TL;DR
This paper introduces GIMDA, a novel computational model that predicts potential associations between microRNAs (miRNAs) and diseases. By leveraging "graphlet interactions" to capture complex relationships within biological networks, GIMDA achieves high accuracy in predicting these associations, outperforming existing methods.
What Problem Does This Paper Address?
MicroRNAs (miRNAs) are small non-coding RNA molecules that play crucial roles in regulating gene expression and are increasingly implicated in various human diseases. Identifying the specific miRNAs associated with particular diseases is vital for understanding disease mechanisms and developing diagnostic or therapeutic strategies. However, experimentally determining these associations is a time-consuming and expensive process. Therefore, there's a pressing need for accurate and efficient computational models to predict these latent miRNA-disease links, guiding experimental research. Existing computational methods often rely on direct network neighbors or miRNA-target interactions, which can be prone to errors. This paper aims to address this by developing a more robust prediction model that captures intricate relationships within biological networks.
Methods
The core innovation of GIMDA lies in its use of graphlet interactions to model relationships between miRNAs and between diseases. Here's a breakdown of the approach:
- Graph Representation: Both miRNAs and diseases are represented as nodes in separate graphs (GM for miRNAs, GD for diseases). Edges connect nodes that exhibit similarity.
- Similarity Measures:
- MiRNA Functional Similarity: Utilizes pre-computed similarity scores for miRNAs.
- Disease Semantic Similarity: Employs two models (SS1 and SS2) based on Directed Acyclic Graphs (DAGs) to quantify disease similarity, considering the hierarchical relationships and specificity of disease terms.
- Gaussian Interaction Profile (GIP) Kernel Similarity: This standard technique measures similarity based on the "interaction profiles" of miRNAs and diseases, essentially looking at which other entities they are associated with.
- Integrated Similarity: The functional/semantic similarities are combined with GIP kernel similarities to create a more comprehensive measure of similarity for both miRNAs (SM) and diseases (SD).
- Graphlet Interaction: This is the novel component. Graphlets are small, connected subgraphs (up to 4 nodes in this study). The "graphlet interaction" quantifies the relationship between two nodes (miRNAs or diseases) by analyzing their participation in various graphlet structures. Specifically, they count the occurrences of 28 different "graphlet interaction isomers" (defined by automorphism orbits) between pairs of nodes. This allows GIMDA to capture complex, indirect relationships beyond simple direct connections.
- Prediction Score: The model calculates an association score for each miRNA-disease pair by combining scores derived from the miRNA graph (Sm) and the disease graph (Sd). These scores are computed using the graphlet interaction counts and learned weights.
- Training: The model is trained using known miRNA-disease associations to learn the weights for the graphlet interactions.
Datasets
- HMDD V2.0: This is the primary dataset used for training and evaluating the GIMDA model. It contains experimentally confirmed human miRNA-disease associations.
- dbDEMC and miR2Disease: These databases were used for validating the predictions made by GIMDA in case studies.
- HMDD V1.0: Used for an additional case study to assess the model's performance on a different version of the known association data.
Key Insights and Ideas
- Graphlets Capture Complex Relationships: The core idea is that analyzing how nodes participate in small subgraph patterns (graphlets) can reveal deeper, indirect relationships that simple similarity measures might miss. This is particularly relevant in biological networks where interactions are often multi-layered.
- Beyond Direct Links: By focusing on graphlet interactions, GIMDA moves beyond relying solely on direct connections or immediate neighbors, allowing for a more nuanced understanding of how miRNAs and diseases are linked.
- Integrated Similarity is Powerful: Combining multiple similarity measures (functional, semantic, and GIP) provides a richer representation of the entities, leading to more accurate predictions.
- Robustness Across Datasets: The model performs well even when tested on an older version of the HMDD database (V1.0), suggesting its generalizability.
Key Results
- High Predictive Accuracy: GIMDA achieved an Area Under the Curve (AUC) of 0.9006 in global leave-one-out cross-validation (LOOCV) and 0.8455 in local LOOCV. The average AUC for five-fold cross-validation was 0.8927 ± 0.0012.
- Superior Performance: GIMDA's AUC values were superior to several previously published models (HGIMDA, RLSMDA, HDMP, WBSMDA, RWRMDA, MCMDA) in both global and local LOOCV.
- Successful Case Studies:
- For colon neoplasms, kidney neoplasms, and prostate neoplasms, GIMDA correctly predicted 45, 45, and 41 of the top 50 potential miRNAs, respectively, which were validated by dbDEMC and miR2Disease.
- In a case study for new diseases (hepatocellular carcinoma) without prior known associations, 50 out of the top 50 predicted miRNAs were validated.
- Robustness: Twofold, threefold, and fourfold cross-validations also demonstrated the model's robustness.
Limitations and Future Work
- Data Sufficiency: The authors acknowledge that the number of known miRNA-disease associations is still limited. They anticipate that as the dataset expands, the model's performance will improve.
- Graphlet Size: GIMDA currently considers graphlets with up to four nodes. This means it might miss complex indirect relationships involving more than three intermediate links. Future work could explore larger graphlets if computational resources allow.
- New miRNA-New Disease Prediction: The current model is "helpless" for predicting associations between entirely new miRNAs and entirely new diseases that have no prior associations in the training data.
- Computational Cost: Counting graphlet interactions for 28 isomers across all nodes can be computationally intensive.
- Future Directions: The authors suggest that future work could focus on developing more rational methods for evaluating model performance and exploring the use of graphlet or network motif concepts for more effective miRNA-disease association prediction. They also highlight the potential for future models to incorporate more nodes into graphlet analysis.
Summary updated