Deciphering the sequence basis and application of transcriptional initiation regulation in plant genomes through deep learning

Transcription initiation is a key checkpoint in plant gene regulation, yet the DNA features that determine where and the frequency of the genes start transcription remain unclear. We develop GenoRetriever, an interpretable deep learning model trained on base pair resolution STRIPE-seq data from multiple crop genomes, to systematically reveal and quantify the sequence code that governs transcription start sites (TSSs). Using TSS profiles from 16 soybean tissues and six additional crops, GenoRetriever identifies 27 core promoter motifs, including canonical TATA box and initiator elements, that together dictate TSS choice and activity. Model interpretation shows how each motif modulates both initiation frequency and precise start site position; these effects are confirmed by in silico motif edits, saturation mutagenesis, and targeted promoter assays. A new telomere-to-telomere assembly of wild soybean, Glycine soja, reveals that 31.85% of natural promoter variants shift dominant motifs relative to cultivated soybean, explaining domestication-driven changes in transcriptional regulation. Cross-species comparisons further indicate that, although many motif functions are conserved, monocots and dicots display distinct motif frequencies and positional preferences. GenoRetriever provides an interpretable, cross species framework for decoding transcription initiation in plants. By linking specific sequence motifs to quantitative transcriptional outcomes and validating these links experimentally, our study advances fundamental knowledge of promoter architecture and supplies a practical platform for rational engineering of gene expression in crop improvement and functional genomics.