Improving genomic selection accuracy using a dual-path convolutional neural network framework: a terpenoid case study

Genomic selection (GS) accelerates forest tree early breeding. Using Litsea cubeba, a representative species of the Lauraceae family, as an example, this study aimed to uncover the genetic basis of terpenoid biosynthesis and to develop deep learning-based GS strategies for efficient genetic improvement. Whole-genome resequencing of 945 L. cubeba germplasms and GC-MS terpenoid quantification for 310 samples were done. Genome-wide association studies (GWAS) identified loci for key compounds (citral, geranial, and neral). To improve GS predictive ability, the PKDP deep learning model was developed: it uses parallel paths for distinct feature extraction from GWAS-identified loci and genome-wide markers, fusing them to combine prior knowledge with broad genomic information. GWAS identified 125 candidate genes for terpenoid biosynthesis, including two terpene synthase (TPS) gene clusters and MVA/MEP pathway enzymes. The PKDP model improved predictive ability for key terpenoids (citral, geranial, and neral) by 2-10% over traditional rrBLUP. Key genetic factors for L. cubeba terpenoid synthesis were identified, aiding molecular mechanism studies. By integrating multiscale genomic information, PKDP substantially enhanced GS prediction, providing an efficient tool for L. cubeba genetic improvement and innovative strategies for breeding complex traits in forest trees.