Predicting rice drought-responsive genes via distance-based prototypical graph neural network with path aggregation mechanism

Drought is a primary factor that adversely impacts rice yield and quality. Identifying drought-responsive genes is essential for developing drought-responsive cultivars. Recently, graph neural network methods based on embedding learning have shown considerable success in biological networks. However, challenges still remain in adequately capturing node attribute features, representing topological structures, and addressing class imbalance, which may constrain the model's predictive capability. To address these issues, we propose a distance-based prototypical graph neural network with path aggregation mechanism (DPGNNPAM) to mine drought-responsive genes in rice. First, we combine gene expression data and protein interaction networks in rice to construct graph-based datasets. Next, we utilize a random walk strategy to generate diverse walk paths and employ a recursive neural network-based path aggregator to encode node attributes along these paths. The prototypical network approach is subsequently employed during training to focus on global information and address the issue of sample imbalance. After that, the weighted similarity is computed by measuring the distance between the node embeddings and the class prototypes. Specifically, class prototypes are representative embeddings that capture the central characteristic of each class. Finally, we transform this value into a predictive probability using the softmax function. Experimental results demonstrate that DPGNNPAM outperforms traditional graph neural network algorithms in identifying drought-responsive genes in rice. Ultimately, we identify 17 candidate genes closely related to drought stress, 12 of which are confirmed in the literature as being involved in the plant's drought stress response.