MORONET: Multi-omics Integration via Graph Convolutional Networks for Biomedical Data Classification

Tongxin Wang, Wei Shao, Zhi Huang, Haixu Tang, Jie Zhang, Zhengming Ding, Kun HuangbioRxiv 2020 · DOI

TL;DR

This paper introduces MORONET, a novel deep learning framework for integrating multiple types of "omics" data (like gene expression, DNA methylation, and microRNA) to improve the accuracy of biomedical classification tasks. MORONET achieves state-of-the-art performance by combining graph convolutional networks (GCNs) for omics-specific learning with a view correlation discovery network (VCDN) for cross-omics integration, and it can also identify important disease-related biomarkers.

What Problem Does This Paper Address?

Understanding complex human diseases requires integrating information from various biological "omics" data sources. Each omics technology provides a partial view of biological processes, and combining them offers a more comprehensive picture. However, effectively integrating these diverse datasets for classification tasks, especially when labeled data is available, remains a challenge. Existing methods often fail to capture the intricate correlations between different omics data types or between samples, potentially leading to biased or suboptimal predictions. This paper aims to bridge this gap by proposing a novel supervised multi-omics integration method that leverages both omics-specific features and cross-omics relationships at the label space.

Methods

MORONET is a deep learning framework that integrates two key components:

  • Omics-Specific Learning via Graph Convolutional Networks (GCNs): For each omics data type (e.g., mRNA, DNA methylation, miRNA), a GCN is employed. GCNs are well-suited for this task because they can learn from both the features of individual samples and the relationships (patient similarity) between them, represented as a graph. This allows for a more nuanced understanding of omics-specific patterns compared to standard fully-connected networks. The paper details how patient similarity networks are constructed using cosine similarity and how GCNs are trained on these networks.
  • Multi-Omics Integration via View Correlation Discovery Network (VCDN): After obtaining initial class predictions from each omics-specific GCN, a VCDN is used to integrate these predictions. VCDN is designed to learn higher-level correlations across different omics data types specifically within the "label space." This means it looks at how the predictions from different omics sources relate to each other to make a final, more robust classification. The paper extends VCDN, originally designed for two views, to handle multiple omics data types by constructing a "cross-omics discovery tensor."

MORONET is an end-to-end model where both GCNs and VCDN are trained jointly.

Datasets

The MORONET framework was evaluated on four different biomedical classification tasks using publicly available datasets:

  • ROSMAP: For Alzheimer's disease (AD) patient classification (AD vs. normal control).
  • LGG: For tumor grade classification in Low Grade Glioma (Grade 2 vs. Grade 3).
  • KIPAN: For kidney cancer type classification (KICH, KIRC, KIRP).
  • BRCA: For breast invasive carcinoma subtype classification (PAM50 subtypes).

The paper notes that KIPAN was used as a proof-of-concept for multi-class applications. All datasets involved matched mRNA expression, DNA methylation, and miRNA expression data.

Key Insights and Ideas

  • The Power of Label-Space Correlation: The paper highlights that integrating omics data not just at the feature level but also by exploring correlations in the label space (via VCDN) significantly boosts classification performance. This suggests that understanding how different omics data types "agree" or "disagree" on class predictions is crucial.
  • GCNs Excel with Patient Similarity: Using GCNs with patient similarity networks allows the model to leverage both individual omics features and the complex relationships between patients, leading to better omics-specific learning.
  • MORONET's Robustness: The method demonstrates consistent outperformance across various datasets and classification tasks, including challenging ones like Alzheimer's disease and breast cancer subtype classification.
  • Biomarker Identification: Beyond classification, MORONET can identify important omics features (genes, miRNAs, methylation sites) that are strongly associated with the diseases of interest, offering valuable biological insights.
  • Deep Learning Design Matters: The paper suggests that while deep learning is powerful, its effectiveness in multi-omics integration relies on carefully designed architectures like MORONET, rather than simply applying standard deep learning models.

Key Results

MORONET consistently outperformed existing supervised multi-omics integration methods, including state-of-the-art approaches like block_plsda and block_splsda, across most classification tasks. For example:

  • On the ROSMAP dataset, MORONET achieved an accuracy of 82.45% ± 2.20, outperforming other methods.
  • For LGG grade classification, MORONET achieved 79.48% ± 2.36 accuracy, slightly better than GCN_NN.
  • In KIPAN, MORONET achieved near-perfect accuracy (99.80% ± 0.25).
  • For BRCA, MORONET achieved 80.61% ± 0.54 accuracy.

The paper also showed that integrating multiple omics data types via MORONET significantly improved performance compared to using any single omics data type alone, often with a reduced standard deviation in performance metrics.

Limitations and Future Work

  • Hyperparameter Sensitivity (k): The performance of MORONET is influenced by the hyperparameter k, which controls the density of the patient similarity network. While MORONET showed robustness across a range of k values, careful selection through cross-validation is still necessary.
  • Binary vs. Multi-class: The paper notes that in binary classification tasks (like LGG), the benefit of cross-view correlation learning might be limited compared to multi-class scenarios.
  • Future Work: The authors suggest that MORONET's GCN component is flexible and can be generalized to include more types of information to further boost classification performance. The framework's ability to identify biomarkers is also a promising avenue for future research.

Summary updated

Abstract

ABSTRACT To fully utilize the advances in omics technologies and achieve a more comprehensive understanding of human diseases, novel computational methods are required for integrative analysis for multiple types of omics data. We present a novel multi-omics integrative method named Multi-Omics gRaph cOnvolutional NETworks (MORONET) for biomedical classification. MORONET jointly explores omics-specific learning and cross-omics correlation learning for effective multi-omics data classification. We demonstrate that MORONET outperforms other state-of-the-art supervised multi-omics integrative analysis approaches from a wide range of biomedical classification applications using mRNA expression data, DNA methylation data, and miRNA expression data. Furthermore, MORONET is able to identify important biomarkers from different omics data types that are related with the investigated diseases.
BibTeX
@misc{wang2020moronet,
  author = {Tongxin Wang and Wei Shao and Zhi Huang and Haixu Tang and Jie Zhang and Zhengming Ding and Kun Huang},
  title = {{MORONET: Multi-omics Integration via Graph Convolutional Networks for Biomedical Data Classification}},
  year = {2020},
  doi = {10.1101/2020.07.02.184705},
  howpublished = {bioRxiv},
}
Figure