Dual Convolutional Neural Network Based Method for Predicting Disease-Related miRNAs

Ping Xuan, Yihua Dong, Yahong Guo, Tiangang Zhang, Yong LiuInternational Journal of Molecular Sciences 2018 · DOI

TL;DR

This paper introduces CNNDMP, a novel dual Convolutional Neural Network (CNN) approach to predict disease-associated microRNAs (miRNAs). By integrating miRNA and disease similarities, their known associations, and network topology information, CNNDMP significantly outperforms existing methods in identifying potential disease miRNAs.

What Problem Does This Paper Address?

MicroRNAs (miRNAs) are small RNA molecules that play crucial roles in regulating gene expression. Their dysregulation is increasingly linked to various human diseases, making the identification of disease-associated miRNAs vital for understanding disease etiology and developing new therapeutic strategies. While many computational methods exist, they often struggle to capture the complex, deep features within miRNA and disease similarities, as well as their intricate association networks. Existing approaches tend to rely on simpler integration of these data sources, failing to fully leverage the rich topological information present in these biological networks. This limitation hinders accurate prediction of novel disease-miRNA links.

Methods

The core of CNNDMP is a dual Convolutional Neural Network (CNN) architecture designed to learn from multiple facets of miRNA-disease relationships. The method can be broken down into three main components:

  1. Embedding Layer: This layer constructs rich feature representations for miRNAs and diseases. It combines:
  2. Original features: Known miRNA-disease associations and pre-computed miRNA and disease similarities.
  3. Biological premises: Leveraging the idea that functionally similar miRNAs are often associated with similar diseases, and vice-versa.
  4. Network Topology: This is a key innovation. The paper uses random walks with restart on miRNA and disease networks to capture their topological structures. These walks generate "convergent vectors" that represent the similarity of a starting node to all other nodes in the network, effectively encoding structural information.
  5. Dual CNN Architecture: The method employs two parallel CNN pathways:
  6. Left CNN: Processes the integrated original features and biological premises from the embedding layer.
  7. Right CNN: Processes the network topology information derived from the random walks. Both CNNs use multiple convolutional and pooling layers to extract deep feature representations.

  8. Combined Strategy: The scores from the left and right CNNs are combined using a weighted sum, allowing the model to balance the importance of original features and network topology for the final prediction.

Datasets

The study primarily utilizes the Human miRNA-disease database (HMDD) [39] for miRNA-disease associations. This database contains thousands of reliable association pairs. For disease similarity measurements, the paper refers to previously published studies [18, 19, 41] that integrated semantic and phenotypic information. For case studies, external databases like dbDEMC [31], miRCancer [32], and PhenomiR [33] were used to validate predicted disease-associated miRNAs.

Key Insights and Ideas

  • Topology Matters: The paper strongly emphasizes that the network topology of miRNA and disease relationships is crucial and often underutilized. Capturing this structural information through random walks significantly enhances prediction accuracy.
  • Dual CNN Synergy: Combining two CNN pathways – one for original features and one for network topology – allows for a more comprehensive understanding of miRNA-disease associations than either approach alone.
  • Deep Feature Learning: CNNs are adept at learning hierarchical features from complex data. Applying them to both the integrated feature embeddings and the network topology representations allows for the extraction of deeper, more nuanced patterns.
  • Embedding as a Foundation: The careful construction of the embedding layer, integrating multiple data sources and biological insights, provides a robust input for the CNNs.

Key Results

CNNDMP demonstrated superior performance compared to several state-of-the-art methods.

  • Cross-validation: On 15 common diseases, CNNDMP achieved an average ROC-AUC of 0.956, which was significantly higher than GSTRW (0.802), DMPred (0.917), PBMDA (0.844), and Liu's Method (0.865). Similarly, its PR-AUC was 0.538, outperforming others by a considerable margin (e.g., 36.1% higher than GSTRW).
  • Case Studies: For breast cancer, colorectal cancer, and lung cancer, CNNDMP successfully identified a high proportion of known disease-associated miRNAs, with many predictions validated by external databases and literature. For breast cancer, 43 out of the top 50 predicted miRNAs were found in the dbDEMC database.

Limitations and Future Work

  • Parameter Sensitivity: While the paper mentions optimizing parameters, the sensitivity of the model to specific parameter choices (like the random walk restart probability $\beta$ and the weighting parameter $\lambda$) could be a limitation.
  • Data Sparsity: The paper acknowledges that miRNA-disease associations are sparse, which is a general challenge in this field. While CNNDMP handles this well, extreme sparsity might still pose difficulties.
  • Future Work: The authors suggest applying the method to a larger number of diseases (they mention extending to all 326 diseases in their dataset) and exploring further integration of diverse biological data sources. The potential for discovering novel disease-related miRNAs remains a key area for future exploration.

Summary updated

Abstract

Identification of disease-related microRNAs (disease miRNAs) is helpful for understanding and exploring the etiology and pathogenesis of diseases. Most of recent methods predict disease miRNAs by integrating the similarities and associations of miRNAs and diseases. However, these methods fail to learn the deep features of the miRNA similarities, the disease similarities, and the miRNA⁻disease associations. We propose a dual convolutional neural network-based method for predicting candidate disease miRNAs and refer to it as CNNDMP. CNNDMP not only exploits the similarities and associations of miRNAs and diseases, but also captures the topology structures of the miRNA and disease networks. An embedding layer is constructed by combining the biological premises about the miRNA⁻disease associations. A new framework based on the dual convolutional neural network is presented for extracting the deep feature representation of associations. The left part of the framework focuses on integrating the original similarities and associations of miRNAs and diseases. The novel miRNA and disease similarities which contain the topology structures are obtained by random walks on the miRNA and disease networks, and their deep features are learned by the right part of the framework. CNNDMP achieves the superior prediction performance than several state-of-the-art methods during the cross-validation process. Case studies on breast cancer, colorectal cancer and lung cancer further demonstrate CNNDMP's powerful ability of discovering potential disease miRNAs.
BibTeX
@article{xuan2018dual,
  author = {Ping Xuan and Yihua Dong and Yahong Guo and Tiangang Zhang and Yong Liu},
  title = {{Dual Convolutional Neural Network Based Method for Predicting Disease-Related miRNAs}},
  journal = {International Journal of Molecular Sciences},
  year = {2018},
  doi = {10.3390/ijms19123732},
}
Figure