Prediction of circRNA-Disease Associations via Graph Isomorphism Transformer and Dual-Stream Neural Predictor

Abstract

Circular RNAs (circRNAs) have attracted increasing attention for their roles in human diseases, making the prediction of circRNA-disease associations (CDAs) a critical research area for advancing disease diagnosis and treatment. However, traditional experimental methods for exploring CDAs are time-consuming and resource-intensive, while existing computational models often struggle with the sparsity of CDA data and fail to uncover potential associations effectively. To address these challenges, we propose a novel CDA prediction method named the Graph Isomorphism Transformer with Dual-Stream Neural Predictor (GIT-DSP), which leverages knowledge graph technology to address data sparsity and predict CDAs more effectively. Specifically, the model incorporates multiple associations between circRNAs, diseases, and other non-coding RNAs (e.g., lncRNAs, and miRNAs) to construct a multi-source heterogeneous knowledge graph, thereby expanding the scope of CDA exploration. Subsequently, a Graph Isomorphism Transformer model is proposed to fully exploit both local and global association information within the knowledge graph, enabling deeper insights into potential CDAs. Furthermore, a Dual-Stream Neural Predictor is introduced to accurately predict complex circRNA-disease associations in the knowledge graph by integrating dual-stream predictive features. Experimental results demonstrate that GIT-DSP outperforms existing state-of-the-art models, offering valuable insights for precision medicine and disease-related research.

Keywords: circRNA–disease associations; graph isomorphism network; knowledge representation learning; transformer.

Figure 1

Flowchart of GIT-DSP. (A) Construct the CDA knowledge graph with multi-source heterogeneous datasets. (B) Construct a fused similarity network based on the association matrix. (C) Generate high-quality embeddings through attention-based information propagation with the Graph Isomorphism Transformer. (D) Predict CDAs.

Figure 2

Visualization of datasets. (A–C) represent the association distributions of Dataset 1, Dataset 2, and Dataset 3, respectively. This contains five associations (lnc-dis (lncRNA–disease), lnc-mi (lncRNA–miRNA), mi-dis (miRNA–disease), circ-mi (circRNA–miRNA), and circ-dis (circRNA–disease)).

Figure 3

Association between CircRNA and diseases. (A) The horizontal axis indicates the circRNA count, while the vertical axis shows the disease count per degree. (B) Sparsity of some circRNA–disease associations.

Figure 4

Overview of the knowledge graph. It encompasses four types of entities: disease, circRNA, lncRNA, and miRNA, as well as five different types of relationships: miRNA–circRNA, miRNA–lncRNA, disease–miRNA, disease–lncRNA, and circRNA–disease.

Figure 5

Ranking of the expressive power of Sum, Mean, and Max on multisets. The input part represents the aggregated neighborhood network; Sum learns the entire multiset, Mean learns the overall distribution, and Max ignores redundant information.

Figure 6

The principle of graph isomorphism network. (A) Weighted self-loops are added to the nodes to ensure they retain their unique feature information during aggregation. (B) Exchange the neighboring nodes with weighted self-loops. (C) The corresponding parts from (B) are aggregated to update each node’s representation.

Figure 7

Performance comparison of AUC (A) and AUPR (B) on Dataset 1. The compared models include information-propagation-based models (KATZHCDA, RWR, and CD-LNLP), traditional machine learning models (RWR-KNN, ICIRCDA, and RNMFLP), and deep learning models (DMFCDA, GMNN2CD, and KGETCDA).

Figure 8

Performance comparison of AUC (A) and AUPR (B) on Dataset 2. The compared models include information-propagation-based models (KATZHCDA, RWR, and CD-LNLP), traditional machine learning models (RWR-KNN, ICIRCDA, and RNMFLP), and deep learning models (DMFCDA, GMNN2CD, and KGETCDA).

Figure 9

Performance comparison of AUC (A) and AUPR (B) on Dataset 3. The compared models include information-propagation-based models (KATZHCDA, RWR, and CD-LNLP), traditional machine learning models (RWR-KNN, ICIRCDA, and RNMFLP), and deep learning models (DMFCDA, GMNN2CD, and KGETCDA).

Figure 10

Comparison of the average number of accurately identified associations for Dataset 1 (A), Dataset 2 (B), and Dataset 3 (C). The compared models include information-propagation-based models (KATZHCDA, RWR, and CD-LNLP), traditional machine learning models (RWR-KNN, ICIRCDA, and RNMFLP), and deep learning models (DMFCDA, GMNN2CD, and KGETCDA).

Figure 11

Violin plots comparing the performance of different aggregation and prediction modules. (A) AUC values for Dataset 1; (B) AUC values for Dataset 2; (C) AUC values for Dataset 3. It covers GIN+MLP, GIN+SMLP, and GIN+DSP(Ours) based on GIN, GCN+DSP based on GCN, BI-Interaction+DSP based on Bi-Interaction, and GraphSage+DSP based on GraphSage.

Figure 12

Comparison of methods for neighborhood feature calculation. (A,B) represent the AUC and AUPR values on the three datasets, respectively.

Figure 13

Performance comparison of different attention heads and layers. (A) AUC heat map for Dataset 1; (B) AUC heat map for Dataset 2.