Predicting phenotypes from novel genomic markers using deep learning.

Area of Science:

• Genomics
• Bioinformatics
• Plant Breeding

Background:

• Genomic selection (GS) models traditionally rely on single nucleotide polymorphism (SNP) markers for phenotype prediction.
• High dimensionality of genome-wide SNP data presents a significant challenge for predictive accuracy in GS.
• Advances in DNA sequencing enable the study of novel genomic variants like structural variations (SVs) and transposable elements (TEs).

Purpose of the Study:

• To develop a deep convolutional neural network, NovGMDeep, for phenotype prediction using SVs and TEs.
• To evaluate the efficacy of SVs and TEs as markers for genomic selection.
• To compare the prediction accuracy of NovGMDeep with traditional SNP-based models and conventional statistical methods.

Main Methods:

• Development of the NovGMDeep deep convolutional neural network model.
• Training and testing the model on *Arabidopsis thaliana* and *Oryza sativa* samples.
• Utilizing k-fold cross-validation for robust evaluation.
• Assessing prediction accuracy using Pearson's Correlation Coefficient (PCC) and mean absolute error (MAE).

Main Results:

• Models trained with SVs and TEs demonstrated higher prediction correlation compared to SNP-based models.
• NovGMDeep achieved superior prediction accuracy over conventional statistical models.
• The study confirmed the utility of SVs and TEs in genotype-to-phenotype association studies.

Conclusions:

• SVs and TEs are valuable, yet often unappreciated, genomic markers for enhancing phenotype prediction in GS.
• The NovGMDeep model offers a promising approach for leveraging SVs and TEs in plant breeding.
• This research highlights the extensive significance of SVs and TEs for accelerating crop development and improvement.