TL;DR
This paper introduces a novel computational method that leverages heterogeneous graph convolutional networks (GCNs) and matrix completion to predict associations between microRNAs (miRNAs) and diseases. By constructing a rich network that integrates miRNA similarities, disease similarities, and known miRNA-disease relationships, the method effectively captures complex biological interactions, enabling accurate prediction of new associations, even for diseases with no prior known miRNA links.
What Problem Does This Paper Address?
Identifying the links between microRNAs (miRNAs) and diseases is crucial for understanding disease mechanisms, developing diagnostic tools, and designing personalized treatments. However, experimentally determining these associations is costly and time-consuming. While computational methods have emerged to address this, they face challenges: (1) the intricate nature of miRNA-disease relationships, requiring models that consider both local and global network influences, and (2) the critical need to predict associations for diseases where no prior miRNA links are known. This paper aims to overcome these limitations by proposing a more comprehensive and robust prediction framework.
Methods
The core of this work lies in a novel computational framework that combines Graph Convolutional Networks (GCNs) with matrix completion.
- Heterogeneous Network Construction: The method begins by constructing a heterogeneous network. This network integrates three key components:
- miRNA Similarity Network (RS): This network captures similarities between miRNAs based on their target associations, families, clusters, and known disease associations.
- Disease Similarity Network (DS): This network represents similarities between diseases, derived from shared gene functional information.
- miRNA-Disease Association Network (RD): This is the known, experimentally validated network of miRNA-disease links.
- Neighborhood Information Aggregation (GCNs): To capture the complex, multi-faceted relationships within the heterogeneous network, GCNs are employed. These networks aggregate information from a node's neighbors, propagating this information to intermediate and distant nodes. This allows the model to learn rich node embeddings that reflect both local and global network influences.
- Topology-Preserving Learning: The GCNs are trained to reconstruct the entire heterogeneous network. By minimizing the error in reconstructing this complex network, the model learns to preserve the topological information and relationships present in all three subnetworks (RS, DS, RD).
- Matrix Completion for Prediction: After learning the node embeddings, matrix completion techniques are used. The reconstructed heterogeneous network, particularly the miRNA-disease subnetwork, is used to predict unknown associations. By comparing the predicted associations with the original known associations, potential new links are identified.
Datasets
The study utilizes several publicly available datasets:
- HMDD database: Used to construct the known miRNA-disease association network (RD). The authors report using 6,441 associations between 577 miRNAs and 336 diseases after removing duplicates.
- HumanNet database: Used to derive disease functional similarity information for the disease similarity network (DS).
- miRTarBase: Used to obtain miRNA-target associations for constructing the miRNA similarity network (RS).
- miRBase: Used to access miRNA family and cluster information for the miRNA similarity network (RS).
The paper does not introduce a new benchmark dataset but rather leverages existing resources to validate its approach.
Key Insights and Ideas
- Holistic Network Representation: The integration of miRNA similarity, disease similarity, and known associations into a single heterogeneous network is crucial. This allows the model to leverage a broader spectrum of biological information than methods relying on single types of relationships.
- GCNs for Capturing Complex Interactions: The use of GCNs is a key innovation, enabling the model to effectively learn node embeddings by considering both direct and indirect neighborhood influences within the complex heterogeneous graph. This addresses the challenge of intricate miRNA-disease relationships.
- Topology Preservation is Key: Training the GCNs to reconstruct the entire heterogeneous network, rather than just the miRNA-disease subnetwork, is a critical insight. This ensures that the learned embeddings are robust and capture the underlying structure of all integrated data sources.
- Predicting the Unknown: The method's demonstrated ability to predict associations for diseases with no prior known links is a significant advancement, addressing a major limitation of many existing approaches.
Key Results
The proposed method demonstrates strong performance across various validation frameworks:
- Cross-Validation Performance: The method achieved impressive Area Under the Curve (AUC) scores:
- 5-fold cross-validation: 0.9616
- Global Leave-One-Out Cross-Validation (LOOCV): 0.9656
- Local LOOCV: 0.9532
- Leave-One-Disease-Out Cross-Validation (LODOCV): 0.9946
- Superiority over Baselines: These AUC values significantly outperformed several state-of-the-art baseline methods, as visualized in the ROC curves provided in Figure 2.
- Case Study Validation:
- For common diseases like breast, lung, and prostate neoplasms, the method accurately predicted top miRNAs, with 27-28 out of the top 30 predictions being verified by external databases (dbDEMC and PhenomiR).
- For diseases with no known associations (e.g., pancreatic neoplasms), the method successfully predicted candidate miRNAs, with 50 out of the top 50 predictions confirmed by at least one database.
Limitations and Future Work
The authors acknowledge the following limitations:
- Data Noise and Outliers: The datasets used to construct the network may contain noise and outliers, which could impact prediction accuracy.
- Network Insufficiency: The current heterogeneous network construction might not fully capture the extremely complex relationships between miRNAs and diseases.
Future work will focus on:
- Exploring Diverse Relationships: Investigating and incorporating a wider range of relationships between miRNAs and diseases.
- Improving Network Representation: Developing more sophisticated methods to represent the complex biological interactions within the network.
Summary updated