TL;DR
This paper introduces a novel method using sparse canonical correlation analysis (SCCA) to predict associations between microRNAs (miRNAs) and diseases. Beyond prediction, the method aims to uncover the underlying molecular mechanisms by identifying correlated sets of miRNA target genes and diseases, offering valuable biological insights for further research.
What Problem Does This Paper Address?
MicroRNAs (miRNAs) are small non-coding RNAs that play crucial roles in regulating gene expression and are implicated in the development and progression of many human diseases. While numerous computational methods exist to predict miRNA-disease associations, they often fall short in explaining why these associations exist. This lack of mechanistic understanding hinders the translation of these predictions into actionable biological experiments. Existing methods either focus on functional similarity between miRNAs or rely on predicted miRNA-target gene interactions, which can suffer from high false positive rates. This paper aims to bridge this gap by providing a method that not only predicts associations but also offers a biologically interpretable explanation for them.
Methods
The core of this paper's methodology lies in Sparse Canonical Correlation Analysis (SCCA). The authors first construct "target gene profiles" and "disease profiles" for miRNAs based on experimentally validated miRNA-gene interactions and miRNA-disease associations from existing databases.
Here's a breakdown of the approach:
- Data Preparation: They gather data from miRTarBase (for miRNA-gene interactions) and HMDD v2.0 (for miRNA-disease associations), focusing on human data and experimentally validated evidence.
- Profile Construction: For each miRNA, they create two vectors: one representing the presence/absence of its target genes and another representing the presence/absence of diseases associated with it.
- Canonical Correlation Analysis (CCA): Standard CCA aims to find linear combinations of two sets of variables that maximize their correlation. However, standard CCA often produces non-sparse weight vectors, making interpretation difficult.
- Sparse CCA (SCCA): To address this, the authors employ a penalized matrix decomposition (PMD) strategy on the matrix formed by the two profiles. This introduces sparsity constraints, leading to more interpretable weight vectors. This allows them to extract multiple "canonical components" (CCs), each representing a correlated set of genes and diseases.
- Prediction: For a new miRNA with a known target gene profile, its disease profile is predicted by combining the weight vectors from the extracted CCs.
Datasets
The study utilizes two primary datasets:
- miRTarBase: Used for experimentally validated miRNA-gene interactions.
- HMDD v2.0: Used for experimentally supported human miRNA-disease associations.
The benchmark dataset for training and cross-validation consists of 404 miRNAs, 2796 target genes, and 362 diseases, with 7999 miRNA-gene interactions and 5117 miRNA-disease associations. The authors also use the newly released HMDD v3.0 to validate their predictions.
Key Insights and Ideas
- Mechanistic Interpretation: The most significant insight is the ability of SCCA to extract correlated sets of genes and diseases. These sets provide a plausible explanation for how a miRNA might influence a disease by linking its target genes to disease outcomes.
- Sparsity for Interpretability: The use of sparse CCA is crucial. It forces the model to select a smaller, more focused set of genes and diseases within each component, making the identified relationships more interpretable and less noisy.
- Cause-and-Effect Hypothesis: The extracted gene-disease sets can be viewed as potential molecular pathways. Diseases within a set are suggested to be outcomes of miRNA perturbations on the target genes that also appear in the same set.
- Validation through Existing Knowledge: The method's predictions are supported by existing literature and databases, reinforcing the biological relevance of the identified associations.
Key Results
- Performance: The method achieved an Area Under the ROC Curve (AUC) of 0.84606 in 5-fold cross-validation experiments, indicating strong predictive performance.
- Extraction of Components: The SCCA successfully extracted 60 canonical components (CCs), each containing a small number of genes and diseases, facilitating interpretation.
- Case Studies: Three specific case studies (CC31, CC40, and CC54) are presented, demonstrating how the extracted gene-disease sets align with known biological mechanisms and literature evidence for diseases like sarcoma, ovarian cancer, and cataracts.
- Novel Predictions: The authors predicted associations for 100 new miRNAs, with some high-scoring predictions validated against HMDD v3.0.
- Comparison with Existing Databases: A comparison of top predicted diseases with those from HMDD v2.0 and v3.0 showed overlap, further supporting the method's capacity.
Limitations and Future Work
- Target Gene Information Dependency: The method relies on the availability of target gene information for new miRNAs. While biological assays are improving, this remains a potential limitation. The authors acknowledge that having detailed target gene profiles for candidate molecules is not always practical.
- Data Quality: The performance is contingent on the quality of the input datasets. While they used curated, experimentally validated data, biases or incompleteness in these databases could affect results.
- Future Work: The authors suggest that their method can serve as an informative tool for discovering disease pathogenesis. Further advancements in biological assays will likely provide more comprehensive miRNA-gene interaction data, enhancing the method's applicability.
Summary updated