TL;DR
This paper surveys computational methods for predicting associations between non-coding RNAs (ncRNAs) and diseases, highlighting a shift towards a "Mode of Action (MoA) network" perspective. This approach integrates ncRNAs with other biological entities like mRNAs and pathways to better understand disease mechanisms, moving beyond simple correlations to uncover deeper biological insights.
What Problem Does This Paper Address?
Understanding how non-coding RNAs (ncRNAs) contribute to diseases is crucial for developing new diagnostics and therapeutics. However, ncRNAs don't act in isolation; they influence gene expression, which in turn affects biological pathways and cellular phenotypes, ultimately leading to disease. The challenge lies in computationally modeling these complex, multi-step regulatory processes to accurately predict ncRNA-disease associations. Traditional methods often focus on direct correlations, missing the intricate biological context. This paper addresses this gap by reviewing computational approaches that integrate diverse biological data layers to build a more holistic "Mode of Action (MoA) network" for ncRNA-disease association prediction.
Methods
The paper categorizes computational methodologies into two main groups:
- Network Mining Methods: These approaches leverage existing biological networks to infer associations.
- Statistical Methods: Primarily differential expression (DE) analysis to identify ncRNAs or genes with significantly altered expression levels between conditions. While fundamental, they often miss complex interactions.
- Network Propagation: Algorithms like Random Walk with Restart (RWR) are used to spread signals (e.g., disease relevance) through a biological network, identifying influential nodes (ncRNAs or genes).
- Random Walk-Based Methods: These methods generate node embeddings by performing random walks on the network, capturing local structural information.
- Network Learning Methods: These employ machine learning, particularly deep learning, to learn complex patterns from network data.
- Matrix Factorization: Decomposes matrices representing known associations (e.g., ncRNA-disease) into latent factors to predict missing associations.
- Graph Neural Networks (GNNs): Directly incorporate graph structures into neural networks, learning node embeddings by aggregating information from neighbors. This allows for end-to-end learning for ncRNA-disease prediction.
The paper emphasizes the MoA network perspective, which integrates ncRNAs with mRNAs, pathways, and phenotypes as nodes in a heterogeneous network.
Datasets
The paper highlights several existing databases that provide experimentally supported ncRNA-disease associations. These databases are crucial for training and evaluating computational models. Examples include:
- HMDD v3.2: For miRNA-disease associations.
- LncRNADisease v2.0: For lncRNA-disease associations.
- CircRNADisease: For circRNA-disease associations.
- MNDR v3.1: A comprehensive database integrating various mammalian ncRNA-disease associations.
The paper does not introduce new datasets but focuses on surveying methods that utilize these existing resources.
Key Insights and Ideas
- The MoA Network is the Future: The paper strongly advocates for the "Mode of Action (MoA) network" perspective as a new trend in ncRNA-disease association prediction. This holistic approach, inspired by drug discovery, moves beyond simple correlations to model the biological cascade from ncRNA to disease.
- Beyond Direct Associations: Simply knowing that an ncRNA is associated with a disease is insufficient. Understanding how it exerts its effect through gene regulation, pathway modulation, and phenotypic changes is key to unlocking its therapeutic potential.
- Integration is Key: The success of MoA network approaches hinges on integrating diverse biological data types (ncRNA, mRNA, pathways, phenotypes). This multi-omics integration provides a richer context for prediction.
- Deep Learning's Rise: Graph Neural Networks (GNNs) are emerging as powerful tools for learning from complex network structures, offering an end-to-end approach for ncRNA-disease prediction.
- Databases are Foundational: The availability of curated databases is critical for advancing computational ncRNA-disease association studies.
Key Results
The paper is a survey, not an experimental study, so it doesn't present novel quantitative results. Instead, it summarizes the performance of various existing computational tools in predicting ncRNA-disease associations. For instance, Table 3 shows that tools like GCNCDA and Nguyen et al.'s method achieve high AUROC scores (e.g., 0.9090 and 0.9882, respectively) in predicting miRNA-disease associations. Table 4 highlights that MOGONET, using multi-omics data, achieved an AUROC of 0.9134 for predicting phenotypes. The key "result" is the demonstration that integrating multiple biological layers (as in the MoA network) generally leads to improved prediction performance compared to methods relying on single data types.
Limitations and Future Work
- Limited MoA Network Coverage: While the MoA network concept is promising, the paper notes that few existing methods fully integrate all the proposed entities (ncRNA, mRNA, pathway, phenotype). Many studies still cover only a fraction of this network.
- Data Availability and Quality: The accuracy and comprehensiveness of the underlying databases are crucial. Future work could focus on improving data curation and standardization.
- Interpretability: While deep learning methods like GNNs show promise, understanding why they make certain predictions (interpretability) remains a challenge.
- Experimental Validation: Computational predictions are a starting point. Future work will require extensive experimental validation to confirm the discovered ncRNA-disease associations and their underlying mechanisms.
- Formulating ncDA into Learning Frameworks: The paper suggests that despite the progress, there's still a need for more studies that explicitly formulate ncRNA-disease association prediction within a robust learning framework that leverages the MoA network. This is identified as a promising future direction.
Summary updated