Application of machine learning to explore the genomic prediction accuracy of fall dormancy in autotetraploid alfalfa

Abstract

Fall dormancy (FD) is an essential trait to overcome winter damage and for alfalfa (Medicago sativa) cultivar selection. The plant regrowth height (PRH) after autumn clipping is an indirect way to evaluate FD. Transcriptomics, proteomics, and QTL mapping have revealed crucial genes correlated with FD, however, these genes can’t predict alfalfa FD very well. Here, we conducted genomic prediction of FD using whole genome SNP markers based on machine learning-related methods support vector machines (SVM) regression and regularization-related methods, such as lasso and ridge regression. The results showed that using SVM regression with linear kernel and the top 3,000 GWAS-associated markers achieved the highest prediction accuracy for FD of 64.1%. For RPH, the prediction accuracy was 59.0% using the 3,000 GWAS-associated markers and the SVM linear model. It is better than that using whole-genome markers (25.0%). Therefore, the method we explored for alfalfa FD prediction outperformed the other models, such as lasso and ElastNet. The study suggests the feasibility of using machine learning to predict FD with GWAS-associated markers, and the GWAS-associated markers combined with machine learning would benefit FD-related traits as well. Application of the methodology may provide potential targets for FD selection, which would accelerate genetic research and molecular breeding of alfalfa with optimized FD.

Glutamine and amino acid metabolism as a prognostic signature and therapeutic target in endometrial cancer

“…Because when the amount of data is large with many missing values, and the number of independent variables is much larger than the sample size, the traditional Cox regression forward, backward, and stepwise method may not be applicable. In addition to these methods, the random forest, Support Vector Machine, principle component analysis, 33 deep learning, and extreme gradient boosting of machine learning are also becoming much popular 32,34,35 . Machine learning are more robust and can outfit imbalanced datasets 35 .…”

“…This model is mainly and widely used for the prognostic analysis of tumors and other chronic diseases, and can also be used for etiology exploration in cohort studies. For high‐dimensional and multicollinear data, Lasso, Ridge, and ElasticNet regressions are more suitable 31,32 . Because when the amount of data is large with many missing values, and the number of independent variables is much larger than the sample size, the traditional Cox regression forward, backward, and stepwise method may not be applicable.…”

Effects of SNP marker density and training population size on prediction accuracy in alfalfa (Medicago sativa L.) genomic selection

“…Our results showed that for all traits, the prediction accuracies using subsets of SNPs were generally slightly higher as compared to that of using all SNPs. Several previous studies showed that a larger number of random SNPs reduced prediction accuracy in GS as compared to the SNPs derived from GWAS studies (Alemu et al, 2021;Singer et al, 2022;Spindel et al, 2016;Yan et al, 2023;Zhang et al, 2022). This high prediction appeared to be common in GS regardless of traits of interest in alfalfa (Zhang et al, 2022) and other crops such as maize (Rice & Lipka, 2019;Yan et al, 2023), wheat (Triticum aestivum L.) (Alemu et al, 2021;Arruda et al, 2016), rice (Oryza sativa L.) (Spindel et al, 2016) There is also a concern about the overestimation of prediction accuracy in the GWAS-derived approach.…”

“…Several previous studies showed that a larger number of random SNPs reduced prediction accuracy in GS as compared to the SNPs derived from GWAS studies (Alemu et al, 2021;Singer et al, 2022;Spindel et al, 2016;Yan et al, 2023;Zhang et al, 2022). This high prediction appeared to be common in GS regardless of traits of interest in alfalfa (Zhang et al, 2022) and other crops such as maize (Rice & Lipka, 2019;Yan et al, 2023), wheat (Triticum aestivum L.) (Alemu et al, 2021;Arruda et al, 2016), rice (Oryza sativa L.) (Spindel et al, 2016) There is also a concern about the overestimation of prediction accuracy in the GWAS-derived approach. Alemu et al (2021) reported that wrr-BLUP method (weighted ridge regression BLUP) could overestimate the GS accuracy due to individuals' inclusion in both the GWAS and test populations.…”

“…The utilization of a larger number of randomly selected single-nucleotide polymorphisms (SNPs) did not result in a significant improvement of GS accuracy for alfalfa forage yield (He et al, 2022) and the traits of other crops such as maize (Zea mays L.) (Li et al, 2019) compared to use of a smaller number of informative SNPs (Fu et al, 2017). Recently, Zhang et al (2022) showed an evidence of uninformative SNPs, which could reduce prediction accuracy in GS. This study demonstrated an improved prediction accuracy of 3000 SNPs based on genome-wide association studies (GWAS) in comparison to a large array of random SNPs for fall plant height (FPH) and regrowth height in alfalfa.…”