Haploid maize seeds prediction using deep learning and using mock reference genomes for genomic prediction of hybrids

Abstract

Haploid maize seeds prediction using deep learning and using mock reference genomes for genomic prediction of hybrids Prediction is a key concept for animal and plant breeding. Accurate estimates of phenotypic and genetic values are crucial for the selection of the best genotypes. For this reason, several tools have been used to improve the accuracy of these estimates, from molecular markers, used to access genetic information, to high-throughput phenotyping, used to increase sample size and phenotypic precision. Here, we present two studies involving the use of different approaches and tools in the prediction process. First, we describe a study using deep learning and images for seed phenotyping. We built a convolutional neural network (CNN) model to classify images from putative and true haploid maize seeds based on the R1-nj phenotype. Our results reveal that the CNN model could classify putative haploid maize seeds with high accuracy (97%). However, the CNN model was unable to recognize true haploid seeds. Finally, we provide a highly accurate and trained CNN model to the scientific community to classify haploid maize seeds via R1-nj. In the latter, we studied using mock genomes to discover markers and their effect on estimates of genetic diversity and genomic prediction of hybrids. Moreover, we compared them with SNP markers from SNP-array and genotyping-bysequencing (GBS) scored in the reference genome B73. Our results show that using mock genomes delivers estimates comparable to standard platforms when considering simple traits and additive effects. However, for complex traits and dominance effects, the estimates were slightly worse. We believe that these studies provide relevant knowledge for the phenotypic and genomic prediction applied to plant breeding.

Optimizing Genomic-Enabled Prediction in Small-Scale Maize Hybrid Breeding Programs: A Roadmap Review

“…In this context, we studied how SNP markers obtained from two genotyping platforms (616K SNP-array and GBS) affect the GP in our germplasm (Sabadin, 2020 ). We also attempted to verify the effect of the use of different reference genomes in SNP calls via GBS (i) using the most common reference genome, line B73 (GBS-B73), (ii) using a simulated reference genome built with GBS data, considering all inbred lines (GBS-Mock-All), and (iii) using a simulated reference genome built with GBS data from a single line, our heterotic pool tester L56 (GBS-Mock-L56).…”