Exploring Machine Learning Algorithms to Unveil Genomic Regions Associated With Resistance to Southern Root-Knot Nematode in Soybeans

Abstract

Southern root-knot nematode [SRKN, Meloidogyne incognita (Kofold & White) Chitwood] is a plant-parasitic nematode challenging to control due to its short life cycle, a wide range of hosts, and limited management options, of which genetic resistance is the main option to efficiently control the damage caused by SRKN. To date, a major quantitative trait locus (QTL) mapped on chromosome (Chr.) 10 plays an essential role in resistance to SRKN in soybean varieties. The confidence of discovered trait-loci associations by traditional methods is often limited by the assumptions of individual single nucleotide polymorphisms (SNPs) always acting independently as well as the phenotype following a Gaussian distribution. Therefore, the objective of this study was to conduct machine learning (ML)-based genome-wide association studies (GWAS) utilizing Random Forest (RF) and Support Vector Machine (SVM) algorithms to unveil novel regions of the soybean genome associated with resistance to SRKN. A total of 717 breeding lines derived from 330 unique bi-parental populations were genotyped with the Illumina Infinium BARCSoySNP6K BeadChip and phenotyped for SRKN resistance in a greenhouse. A GWAS pipeline involving a supervised feature dimension reduction based on Variable Importance in Projection (VIP) and SNP detection based on classification accuracy was proposed. Minor effect SNPs were detected by the proposed ML-GWAS methodology but not identified using Bayesian-information and linkage-disequilibrium Iteratively Nested Keyway (BLINK), Fixed and Random Model Circulating Probability Unification (FarmCPU), and Enriched Compressed Mixed Linear Model (ECMLM) models. Besides the genomic region on Chr. 10 that can explain most of SRKN resistance variance, additional minor effects SNPs were also identified on Chrs. 10 and 11. The findings in this study demonstrated that overfitting in GWAS may lead to lower prediction accuracy, and the detection of significant SNPs based on classification accuracy limited false-positive associations. The expansion of the basis of the genetic resistance to SRKN can potentially reduce the selection pressure over the major QTL on Chr. 10 and achieve higher levels of resistance.

Genetic architecture of soybean tolerance to off-target dicamba

“…Given the average of squared VIP scores are equal to 1.0, a threshold higher than 1.0 is employed to select features that make the most substantial contribution to D e ( Chong and Jun, 2005 ; Cocchi et al., 2018 ). In scenarios where the number of independent variables significantly exceeds the number of observations and there is considerable multicollinearity, a threshold of 2.0 is suggested to filter significant predictors ( Cocchi et al., 2018 ; Canella Vieira et al., 2022c ). A total of 113 SNPs with VIP scores above 2.0 were distributed across chromosomes 1 (7 SNPs, LG D1a), 2 (6 SNPs, LG D1b), 3 (6 SNPs, LG N), 4 (1 SNP, C1), 6 (25 SNPs), 7 (1 SNP, LG M), 8 (2 SNPs), 9 (2 SNPs), 10 (4 SNPs), 11 (1 SNP, LG B1), 13 (6 SNPs, LG F), 17 (16 SNPs), and 19 (36 SNPs) ( Figure 4 ).…”

“…The ML-GWAS pipeline to identify the combination of predictors yielding the highest prediction accuracy was implemented following the protocol first described by Canella Vieira et al. (2022c) ( Figure 1 ).…”

“…The optimal combination of trainable parameters was provided using the function ‘ tune ’. The final model was tunned using a grid search for the cost ranging from 0.01, 0.1, 1, 10, 100, and 1000, and gamma ranging from 0.0001, 0.001, 0.01, 0.5, and 1 ( Canella Vieira et al., 2022c ).…”

“…Genome-wide association studies (GWAS) are performed to detect significant associations between a trait of interest and molecular markers using linear regression models ( Yu et al., 2006 ; Hwang et al., 2014 ; Vuong et al., 2015 ; Zhang et al., 2015 ) as well as machine and deep learning algorithms ( Liu et al., 2019 ; Yoosefzadeh-Najafabadi et al., 2021 ; Canella Vieira et al., 2022c ; Yoosefzadeh-Najafabadi et al., 2022 ). Using a panel of genetically diverse soybean accessions, significant associations were reported between off-target dicamba response and single nucleotide polymorphisms (SNPs) on chromosomes 10 [Linkage Group (LG) O], 11 (LG B1), 15 (LG E), 18 (LG G), and 19 (LG L) ( Canella Vieira et al., 2022a ).…”

Recent advances in artificial intelligence, mechanistic models, and speed breeding offer exciting opportunities for precise and accelerated genomics‐assisted breeding

“…Zhou et al (2019) combined GWAS and ML feature selection, revealing significant identification power for mining minor QTL to help understand biological activities between genotypes and phenotypes related to the causal networks of interacting genes. Other studies have combined ML‐based algorithms with GWAS to identify genetic variants associated with traits of interest, such as disease susceptibility (Silva et al, 2022), root‐knot nematode resistance (Vieira et al, 2022), and yield‐related traits (Yoosefzadeh‐Najafabadi et al, 2022).…”