Rice Drought Mutation Designer (RDMD): A Deep Learning–Guided Framework for De Novo Engineering of Drought-Responsive Cis-Regulatory Elements
In this study, we present RDMD, an integrated bioinformatics pipeline that combines ATAC-seq data, convolutional neural network (CNN)–based chromatin-openness scoring, motif interpretation, and heuristic mutagenesis to engineer drought-responsive regulatory sequences in rice. First, public rice root ATAC-seq datasets under drought and control conditions were used to train a CNN that distinguishes open chromatin regions specific to water-deficit stress. The model achieved high discrimination performance (AUC-ROC 0.99 training, 0.94 testing) on 240 bp one-hot–encoded inputs. Next, we interpreted the first-layer convolutional filters as position weight matrices (PWMs), annotating 51 of 300 filters against the JASPAR Plants database via Tomtom. Hierarchical clustering of these motifs revealed clusters enriched for known drought-related elements (e.g., bZIP910, MYB15) alongside uncharacterized motifs, pointing to novel regulatory patterns. We then developed a heuristic mutation algorithm that introduces 3–5 clustered base substitutions within a 6 bp window to maximize the CNN’s predicted openness score. Applying RDMD to the upstream region of the drought-responsive DRG9 gene, we identified a DOF TF binding site–creating mutation (ATTAA → TTTCA) in window 6 that boosts the drought-openness score from 0.656 to 0.829. This mutation aligns with known OsDof12–mediated activation of phenylpropanoid-pathway genes to enhance drought tolerance. Overall, RDMD enables de novo design of stress-responsive cis-elements, offering a versatile strategy for engineering improved abiotic stress resilience in crops.