VGNPred: Improving ncRNA family prediction using variational graph neural network with diffusion prior
Ang Li, Yuanning Liu, Zhaoyang Zhang, Zhiyong Zhou, Shaoqiang Zhang, Dawei Lin · Expert Systems with Applications · 2026
Non-coding RNA (ncRNA) is a class of RNA molecules that are transcribed from the genome but do not directly encode proteins. They play important roles in many biological processes, such as gene expression regulation, cell structure maintenance, and cell signaling. Family prediction of ncRNA is important for understanding gene regulatory mechanisms, promoting genomics research, and supporting disease diagnosis and treatment. Existing methods based on sequence comparison or unimodal feature learning may be limited in capturing comprehensive sequence-structure representations, especially for complex non-nested structural patterns such as pseudoknots and non-canonical pairings, which can affect their performance in ncRNA family classification tasks. In this paper, we propose an ncRNA family prediction method called VGNPred. The proposed framework uses a BERT model pre-trained on large-scale ncRNA data to extract sequence features. A variational graph neural network with diffusion prior, termed VGAE-DP, is introduced to extract structural features of ncRNA and to provide a more flexible latent prior for structural representation learning. To reduce the representation gap between sequence and structure modalities, a cross-modal prompt interaction module is employed. In addition, a modality balance learning strategy is used to dynamically adjust the optimization process of different modalities during training. Finally, sequence and structural features are integrated for ncRNA family prediction. Experiments on the NCY and nRC datasets show that VGNPred achieves competitive performance compared with representative ncRNA family prediction methods and obtains consistent improvements under the adopted evaluation protocol. The code for this research is available at https://github.com/Endav99/VGNPred .