TL;DR
This paper introduces BRWLDA, a novel computational method that uses "bi-random walks" on a directed bi-relational network to predict associations between long non-coding RNAs (lncRNAs) and diseases. By accounting for the distinct structures of lncRNA and disease networks, as well as the hierarchical nature of diseases, BRWLDA significantly outperforms existing methods in identifying these crucial biological links.
What Problem Does This Paper Address?
Understanding the complex relationship between lncRNAs and human diseases is a critical area of biomedical research. While thousands of lncRNAs have been identified, only a fraction have been experimentally linked to specific diseases. Identifying these associations is vital for disease diagnosis, prognosis, and treatment. However, experimental validation is costly and time-consuming. Computational models aim to bridge this gap, but many existing approaches fail to adequately consider the inherent structural differences between lncRNA interaction networks and disease hierarchical structures. They also often overlook the pattern that newly discovered disease associations for a lncRNA tend to be more specific subtypes of diseases already linked to it. This paper tackles the challenge of developing a more accurate and robust computational method that addresses these limitations.
Methods
The core innovation of this paper is the BRWLDA (Bi-Random Walks for predicting lncRNA-Disease Associations) model. It constructs a directed bi-relational network that integrates three key components:
- lncRNA Functional Similarity Network: This network captures how similar lncRNAs are based on their functional roles. It's built using multiple data sources, including lncRNA-miRNA interactions, miRNA-disease associations, and lncRNA-gene functional associations.
- Disease Similarity Network: This network leverages the Disease Ontology (DO), a hierarchical classification of diseases. This allows the model to understand the relationships between diseases (e.g., a specific cancer being a subtype of a broader category).
- Known lncRNA-Disease Associations: This forms the initial links between the lncRNA and disease subnetworks.
The key algorithmic novelty is the bi-random walk approach. Unlike traditional random walks that treat the entire network uniformly, BRWLDA employs two distinct random walkers:
- One starts in the lncRNA subnetwork and walks towards the disease subnetwork.
- The other starts in the disease subnetwork and walks towards the lncRNA subnetwork.
These walks are "bi-directional" and account for the directed nature of the disease hierarchy and the functional similarity within the lncRNA network. The model also incorporates a restart probability to prevent walkers from getting lost in deep parts of the network and to ensure they can return to known associations. This bi-random walk strategy allows BRWLDA to explore potential associations by considering both the functional similarity of lncRNAs and the hierarchical relationships of diseases, and importantly, it can predict associations even when a lncRNA or disease has no prior known links.
Datasets
The study utilizes several publicly available biological databases:
- starBase v2.0: For lncRNA-miRNA interactions.
- HMDD (Human microRNA and Disease Database): For miRNA-disease associations.
- GeneRIFs (Gene Reference into Function): For lncRNA-gene functional associations.
- LncRNADisease database: For known lncRNA-disease associations.
- Disease Ontology (DO): For constructing the disease hierarchical network.
The paper also mentions using data from GeneRIFs and LncRNADisease for experimental verification and evaluation.
Key Insights and Ideas
- Structural Differences Matter: The paper strongly argues that treating lncRNA and disease networks as homogenous is a significant limitation. BRWLDA's success stems from explicitly modeling the distinct structures: functional similarity for lncRNAs and hierarchical relationships for diseases.
- Hierarchical Disease Information is Crucial: Leveraging the Disease Ontology is not just about grouping diseases; it's about understanding the "pattern of newly discovered associations." The model capitalizes on the observation that new lncRNA-disease links often involve more specific disease subtypes.
- Bi-directional Exploration is Powerful: The bi-random walk approach, with walkers moving between the lncRNA and disease subnetworks, allows for a more comprehensive exploration of potential associations than a single, global random walk. This is particularly effective for identifying novel links.
- Predicting the Unknown: A key strength is BRWLDA's ability to predict associations for lncRNAs or diseases with no prior known links, making it highly valuable for discovery.
Key Results
- Superior Performance in LOOCV: BRWLDA achieved the highest Area Under the Curve (AUC) values in both lncRNA-oriented (0.7952) and disease-oriented (0.7940) Leave-One-Out Cross-Validation (LOOCV) compared to five other state-of-the-art methods.
- Robustness in Simulated Experiments: In a more challenging scenario where known associations were masked, BRWLDA consistently outperformed other methods, achieving AUCs of 0.9888, 0.9548, and 0.9275 for masking 1, 3, and 5 diseases, respectively.
- Effective Prediction of Novel Associations: When tested on diseases with completely unknown associated lncRNAs, BRWLDA identified more confirmed associations than its competitors.
- Case Study Validation: Manual literature mining confirmed that a significant portion of the top-ranked predicted lncRNAs by BRWLDA for breast, colon, and lung cancers were indeed associated with these diseases, demonstrating its practical utility.
Limitations and Future Work
- Data Sparsity and Noise: The authors acknowledge that the available lncRNA-miRNA, miRNA-disease, and lncRNA-gene functional associations are sparse and may contain noise, which can impact performance.
- Empirical Estimation of Transition Probabilities: The semantic similarity between diseases is empirically estimated, and further improvements could be made with more accurate estimations and a more refined Disease Ontology.
- Integration of More Data: The paper suggests that incorporating additional biological information, such as amino acid sequences, transcription information, and other interaction data (e.g., protein-protein interactions), could further enhance BRWLDA's performance.
- Modeling Complex Diseases: The authors propose that modeling all related biological molecules (miRNAs, genes, proteins) within a complex network could provide a more comprehensive understanding of disease mechanisms.
Summary updated