (panicle length). All selected SNPs met the Fluidigm array design requirements, supporting its deployment as a compact, genotype- based panel for early selection and a practical step toward digital breeding in rice.