TL;DR
This paper introduces MSPCD, a novel computational framework that accurately predicts associations between circular RNAs (circRNAs) and diseases. By integrating diverse biological data and employing a hierarchical neural network, MSPCD significantly outperforms existing methods, offering a promising tool for understanding disease mechanisms and identifying potential biomarkers.
What Problem Does This Paper Address?
Circular RNAs (circRNAs) are a fascinating class of non-coding RNA molecules that have emerged as crucial players in various biological processes and disease development. Their stable, circular structure makes them resistant to degradation, allowing them to exert regulatory functions, often by interacting with microRNAs (miRNAs). Identifying the specific associations between circRNAs and diseases is vital for precise diagnosis and effective treatment strategies. However, traditional experimental methods for discovering these associations are time-consuming, expensive, and labor-intensive. This necessitates the development of efficient and accurate computational approaches to predict these links. While several computational methods exist, they often rely on limited biological information or simpler modeling techniques, leaving room for improvement in prediction accuracy.
Methods
MSPCD tackles this challenge by adopting a multi-pronged approach:
- Multi-Source Data Integration: The core of MSPCD lies in its ability to leverage a rich tapestry of biological data. It goes beyond simple circRNA-disease associations by incorporating:
- CircRNA-miRNA associations: Understanding how circRNAs interact with miRNAs, which are known regulators of gene expression.
- CircRNA-Gene Ontology (GO) associations: Utilizing functional annotations of circRNAs to infer their biological roles.
- CircRNA sequence similarity: Measuring similarity based on the edit distance of circRNA sequences.
- CircRNA functional similarity: Assessing similarity based on GO terms and miRNA interactions.
- CircRNA GIP kernel similarity: Capturing network topology information.
- Disease GIP kernel similarity: Measuring disease similarity based on network topology.
-
Disease semantic similarity: Leveraging the MeSH database to understand disease relationships.
These diverse data sources are fused to create comprehensive similarity matrices for both circRNAs and diseases.
-
Hierarchical Neural Network for High-Order Feature Extraction: Instead of directly using concatenated similarity features, MSPCD employs a hierarchical neural network (specifically, three layers of fully connected neural networks with ReLU activation) to extract "high-order" latent features for both circRNAs and diseases. This allows the model to learn more complex and abstract representations of these entities, capturing intricate relationships that might be missed by simpler methods.
- Feature Interaction and DNN Prediction: The extracted high-order features are then combined. A dot product is used to capture interaction features between circRNA and disease features. These interaction features are then concatenated with the original high-order features to form a comprehensive representation of a circRNA-disease pair. This combined vector is fed into a Deep Neural Network (DNN) with three fully connected layers (the first two using ReLU, and the last using a sigmoid activation) to predict the final circRNA-disease association probability.
Datasets
The study utilizes two publicly available datasets:
- CircFunBase: This dataset was used for training and evaluating the primary model. It contains 2984 verified circRNA-disease associations, involving 2597 circRNAs and 67 diseases.
- circR2Disease: This dataset was used for further validation and comparison. It includes 612 circRNA-disease associations, involving 533 circRNAs and 89 diseases.
Key Insights and Ideas
- The Power of Data Fusion: The paper strongly emphasizes that integrating multiple sources of biological information is critical for accurate circRNA-disease association prediction. Simply relying on one or two types of similarity is insufficient.
- Hierarchical Feature Learning: The use of a hierarchical neural network to extract high-order features is a key innovation. This approach allows the model to learn more abstract and robust representations, leading to improved predictive performance compared to methods that directly concatenate raw similarity features.
- Beyond Simple Associations: The study highlights that understanding the functional roles of circRNAs (via GO terms and miRNA interactions) and the semantic relationships between diseases is as important as direct association data for accurate prediction.
Key Results
- Superior Performance: MSPCD achieved a remarkable average AUC of 0.9904 on the CircFunBase dataset, significantly outperforming seven state-of-the-art methods. The AUC value was 4.12% higher than the second-best method.
- Robustness Across Datasets: MSPCD also demonstrated strong performance on the circR2Disease dataset, achieving an average AUC of 0.9526, further validating its generalizability.
- Effective Feature Extraction: The analysis of high-order feature lengths showed that the model's performance remained stable across a range of lengths (16-128), suggesting effective regularization against overfitting.
- Case Study Validation: A case study identified 15 potential circRNA-disease associations, with five being confirmed through literature validation, demonstrating MSPCD's practical utility in discovering novel associations.
Limitations and Future Work
- Sparse Supervisory Signal: The authors acknowledge that the circRNA-disease association prediction is modeled as a supervised learning problem with a sparse supervisory signal. This limitation can affect performance.
- Future Work - Self-Supervised Learning: To address the sparse signal issue, the authors propose exploring self-supervised learning techniques. This would involve pretraining the model on large-scale datasets without manual annotations to learn more robust representations before fine-tuning on the specific circRNA-disease association task.
- Future Work - Unconfirmed Associations: The case study identified several unconfirmed associations. Future work could involve further experimental validation of these potential links.
Summary updated