Novel cis-trans regulatory networks for kiwifruit ripening uncovered with explainable deep neural network

Variations in cis-regulatory elements (CREs) resulting in diversification of gene expression have played central roles in establishing lineage-specific traits in plants. Even for an apparently identical trait, the responsible expression networks often evolved independently because of wide CRE variations in each lineage. Ripening has long been a focus for many varieties of fruit crops, including how regulatory networks might evolve in lineage-specific manners. In this study, with explainable deep learning (DL) models, we aimed to predict and model expression behaviors in the kiwifruit ripening process with the focus on CREs in gene promoters. For this, we applied a one-dimensional convolutional neural network (CNN) trained with 370 transcription factor (TF)-channeled CRE arrays of the promoter regions of whole genes in the kiwifruit genome to predict expression changes in kiwifruit ripening triggered by ethylene treatment. We achieved statistically significant predictions for ethylene-induced gene expression. Guided backpropagation of the trained CNN models for the genes with high-confidence predictions identified novel CREs relevant to the prediction of expression activation in kiwifruit ripening, including genes recognized by a bZIP TF family. Transient reporter assays and DNA affinity purification sequencing (DAP-Seq) analysis experimentally validated the CNN predictions, where the cis-trans interaction involving a bZIP TF was a key for gene activation in kiwifruit ripening. Comparative analysis with co-expression networking suggested that this AI-based method spotted the regulatory networks independent of co-expression patterns. Our results suggest that the application of explainable DL models provides a novel aspect to understand lineage-specific cis-trans regulatory networks.