gsGNA: A Universal Pipeline for Mining Plant Stress-Resistance Regulators

Technical Summary of "gsGNA: A Universal Pipeline for Mining Plant Stress-Resistance Regulators"

Problem Statement
Global warming has intensified abiotic stresses such as drought, alkalinity, and cold, severely impacting rice productivity. While individual studies have explored gene regulatory networks (GRNs) under specific stress conditions, there is a lack of systematic, multi-stress investigations from a systems biology perspective. Existing GRN inference methods face significant challenges: single algorithms often suffer from inherent biases (e.g., overlooking weak signals or overestimating indirect relationships), and deep learning models frequently struggle with generalization on small sample sizes. Furthermore, existing ensemble strategies are often species-specific or rely on supervised learning paradigms that require gold-standard datasets, which are scarce in rice stress biology. Consequently, there is an urgent need for a robust, universal computational pipeline capable of integrating diverse algorithms to construct high-confidence GRNs for rice under multiple abiotic stresses.

Methodology
The authors propose gsGNA (General Stress-resistance Gene Regulatory Network), a universal pipeline designed to identify plant stress-resistance regulators using publicly available RNA-seq data. The methodology proceeds through the following stages:

1. Data Integration: The pipeline utilizes rice leaf RNA-seq datasets under aridity, alkaline, and cold stress conditions. Transcription factor (TF) lists are sourced from PlantRegMap/PlantTFDB, while ChIP-seq data and TF interference experimental data serve as ground truth for validation.
2. Base GRN Inference: Six distinct inference algorithms are employed as base models: two R-based (GENIE3, KBoost) and four Python-based (GRNBoost2, 3DCEMA, DeepRIG, IGEGRNS). These models generate initial regulatory networks using normalized gene expression matrices.
3. Ensemble Fusion Strategy: To overcome the limitations of individual algorithms, six ensemble strategies (AdaBoost, Bagging, Stacking, XGBoost, hard voting, and quantile rank fusion) are tested to integrate the feature weights of the base GRNs.
4. Optimization and Selection: The performance of these ensemble strategies is evaluated using ChIP-seq data, TF binding site (TFBS) data, and differentially expressed gene (DEG) data from TF interference experiments. Metrics include AUROC and AUPR, assessed via a scenario ranking scoring system and permutation null models.
5. Network Analysis: The optimal ensemble GRNs are analyzed for topological features (e.g., scale-free distribution, small-world properties) and functional modules using network clustering algorithms.
6. Regulator Mining: Key TFs are identified based on degree centrality. A "reverse screening" strategy is employed, targeting core genes in conserved pathways (e.g., MAPK signaling, hormone transduction) to predict upstream regulators, distinguishing between known and potentially novel TFs.

Key Contributions and Results

• Optimal Ensemble Strategy: Comparative analysis revealed that the XGBoost ensemble strategy consistently outperformed other methods (including Stacking and Bagging) across all three stress conditions. XGBoost achieved the highest cumulative ranking scores and global average AUROC/AUPR values. Crucially, ablation experiments demonstrated that the optimal performance was not achieved by stacking all algorithms but by selecting complementary subsets tailored to specific stress types (e.g., 3DCEMA and GRNBoost2 were core components across all stresses).
• Robustness and Generalization: The XGBoost-based ensemble GRNs demonstrated superior stability compared to single algorithms. In external validation using independent biotic stress datasets (Magnaporthe oryzae, Nephotettix cincticeps, and Rice stripe virus), XGBoost achieved the highest AUROC values (up to 0.94), confirming its generalization capability beyond the training abiotic stress conditions.
• Biological Plausibility: The constructed GRNs exhibited typical biological topological features, including power-law degree distributions (scale-free) and small-world characteristics (high clustering, short path lengths), distinguishing them from random networks. Validation against TFBS data and TF interference DEGs confirmed significant enrichment of predicted regulatory pairs, establishing the networks' biological relevance.
• Identification of Core and Novel Regulators:
  • Shared Core TFs: The analysis identified 20 core transcription factors shared across aridity, alkaline, and cold stresses. These TFs and their targets were enriched in key pathways such as plant hormone signal transduction and the MAPK signaling pathway.
  • Network Architecture: The study revealed an organizational pattern of "conserved core and divergent periphery." While the overall network topology is highly conserved across stresses, functional modules show significant divergence, reflecting stress-specific adaptive responses.
  • Novel Predictions: Through a reverse screening strategy targeting core pathway genes, the pipeline predicted 45 potentially novel regulators. These TFs were present under multiple stress conditions but lacked previously reported stress-related functions, offering new candidates for experimental validation.

Significance
The paper claims that the gsGNA pipeline provides a theoretical foundation and a valuable target repository for rice stress-tolerance breeding. By systematically mapping the GRN landscape of rice under common abiotic stresses, the study elucidates the "conserved core and divergent periphery" organizational pattern of plant stress responses. The identification of 20 core TFs and 45 novel candidates offers a resource for dissecting the molecular mechanisms of abiotic stress adaptation. The authors position gsGNA not only as a tool for rice but as a reusable computational framework for plant stress GRN research, addressing the scarcity of gold-standard datasets through a multi-algorithm fusion approach that balances accuracy with robustness.