Deep learning based genomic breeding of pest-resistant grapevine

Crop pests have profoundly deleterious effects on crop yield and food security. However, conventional pest control depends heavily on the utilization of insecticides, which develops strong pesticide resistance and concerns of food safety. Crop and their wild relatives display diverse levels of pest resistance, indicating the feasibility for breeding of pest-resistant crop varieties. In this study, we integrate deep learning (DL)/machine learning (ML) algorithms, plant phenomics and whole genome sequencing (WGS) data to conduct genomic selection (GS) of pest-resistance in grapevine. We employ deep convolutional neural networks (DCNN) to accurately calculate the severity of damage by pests on grape leaves, which achieves a classification accuracy of 95.3% (Visual Geometry Group 16, VGG16, for binary trait) and a correlation coefficient of 0.94 in regression analysis (DCNN with Pest Damage Score, DCNN-PDS, for continuous trait). We apply DL models to predict and integrate phenotype (both binary and continuous) along with WGS data from 231 grape accessions, conducting Genome-Wide Association Studies (GWAS). This analysis detects a total of 69 QTLs, encompassing 139 candidate genes involved in pathways associated with pest resistance, including jasmonic acid (JA), salicylic acid (SA), ethylene, and other related pathways. Furthermore, through the combination with transcriptome data, we identify specific pest-resistant genes, such as ACA12 and CRK3, which play distinct roles in resisting herbivore attacks. Machine learning-based GS demonstrates a high accuracy (95.7%) and a strong correlation (0.90) in predicting the leaf area damaged by pests as binary and continuous traits in grapevine, respectively. In general, our study highlights the power of DL/ML in plant phenomics and GS, facilitating genomic breeding of pest-resistant grapevine. ### Competing Interest Statement The authors have declared no competing interest.