PTGCDA: a PU-learning-driven Transformer and GraphSAGE framework for circRNA-disease association prediction

Abstract Motivation Circular RNAs play key regulatory roles in many human diseases, making accurate prediction of circRNA-disease associations essential for understanding diseases and discovering biomarkers. Existing methods are limited by insufficient sequence utilization, network sparsity, and noisy negative samples from unlabeled data. Results We propose PTGCDA, a computational framework that integrates sequence representation learning, graph representation learning, Transformer, and positive-unlabeled learning for circRNA-disease association prediction. PTGCDA extracts circRNA sequence features using k-mer embeddings, Word2Vec, and TextCNN; GraphSAGE learns node representations of circRNAs and diseases by aggregating neighborhood information from the heterogeneous network; and multi-head attention in Transformer further captures interactions between circRNA and disease representations. Meanwhile, a two-step positive-unlabeled learning strategy identifies reliable negative samples to reduce label noise during training. Five-fold cross-validation on four benchmark datasets and one integrated dataset demonstrated that PTGCDA outperformed existing methods, achieving an AUROC of 0.9202 on Dataset4. De novo prediction on the same dataset achieved an AUROC of 0.8646 and an AUPR of 0.1874, while ablation and case studies further supported the effectiveness of PTGCDA and its ability to identify potential circRNA-disease associations. Availability and implementation The PTGCDA code is freely available at https://github.com/WANGSicong1113/PTGCDA.