Identification and Analysis of Human Microbe-Disease Associations by Matrix Decomposition and Label Propagation

Jia Qu, Yan Zhao, Jun YinFrontiers in Microbiology 2019 · DOI

TL;DR

This paper introduces MDLPHMDA, a novel computational model that leverages matrix decomposition and label propagation to predict associations between human microbes and diseases. By cleaning noisy microbe-disease data and integrating various similarity measures, MDLPHMDA achieves high prediction accuracy and demonstrates its utility in identifying potential microbe-disease links for conditions like asthma and colorectal cancer.

What Problem Does This Paper Address?

The human body hosts a vast and complex microbial ecosystem, known as the microbiota, which plays a crucial role in maintaining health. Emerging research increasingly links alterations in this microbiota (dysbiosis) to a wide range of human diseases, from infectious to chronic conditions like inflammatory bowel disease and metabolic disorders. Understanding these microbe-disease associations is vital for deciphering disease mechanisms and developing new diagnostic and therapeutic strategies. However, experimentally identifying these links is a laborious, costly, and time-consuming process. Therefore, there's a pressing need for effective computational methods to predict these associations and guide experimental research. While several computational models exist, the authors aim to improve prediction accuracy by developing a more robust approach that addresses data noise and integrates diverse biological information.

Methods

The core of the MDLPHMDA model lies in a two-stage approach:

  1. Matrix Decomposition for Data Cleaning: The authors first address potential noise and redundancy in the known microbe-disease association data. They employ Sparse Learning Method (SLM) to decompose the initial adjacency matrix (representing known associations) into a low-rank matrix (capturing essential patterns) and a sparse matrix (representing noise). This process effectively "cleans" the data, creating a more reliable adjacency matrix for subsequent steps.
  2. Label Propagation for Prediction: Using the cleaned adjacency matrix, the model then applies a Label Propagation Algorithm (LPA). LPA works by propagating "labels" (in this case, association information) across a graph. The authors construct two types of graphs:
  3. Disease Graph: Nodes represent diseases, and edge weights are derived from an integrated disease similarity score. This score combines Gaussian interaction profile kernel similarity for diseases and disease symptom similarity.
  4. Microbe Graph: Nodes represent microbes, and edge weights are based on Gaussian interaction profile kernel similarity for microbes.

    LPA is run on both graphs, and the final prediction scores for microbe-disease associations are obtained by averaging the results from both perspectives. This dual approach aims to capture associations from both disease and microbe-centric viewpoints.

Datasets

The primary dataset used in this study is the Human Microbe-Disease Association Database (HMDAD). This database contains confirmed microbe-disease associations collected from published literature. The authors utilized a refined version of this dataset, comprising 450 associations between 39 diseases and 292 microbes. No new benchmarks were introduced; the focus was on leveraging and enhancing existing data.

Key Insights and Ideas

  • Data Noise is a Major Hurdle: The paper highlights that raw microbe-disease association data can be noisy and redundant, potentially hindering prediction accuracy. Matrix decomposition is presented as an effective strategy to mitigate this issue.
  • Integrating Diverse Similarity Measures is Crucial: The model's success stems from combining multiple sources of information: known associations, Gaussian interaction profiles (which capture how microbes/diseases interact with others), and disease symptom similarity. This multi-faceted approach provides a richer context for prediction.
  • Dual Perspective Prediction Enhances Robustness: By performing label propagation from both disease and microbe perspectives and then averaging the results, MDLPHMDA gains a more comprehensive understanding of potential associations.
  • Unsupervised Learning for Novel Discoveries: The method does not require negative samples, making it suitable for predicting associations for new diseases or microbes where negative examples are unknown.

Key Results

  • High Prediction Accuracy: MDLPHMDA achieved impressive performance in cross-validation:
  • Leave-One-Out Cross Validation (LOOCV): Area Under the Curve (AUC) of 0.9034.
  • Five-fold Cross Validation: AUC of 0.8954 ± 0.0030. These AUC values are competitive and, in the case of LOOCV, superior to other established methods like LRLSHMDA and KATZHMDA. Statistical significance tests confirmed MDLPHMDA's superior performance.

  • Successful Case Studies: The model demonstrated its practical utility in identifying known microbe-disease associations for several conditions:

  • Asthma: 10 out of the top 10 predicted microbes were confirmed by literature.
  • Colorectal Carcinoma (CRC): 9 out of the top 10 predicted microbes were confirmed.
  • Liver Cirrhosis: 9 out of the top 10 predicted microbes were confirmed.
  • Type 1 Diabetes: 8 out of the top 10 predicted microbes were confirmed.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Limited Data: The current dataset of known microbe-disease associations is finite. The availability of more confirmed associations would further enhance the model's predictive power.
  • Parameter Optimization: Finding the optimal values for all parameters in the model can be challenging, impacting prediction accuracy.
  • SLM's Potential for Noise: While SLM aims to remove noise, it might also introduce or miss some association information, potentially affecting LPA's performance.
  • Feature Expansion: The current model relies heavily on known associations for similarity calculations. Future work could incorporate other features, such as microbe-drug associations or other omics data, to build a more comprehensive understanding of microbe similarity.
  • Experimental Validation: The authors emphasize the need for experimental validation of the predicted associations to confirm their biological relevance.

Summary updated

Abstract

Studies have shown that microbes exist widely in the human body and are closely related to human complex diseases. Predicting potential associations between microbes and diseases is conducive to understanding the mechanisms of complex diseases and can also facilitate the diagnosis and prevention of human diseases. In this paper, we put forward the Matrix Decomposition and Label Propagation for Human Microbe-Disease Association prediction (MDLPHMDA) on the basis of the dataset of known microbe-disease associations collected from the database of HMDAD and the Gaussian interaction profile kernel similarity for diseases and microbes, disease symptom similarity. Moreover, the performance of our model was evaluated by means of leave-one-out cross validation and five-fold cross validation, and the corresponding AUCs of 0.9034 and 0.8954 ± 0.0030 were gained, respectively. In case studies, 10, 9, 9, and 8 out of the top 10 predicted microbes for asthma, colorectal carcinoma, liver cirrhosis, and type 1 diabetes were confirmed by literatures, respectively. Overall, evaluation results showed that MDLPHMDA has good performance in potential microbe-diseasepositive free parameter, which associations prediction.
BibTeX
@inproceedings{qu2019identification,
  author = {Jia Qu and Yan Zhao and Jun Yin},
  title = {{Identification and Analysis of Human Microbe-Disease Associations by Matrix Decomposition and Label Propagation}},
  booktitle = {Frontiers in Microbiology},
  year = {2019},
  doi = {10.3389/fmicb.2019.00291},
}
Figure