Direction-aware graph attention modeling and co-expression integration reveal novel core regulators of shared drought...

Although differential expression analysis has identified numerous drought- and salinity-responsive genes in rice, it cannot resolve the directional, transcription factor (TF)-driven regulatory architecture underlying responses across independent drought and salinity datasets. To address this, we applied TF-Anchored Directed SiGaResNet, a signed, direction-aware graph attention residual network with a TF-anchored edge-retention rule and permutation-based validation, to RNA-seq profiles from 92 rice leaf samples (51 stress, 41 control) derived from independent drought and salinity studies. Unlike conventional correlation- or co-expression-based approaches, the framework constrains regulatory inference around TF nodes while learning directed, signed relationships through supervised graph representation learning. Differential expression analysis identified 719 genes (405 upregulated, 314 downregulated), with induced transcripts enriched in stress-response and hormone-signaling processes and repressed transcripts dominated by photosynthesis- and secondary-metabolism-related processes. The framework showed strong cross-validated classification performance (MCC = 0.78 ± 0.17; AUC = 0.95 ± 0.05). The resulting directed gene regulatory network comprised 137 nodes and 163 edges and exhibited a sparse, modular architecture (modularity = 0.77; 15 communities), with TFs significantly enriched among regulatory source nodes (P < 0.001). Hub regulators were distributed across multiple TF families, predominantly bZIP, MYB-related, and NAC. Eighteen of thirty-two prioritized TF-encoding genes were independently corroborated within stress-associated WGCNA modules, and four comparatively uncharacterized candidates (Os07g0176200, Os05g0569300, Os03g0231000, and Os02g0187700) received convergent computational support from promoter cis-element enrichment, protein functional-association topology, and leave-one-out cross-validation classification. Quantitative real-time PCR further confirmed stress-responsive expression of all four candidates under separate drought and salinity treatments, supporting their prioritization as putative candidate regulators. These findings indicate that direction- and sign-aware graph attention modeling can integrate transcriptomic information from independent drought and salinity studies to identify biologically coherent, TF-centered regulatory patterns and prioritize candidate regulators for further functional investigation, without establishing causal regulatory mechanisms.