GPTransformer: A Transformer-Based Deep Learning Method for Predicting Fusarium Related Traits in Barley

Abstract

Fusarium head blight (FHB) incited by Fusarium graminearum Schwabe is a devastating disease of barley and other cereal crops worldwide. Fusarium head blight is associated with trichothecene mycotoxins such as deoxynivalenol (DON), which contaminates grains, making them unfit for malting or animal feed industries. While genetically resistant cultivars offer the best economic and environmentally responsible means to mitigate disease, parent lines with adequate resistance are limited in barley. Resistance breeding based upon quantitative genetic gains has been slow to date, due to intensive labor requirements of disease nurseries. The production of a high-throughput genome-wide molecular marker assembly for barley permits use in development of genomic prediction models for traits of economic importance to this crop. A diverse panel consisting of 400 two-row spring barley lines was assembled to focus on Canadian barley breeding programs. The panel was evaluated for FHB and DON content in three environments and over 2 years. Moreover, it was genotyped using an Illumina Infinium High-Throughput Screening (HTS) iSelect custom beadchip array of single nucleotide polymorphic molecular markers (50 K SNP), where over 23 K molecular markers were polymorphic. Genomic prediction has been demonstrated to successfully reduce FHB and DON content in cereals using various statistical models. Herein, we have studied an alternative method based on machine learning and compare it with a statistical approach. The bi-allelic SNPs represented pairs of alleles and were encoded in two ways: as categorical (–1, 0, 1) or using Hardy-Weinberg probability frequencies. This was followed by selecting essential genomic markers for phenotype prediction. Subsequently, a Transformer-based deep learning algorithm was applied to predict FHB and DON. Apart from the Transformer method, a Residual Fully Connected Neural Network (RFCNN) was also applied. Pearson correlation coefficients were calculated to compare true vs. predicted outputs. Models which included all markers generally showed marginal improvement in prediction. Hardy-Weinberg encoding generally improved correlation for FHB (6.9%) and DON (9.6%) for the Transformer network. This study suggests the potential of the Transformer based method as an alternative to the popular BLUP model for genomic prediction of complex traits such as FHB or DON, having performed equally or better than existing machine learning and statistical methods.

Crop genomic selection with deep learning and environmental data: A survey

“…Most of the multi-environment deep learning architecture we discussed so far sought to capture the spatial and/or temporal effect of environmental variables on traits and later incorporated genomic data into the model for estimating phenotypes. Though a few deep learning models were developed by employing attention for genomic selection (Gangopadhyay et al, 2020;Jubair et al, 2021;Måløy et al, 2021), we believe attention-based architectures are the most promising approach for genomic selection. Attention-based methods can capture both temporal and spatial information and summarize the input data by aggregating them based on importance scores.…”

“…The preprocessing steps may involve removing uninformative markers, imputation of missing values and representing the features in some other forms. If the minor allele frequency ≤ 5% (Ma et al, 2018 ; Jubair et al, 2021 ) or more than 30% values are missing, then the marker is usually removed as those markers do not bear any relevant information. To replace the missing values, one popular imputation techniques is k-nearest neighbor.…”

“…GS for single environment trials employs GS to identify top individuals to create a new variety for a specific environment (Meuwissen et al, 2001 ; Heffner et al, 2009 ; Crossa et al, 2017 ; Jubair et al, 2021 ). If the environment changes, single environment GS does not guarantee that the new variety will have the desired outcome in that new environment (Oakey et al, 2016 ).…”

“…In a single environment trial, it is typically assumed that the environment is constant, thus, there is no effect of environment on genotypes. A number of deep learning methods for single environment trials have been published (McDowell, 2016 ; Rachmatia et al, 2017 ; Ma et al, 2018 ; Jubair and Domaratzki, 2019 ; Zingaretti et al, 2020 ; Jubair et al, 2021 ; Montesinos-Lopez et al, 2021 ). These methods differ in their deep learning architectures and focus on how they capture the genetic information.…”

Genetic relationships and low diversity among the tea-oil Camellia species in Sect. Oleifera, a bulk woody oil crop in China

“…Assuming valid Hardy-Weinberg equilibrium (HWE; Jubair et al, 2021;Graffelman and Weir, 2022;Yarar et al, 2022), we used POPGENE 1.32 software to analyze the original ISSR-PCR and SRAP-PCR data and determine genetic diversity indexes, including the percentage of polymorphic sites (P), observed number of alleles (Na), effective number of alleles (Ne), Shannon's information index (I), and Nei's gene diversity (H). DnaSP software was used to for the haploid analysis of 5 chloroplast sequences in different individuals and populations.…”