Sguq: Staged Graph Convolution Neural Network for Alzheimer's Disease Diagnosis Using Multi-Omics Data

Liang Tao, Yixin Xie, Juan Deng, Hui Shen, Hong‐Wen Deng, Weihua Zhou, Chen Zhao

TL;DR

This paper introduces SGUQ, a novel staged graph convolutional neural network that progressively integrates multi-omics data for Alzheimer's disease (AD) diagnosis. By intelligently incorporating data types only when necessary, SGUQ aims to reduce costs and improve diagnostic accuracy, achieving a remarkable 0.858 accuracy on the ROSMAP dataset.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a devastating neurodegenerative disorder with a significant global impact. While high-throughput "omics" technologies (genomics, transcriptomics, proteomics, epigenomics) have revolutionized our understanding of AD, conventional AI approaches for diagnosis often require the complete set of omics data upfront. This is not only expensive and time-consuming but also potentially unnecessary, as not all data types may be required for every patient's diagnosis. The challenge lies in developing a cost-effective and accurate diagnostic method that leverages multi-omics data without incurring excessive costs or patient burden.

Methods

The core of the proposed solution is the Staged Graph Convolutional Network with Uncertainty Quantification (SGUQ). This method operates in stages:

  • Stage 1 (Single-View Classification): The model begins by using a single, most informative omics data type (mRNA in this case) to make an initial prediction. A graph convolutional network (GCN) is employed to learn features from this data, considering the similarity between patients.
  • Uncertainty Quantification: Crucially, SGUQ quantifies the uncertainty associated with each prediction. This is achieved by performing multiple independent predictions for each sample and calculating the standard deviation of these predictions. A higher standard deviation indicates higher uncertainty.
  • Staged Integration: If the uncertainty from Stage 1 exceeds a predefined threshold (t₁), the model proceeds to Stage 2.
  • Stage 2 (Bi-Modal Classification): A second omics data type (e.g., DNA methylation) is incorporated. A View Correlation Discovery Network (VCDN) is used to integrate the information from the first two omics types, learning higher-level cross-omics correlations.
  • Stage 3 (Tri-Modal Classification): If uncertainty remains high after Stage 2 (above threshold t₂), a third omics data type (miRNA) is added, and the VCDN is extended to handle three modalities.
  • Threshold Optimization: The optimal thresholds (t₁ and t₂) are determined through an iterative process that maximizes classification accuracy, ensuring a balance between cost-effectiveness and diagnostic performance.

Datasets

The study utilizes the ROSMAP dataset, a comprehensive resource derived from two longitudinal clinical studies. This dataset includes brain tissue samples from older adults and provides multi-omics data. For this study, three types of omics data were used:

  • mRNA Expression Data (Transcriptomic)
  • DNA Methylation Data (Epigenomics)
  • miRNA Expression Data (Non-Coding RNA Transcriptomics)

The pre-processed dataset was obtained from MOGONET and comprises 169 normal control patients and 182 AD patients.

Key Insights and Ideas

  • Staged Approach is Key: The most significant insight is that not all patients require all omics data for accurate AD diagnosis. A staged approach, starting with the most cost-effective data and progressively adding more complex data only when needed, is both efficient and effective.
  • Uncertainty as a Guide: Quantifying prediction uncertainty is vital. It acts as a smart "gatekeeper," signaling when more information is needed, thereby optimizing resource allocation and reducing unnecessary testing.
  • mRNA as a Strong Starting Point: The study highlights that mRNA data alone provides a surprisingly strong predictive capability for AD diagnosis, suggesting its importance as a primary biomarker.
  • Complementary Nature of Omics: While mRNA is powerful, integrating DNA methylation data with it significantly boosts performance, demonstrating the complementary value of different omics layers.

Key Results

  • High Overall Accuracy: The proposed SGUQ model achieved an impressive accuracy of 0.8584, with an F1-score of 0.8571 and an AUC of 0.8600 on the ROSMAP dataset.
  • Cost-Effectiveness: The staged approach proved highly effective:
  • 46.23% of samples could be reliably predicted using only single-modal mRNA data.
  • An additional 16.04% of samples benefited from combining mRNA and DNA methylation data.
  • Outperformance: SGUQ significantly outperformed various state-of-the-art algorithms, including classical machine learning models and other multi-omics integration methods, as detailed in Table 3. For instance, it achieved a higher accuracy than MMDynamics (0.842) and MLCLNet (0.844).
  • Optimal Thresholds: The identified optimal thresholds for the ROSMAP dataset were t₁ = 0.0527 and t₂ = 0.1024.

Limitations and Future Work

  • Uncertainty and Complexity: While the tri-view model achieves the highest accuracy, it may introduce higher uncertainty due to the complexity of integrating multiple data types. This model is best suited for scenarios where maximizing accuracy is paramount and computational complexity is manageable.
  • Future Directions: The paper suggests that samples with high uncertainty might reflect complex biological patterns that warrant further investigation. This could guide future data collection, experimental design, and the discovery of new biomarkers or disease mechanisms. The staged approach is also highlighted as a promising direction for broader clinical decision-making using multi-view data, paving the way for more efficient and personalized medicine.

Summary updated

Abstract

Alzheimer's disease (AD) is a chronic neurodegenerative disorder and the leading cause of dementia, significantly impacting cost, mortality, and burden worldwide. The advent of high-throughput omics technologies, such as genomics, transcriptomics, proteomics, and epigenomics, has revolutionized the molecular understanding of AD. Conventional AI approaches typically require the completion of all omics data at the outset to achieve optimal AD diagnosis, which are inefficient and may be unnecessary. To reduce the clinical cost and improve the accuracy of AD diagnosis using multi-omics data, we propose a novel staged graph convolutional network with uncertainty quantification (SGUQ). SGUQ begins with mRNA and progressively incorporates DNA methylation and miRNA data only when necessary, reducing overall costs and exposure to harmful tests. Experimental results indicate that 46.23% of the samples can be reliably predicted using only single-modal omics data (mRNA), while an additional 16.04% of the samples can achieve reliable predictions when combining two omics data types (mRNA + DNA methylation). In addition, the proposed staged SGUQ achieved an accuracy of 0.858 on ROSMAP dataset, which outperformed existing methods significantly. The proposed SGUQ can not only be applied to AD diagnosis using multi-omics data but also has the potential for clinical decision-making using multi-viewed data. Our implementation is publicly available at https://github.com/chenzhao2023/multiomicsuncertainty.
BibTeX
@misc{tao2024sguq,
  author = {Liang Tao and Yixin Xie and Juan Deng and Hui Shen and Hong‐Wen Deng and Weihua Zhou and Chen Zhao},
  title = {{Sguq: Staged Graph Convolution Neural Network for Alzheimer's Disease Diagnosis Using Multi-Omics Data}},
  year = {2024},
  eprint = {2410.11046},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.11046},
}

Loading related content...

Figure