#34 · Genomics

Deep networks map genome-wide SNPs to quantitative phenotypes in grapevine and plants

The dominant thread is the use of deep architectures to map genome-wide SNP genotypes onto quantitative phenotypes, with each item adding a mechanism aimed at the high-dimensionality and generalization problems of genomic prediction. The LiT-G2P hybrid combines a linear-Transformer architecture to predict grapevine leaf hair and trichome density from SNPs, reporting improved cross-year robustness. The proximal residual network applies convex and non-convex regularizers within a stochastic adaptive proximal gradient framework to control overfitting on high-dimensional genomic data. CLCNet pairs contrastive representation learning with chromosome-aware SNP feature selection, exploiting the genomic organization of markers. What is new is the explicit encoding of structure, whether through regularization, chromosome partitioning, or contrastive objectives, rather than treating SNPs as an undifferentiated input vector. The residual network is framed for genomic data generally rather than for a specific crop.