Prediction of drug-disease associations based on ensemble meta paths and singular value decomposition

Guangsheng Wu, Juan Liu, Xiang YueBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces EMP-SVD, a novel method for predicting drug-disease associations that bypasses the need for explicit drug or disease similarity data. By leveraging heterogeneous interaction networks and an ensemble of meta-path-based classifiers combined with Singular Value Decomposition (SVD), EMP-SVD effectively identifies potential drug repurposing opportunities.

What Problem Does This Paper Address?

Drug repositioning, the process of finding new therapeutic uses for existing drugs, is a cost-effective and time-efficient alternative to de novo drug discovery. However, accurately predicting these associations computationally is challenging. Many existing methods rely on calculating drug-drug or disease-disease similarities, which can be difficult to obtain, vary significantly between methods, or be unavailable for certain entities. Furthermore, supervised learning approaches often struggle with the scarcity of experimentally validated negative drug-disease pairs, leading to the common practice of treating all unlabeled pairs as negative, which can introduce noise and reduce model reliability. This paper aims to address these limitations by proposing a method that doesn't require similarity information and employs a more sophisticated approach to selecting negative samples.

Methods

The core of the EMP-SVD method lies in its ability to integrate diverse interaction data (drug-disease, drug-protein, disease-protein) without relying on similarity measures. The approach involves:

  • Ensemble of Meta Paths: Five distinct "meta paths" are defined to capture different indirect relationships between drugs and diseases through proteins. These meta paths represent sequences of interactions, such as "Drug → Protein → Disease" or "Drug → Disease → Protein → Disease."
  • Commuting Matrices: For each meta path, a "commuting matrix" is constructed. The elements of this matrix quantify the number of distinct interaction paths between a given drug and disease following that specific meta path.
  • Singular Value Decomposition (SVD): To reduce the dimensionality of the feature space and extract latent representations, SVD is applied to each commuting matrix. This process decomposes the matrix into components that capture the most significant underlying patterns of drug and disease features. The dimensionality reduction is controlled by a latent_feature_percent parameter.
  • Ensemble Classifier: Five base classifiers, one for each meta path, are trained using the Random Forest algorithm. These base classifiers are then combined into an ensemble classifier, which outputs a final prediction score for a drug-disease pair.
  • Likely Negative Sample Selection: Instead of randomly selecting unlabeled pairs as negative, EMP-SVD identifies "more likely negative" samples. This is based on the assumption that if a drug and a disease share no common interacting proteins, they are less likely to have a therapeutic relationship. This is achieved by analyzing the product of drug-protein and disease-protein interaction matrices.

Datasets

The study utilized the following datasets:

  • DrugBank [3, 22, 23]: For drug-protein interactions and chemical structure data.
  • OMIM [24]: For disease-protein interactions.
  • Gottlieb's dataset [11]: For known drug-disease interactions.

The paper notes that the collected heterogeneous network data is sparse. For comparison purposes with other methods, they also collected drug-drug chemical similarities (using SMILES strings and Openbabel) and protein-protein similarities (using Smith-Waterman algorithm on sequence data), as well as disease-disease similarities from MimMiner.

Key Insights and Ideas

  • Similarity-Free Prediction: The most significant insight is that accurate drug-disease association prediction can be achieved without explicit drug or disease similarity data. This overcomes a major bottleneck in many existing methods.
  • Power of Indirect Relationships: The ensemble of meta paths demonstrates that combining multiple indirect interaction pathways provides a richer understanding of drug-disease relationships than any single pathway alone.
  • SVD for Feature Engineering: Using SVD to extract latent features from commuting matrices is an effective way to handle high-dimensional interaction data and create robust representations for drug-disease pairs.
  • Smarter Negative Sampling: The "likely negative sample" selection strategy is a crucial improvement over random assignment, leading to more reliable training data and potentially better model performance.
  • Ensemble Effectiveness: Combining multiple base classifiers trained on different meta paths significantly improves prediction performance compared to individual classifiers.

Key Results

  • Superior Performance: EMP-SVD outperformed several state-of-the-art methods (PREDICT, TL-HGBI, MBiRW, LRSSL, SCMFDD) in terms of key evaluation metrics, particularly AUPR (0.956) and AUC (0.951).
  • Robustness: The ensemble classifier's performance was robust across different settings of the latent_feature_percent parameter.
  • Negative Sample Filtering Benefit: The negative sample filtering strategy consistently improved performance across all meta-path-based classifiers.
  • Practicality Demonstrated: Case studies identified 13 out of the top 20 predicted drug-disease associations that were validated by literature, showcasing the method's ability to uncover novel and potentially useful associations.

Limitations and Future Work

  • Limited Data Integration: The current EMP-SVD model only integrates drug, protein, and disease information. The authors suggest that incorporating other data sources like side effects, pathways, and tissue information could further enhance performance.
  • Simplified Negative Sampling: The negative sample selection currently relies solely on common proteins. Future work could explore using other information, such as gene expression data, or employing positive and unlabeled sample learning methods.
  • Parameter Tuning: While EMP-SVD has fewer parameters than many methods, the latent_feature_percent still requires setting. Future research could explore methods for automatic parameter optimization or learning.

Summary updated

BibTeX
@article{wu2019prediction,
  author = {Guangsheng Wu and Juan Liu and Xiang Yue},
  title = {{Prediction of drug-disease associations based on ensemble meta paths and singular value decomposition}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-2644-5},
}
Figure