a.jighly@agrisapiens.com.au; r.joukhadar@agrisapiens.com.au; germancspangenberg@qau.edu.cn

Despite the promise of Deep Learning (DL) models in genomic prediction, their success in improving the prediction accuracy over standard linear models has been limited. Single nucleotide polymorphism (SNP) data is inherited in a 1D structure (along the chromosome) which typically fits the 1D convolutional neural networks (CNN) structure. Several attempts have been made to adapt SNP data to a 2D structure, suitable for 2D CNNs, which were originally designed to analyse graphic data.

These attempts have involved splitting allelic configurations into three channels (two homozygotes and one heterozygote), stacking multiple genomic relatedness matrices, or generating a 2D grayscale graph based on the Linkage Disequilibrium (LD) relations among SNPs.

However, these models have not consistently improved prediction accuracy. Here, we propose SNP-Portrait, a novel structure for SNP data that enhances the utilisation of loci interactions, thereby improving prediction accuracy. SNP-Portrait is designed to accommodate complex loci interaction scenarios and investigate epistatic interactions among two or more loci. SNP-Portrait was tested on a wheat dataset of 599 individuals phenotyped for grain yields in four environments.

We conducted 100 random replicates, each time using 20% of the population for validation. Our results showed that the model assuming 3-loci interaction resulted in the highest prediction accuracy with 4% improvement over the standard GBLUP model and 7% over the equivalent 1D CNN model. Unlike other DL models, SNP-Portrait consistently improved prediction accuracy across different tested scenarios.

Testing SNP-Portrait on more datasets is currently underway with promising results of achieving consistent improvements of prediction accuracy over standard models.