New Empirical Bayes Models to Jointly Analyze Multiple RNA-Sequencing Data in a Hypophosphatasia Disease Study

Dawson Kinsman, Jian Hu, Zhi Zhang, Gengxin LiGenes 2024 · OpenReview

TL;DR

This paper introduces novel empirical Bayes (EB) integration methods to analyze multiple RNA-sequencing (RNA-seq) datasets for improved disease gene discovery and risk prediction. By combining data from spinal cord and neocortex tissues in a hypophosphatasia (HPP) mouse model, their weighted gene and weighted Z methods significantly enhance the detection of disease-associated genes and improve prediction accuracy compared to analyzing each tissue independently.

What Problem Does This Paper Address?

Hypophosphatasia (HPP) is a rare genetic disorder characterized by impaired bone mineralization and, in severe cases, debilitating neurological symptoms like muscle weakness and seizures. Understanding the genetic underpinnings of these neurological manifestations is crucial for developing effective treatments, but it's challenging due to the complexity of gene expression and the limited availability of high-resolution transcriptomic data. Analyzing RNA-seq data, while powerful, faces a common challenge: a high-dimensional feature space (many genes) with a limited number of samples. This imbalance can lead to biased estimates and an overestimation of genetic effects. Existing methods often struggle to effectively integrate multiple RNA-seq datasets to maximize the detection of subtle genetic signals, especially when disease-related genes might be differentially expressed in specific tissues rather than all of them.

Methods

The core of this work lies in developing and applying empirical Bayes (EB) models. The authors first apply standard EB methods to analyze RNA-seq data from the spinal cord and neocortex separately. The novelty comes with their proposed integration methods:

  • Weighted Gene Approach: This method assigns a weight to each gene based on its significance in each tissue's analysis. Genes with stronger signals in one tissue receive a higher weight, allowing the combined model to prioritize more informative genetic effects. The weighted gene expression is then used in the EB framework.
  • Weighted Z Method: This approach combines the Z-statistics (a measure of gene effect) from individual tissue analyses using weights derived from the absolute values of these Z-statistics. This aims to capture the relative importance of a gene's effect in each tissue to create a more robust combined statistic for EB analysis.

Both integration methods aim to leverage the complementary information from different tissue types to enhance the detection of disease-associated genes and improve the overall prediction performance.

Datasets

The study utilizes RNA-sequencing data from two distinct tissues of a mouse model of infantile hypophosphatasia:

  • Spinal Cord RNA-seq data: Collected from Alpl+/+ (wild type) and Alpl-/- (Global TNAP knockout) mice.
  • Neocortex RNA-seq data: Also collected from Alpl+/+ and Alpl-/- mice.

These datasets consist of approximately 36,500 gene markers and 16 samples (8 wild-type and 8 knockout). The data was cleaned to reduce dimensionality, resulting in 28,426 genes for analysis. The authors also generated simulated RNA-seq data using the SPsimSeq method to benchmark their proposed approaches.

Key Insights and Ideas

  • Tissue-Specific Signals Matter: The study highlights that disease-related genes might not be uniformly expressed across all tissues. Integrating data from different tissues (spinal cord and neocortex) is crucial for capturing a more complete picture of the genetic landscape of HPP.
  • Empirical Bayes for High-Dimensionality: EB methods are well-suited for handling the high-dimensional nature of RNA-seq data, providing more stable and less biased estimates compared to traditional frequentist approaches.
  • Integration Enhances Discovery: Combining data through weighted approaches significantly boosts the power to detect differentially expressed genes (DEGs) and improves the accuracy of disease risk prediction. The integration methods are more effective than analyzing each tissue in isolation.
  • Spinal Cord Data is More Informative: In this specific study, the spinal cord RNA-seq data generally showed stronger genetic signals and better performance than the neocortex data when analyzed individually.

Key Results

  • Improved Gene Discovery: The integration methods identified a substantially larger number of true DEGs compared to single-tissue analyses. For instance, the weighted gene approach found 53 DEGs among the top 100 genes, and weighted Z found 56, significantly outperforming the 21 from spinal cord and 5 from neocortex.
  • Reduced Prediction Error: The weighted Z method achieved the lowest average prediction error (0.1625), followed by the weighted gene approach (0.225), compared to spinal cord (0.275) and neocortex (0.3) alone.
  • Higher Prediction Accuracy: The weighted Z method demonstrated the highest prediction accuracy (AUC: 93.75%), with the weighted gene approach close behind (90%). This is a notable improvement over the individual tissue analyses (Spinal Cord: 87.5%, Neocortex: 85%).
  • Detection of Key Genes: The Alpl gene, a known causal gene for HPP, was successfully detected by all methods, but integration methods helped identify it and other potentially important genes more consistently and at higher ranks.

Limitations and Future Work

  • Small Sample Size: The authors acknowledge that the small sample size (n=16) is a significant limitation, potentially leading to unstable genetic estimates and requiring a less rigorous cross-validation procedure.
  • Future Directions:
    • Multi-omics Integration: The authors plan to integrate RNA-seq data with metabolomics data to gain a more comprehensive understanding of HPP's complex genetic and metabolic architecture.
  • New Theoretical Methods: To address the small sample size issue, they propose exploring Distributionally Robust Optimization (DRO) approaches for more stable genetic estimates.
  • Gene-Gene and Gene-Environment Interactions: Investigating these interactions is suggested as a beneficial next step for a more complete understanding of HPP etiology.

Summary updated

BibTeX
@inproceedings{kinsman2024new,
  author = {Dawson Kinsman and Jian Hu and Zhi Zhang and Gengxin Li},
  title = {{New Empirical Bayes Models to Jointly Analyze Multiple RNA-Sequencing Data in a Hypophosphatasia Disease Study}},
  booktitle = {Genes},
  year = {2024},
  url = {https://openreview.net/forum?id=G2GfhQ5ctj},
}
Figure