CGAN-SAE: A Gated Axial-Attention Enhanced Framework for Identifying Key Disease Resistance Genes in Rice During Magnaporthe oryzae Infection with Limited Samples
Yinfei Dai, Jie Fan, Shihao Lu, Mengjiao Qiao, Yuheng Zhu, Benrui Wang, Rongyuan Liu, Yubao Liu, Hao ZhangIntroduction:
Rice blast, caused by the fungal pathogen Magnaporthe oryzae, poses a significant threat to global rice production. Conventional methods for identifying resistanceassociated genes are often laborious and time-consuming. Machine learning has proven effective in gene prioritization for complex traits, especially in human genomics, yet its application in plant immunity remains limited. To address this gap, we present CGAN-SAE, a novel deep learning framework that integrates a Conditional Generative Adversarial Network (CGAN) with a Sparse Autoencoder (SAE) to predict candidate genes involved in rice immune responses. The approach enables data-efficient and interpretable modeling, offering a powerful strategy to accelerate the development of blast-resistant rice varieties.
Methods:
The CGAN-SAE framework combines a Conditional Generative Adversarial Network (CGAN) to generate biologically plausible synthetic gene expression profiles, thereby alleviating limitations imposed by small sample sizes, with a Gated Axial-Attention-enhanced Stacked Autoencoder (SAE) for robust feature extraction and global pattern recognition. Principal Component Analysis (PCA) is then applied to interpret the learned latent features and evaluate gene-level importance.
Results:
Transcriptomic analysis identified 891 differentially expressed genes (DEGs). Using CGAN-SAE, we prioritized five high-confidence candidate genes potentially associated with disease resistance. The model achieved an AUC of 0.80, accuracy of 0.75, F1-score of 0.77, precision of 0.83, and recall of 0.71, outperforming multiple baseline and state-of-the-art models. These results demonstrate its superior ability to extract meaningful biological signals from limited transcriptomic data and identify key regulatory factors.
Discussion:
CGAN-SAE represents a significant advancement in deep learning-based gene prioritization under data-scarce conditions. Although performance depends on the quality of input DEGs and may require species-specific tuning, the framework shows strong potential for extension to other crops such as maize and soybean, and even to human disease genomics. Future integration of protein- protein interaction networks and multi-omics data could further improve its biological interpretability and broad applicability across host-pathogen systems.
Conclusions:
This study demonstrates that deep learning–driven gene prioritization is feasible and effective in plant disease resistance research, offering a scalable framework for candidate gene discovery under limited data conditions.