TL;DR
This paper introduces CNNDLP, a novel deep learning framework that leverages convolutional neural networks and autoencoders with an attention mechanism to predict associations between long non-coding RNAs (lncRNAs) and diseases. By integrating diverse biological data and learning nuanced representations, CNNDLP significantly outperforms existing methods in identifying potential disease-related lncRNAs, offering a valuable tool for biological research.
What Problem Does This Paper Address?
Long non-coding RNAs (lncRNAs) are increasingly recognized for their crucial roles in biological processes and their strong links to various diseases. Accurately predicting which lncRNAs are associated with specific diseases is vital for understanding disease mechanisms and developing new therapeutic strategies. However, previous methods often relied on shallow learning models, struggled to effectively integrate heterogeneous biological data (like lncRNA-disease associations, lncRNA-miRNA interactions, and their respective similarities), and failed to capture complex, low-dimensional feature representations. This paper aims to bridge this gap by proposing a more sophisticated deep learning approach that can better handle multi-source data and learn richer representations for more accurate lncRNA-disease association prediction.
Methods
CNNDLP employs a dual-pathway deep learning architecture:
- Left Pathway (Attention Mechanism): This pathway focuses on learning an "attention representation" for lncRNA-disease pairs. It uses a convolutional neural network (CNN) with an "adjacent edge attention" mechanism. This attention mechanism is key, as it allows the model to weigh the importance of different connections (edges) between lncRNAs, miRNAs, and diseases when learning representations. This helps the model focus on the most informative relationships.
- Right Pathway (Convolutional Autoencoder): This pathway aims to learn "low-dimensional network representations." It utilizes a convolutional autoencoder to compress and reconstruct the input data, effectively learning compact and meaningful feature representations. This is particularly useful for handling the sparsity of biological data.
Both pathways are built upon two novel embedding layers that integrate various biological data sources:
- Left Embedding Layer: Integrates lncRNA similarities, lncRNA-disease associations, and miRNA-disease associations.
- Right Embedding Layer: Incorporates new types of lncRNA and disease similarities calculated from multiple bipartite networks, along with lncRNA-miRNA interactions and miRNA-disease associations.
The final prediction score is a weighted combination of the scores from both pathways.
Datasets
The study utilized the following datasets:
- LncRNA-disease associations: Obtained from the LncRNADisease database, comprising 2687 associations covering 240 lncRNAs and 405 diseases.
- LncRNA-miRNA interactions: Extracted from starBasev2.0, including 1002 interactions.
- miRNA-disease associations: Sourced from the HMDD database, containing 13,559 associations.
- Disease similarities: Calculated based on directed acyclic graphs (DAGs) from the U.S. National Library of Medicine (MeSH).
The paper does not introduce new benchmark datasets but rather uses existing, well-established biological databases for training and evaluation.
Key Insights and Ideas
- Attention at the Edge Level: The core innovation is the "adjacent edge attention" mechanism. This allows the model to dynamically assign importance to different biological relationships (edges) when learning representations, recognizing that not all connections are equally informative for predicting lncRNA-disease associations.
- Dual Representation Learning: The separation into an "attention representation" (left pathway) and a "low-dimensional network representation" (right pathway) allows CNNDLP to capture different facets of the complex lncRNA-disease relationship.
- Novel Similarity Measures: The paper introduces new ways to calculate lncRNA and disease similarities by considering topological structures across multiple bipartite networks, going beyond simpler similarity calculations.
- Heterogeneous Data Integration: CNNDLP excels at integrating diverse biological data sources (associations, interactions, similarities) from lncRNAs, miRNAs, and diseases, which is a significant challenge in bioinformatics.
Key Results
- Superior Prediction Performance: CNNDLP achieved the highest average Area Under the ROC Curve (AUC) of 0.969 across 405 diseases, outperforming state-of-the-art methods like SIMCLDA, Ping's method, MFLDA, LDAP, and CNNLDA.
- Improved Precision-Recall: The model also demonstrated superior performance in terms of Area Under the Precision-Recall Curve (AUPR), with an average AUPR of 0.286, significantly higher than other methods.
- High Recall at Top Ranks: CNNDLP showed excellent recall rates, correctly identifying a high percentage of known lncRNA-disease associations within the top-ranked predictions (e.g., 88.6% in the top 30 candidates). This is crucial for biologists seeking to prioritize candidates for experimental validation.
- Case Study Validation: Case studies on stomach cancer, breast cancer, and prostate cancer demonstrated CNNDLP's ability to identify known and potentially novel disease-related lncRNAs, with a high proportion of its top predictions being supported by existing databases and literature.
Limitations and Future Work
- Computational Cost: While not explicitly detailed as a limitation, deep learning models with multiple layers and attention mechanisms can be computationally intensive.
- Data Sparsity: Although the model is designed to handle sparse data, the inherent sparsity of biological interaction networks remains a general challenge.
- Experimental Validation: The paper relies on existing curated databases and literature for validation. Future work would ideally involve wet-lab experimental validation of the newly predicted associations.
- Generalizability: While tested on 405 diseases, further evaluation on a broader range of diseases and biological contexts could be beneficial.
The authors suggest that CNNDLP can serve as a powerful prioritization tool for screening prospective candidates, paving the way for subsequent wet-lab experiments to confirm actual lncRNA-disease associations.
Summary updated