Deep learning reveals regulatory logic of water stress adaptation in grapevine

Abstract

Predicting how sequence variants affect regulatory activity is critical for crop improvement, but confidence in these predictions, especially for diverse varieties, remains insufficient for breeding applications. Reference genomes collapse natural genetic diversity, creating reference bias that limits prediction accuracy across heterozygous crop genomes. We demonstrate the application of interpretable deep learning to a non-model perennial crop, Vitis vinifera (grapevine), and explore how layering computational evidence can build confidence in regulatory predictions.

We retrained the DeepCRE convolutional neural network on 163 public RNA-seq samples from water-stressed grapevine to predict genome-wide gene expression (high versus low) from flanking sequence. The model learns sequence-to-expression relationships, which we then interpret using gradient-based methods to extract the regulatory motifs driving predictions. Using chromosome-wise cross-validation to prevent homology inflation, the model achieved 91% accuracy and 92% F1 score. TF-MoDISco interpretation revealed 29 expression-predictive motifs organised into two functional meta-clusters.

Multiple lines of computational evidence support these predictions: genome-wide motif mapping shows enrichment consistent with expected expression patterns; gene ontology analysis reveals significant enrichment of co-occurring motifs in canonical drought response pathways including ABAsignalling, osmotic stress response, and flavonoid biosynthesis; integration with co-expression networks identifies candidate hub transcription factors. Comparative analysis between water stress and control models reveals patterns of regulatory rewiring under drought conditions.

This work explores how integrating multiple computational approaches can support confident prioritisation of regulatory elements and variants for targeted validation – a critical step toward explainable AI predictions for breeding applications in heterozygous crop genomes.