Covariance-Based Sample Selection for Heterogeneous Data: Applications to Gene Expression and Autism Risk Gene Detection

Kevin Lin, Han Liu, Kathryn RoederJournal of the American Statistical Association 2020 · DOI

TL;DR

This paper introduces COBS, a novel method for selecting homogeneous subsets of gene expression data from heterogeneous sources. By identifying samples with similar covariance structures, COBS enhances the power of existing "guilt-by-association" methods like DAWN for detecting autism risk genes, leading to more accurate and enriched gene discoveries.

What Problem Does This Paper Address?

Identifying genes associated with complex diseases like autism is challenging due to the vast number of potential genetic contributors and the heterogeneity of biological data. Gene expression data, for instance, can vary significantly based on tissue type, developmental stage, and other factors. Previous methods for identifying autism risk genes, such as DAWN, often rely on gene co-expression networks derived from gene expression data. However, these methods typically focus on a small, pre-selected subset of samples (e.g., from a specific brain region and developmental period) to avoid issues arising from data heterogeneity. This approach, while simplifying analysis, can lead to a loss of statistical power by excluding potentially valuable data. The authors aim to address this by developing a data-driven method to select a larger, more homogeneous subset of samples that share a common covariance structure, thereby improving the downstream analysis.

Methods

The core of the paper is the COBS (COvariance-Based sample Selection) method, a two-stage procedure:

  1. Stepdown Method for Covariance Testing: This stage uses a hypothesis testing framework to simultaneously assess the equality of covariance matrices across pairs of partitions (subsets of samples). It leverages a multiplier bootstrap approach to construct empirical null distributions, allowing for the control of the family-wise error rate (FWER) while accounting for dependencies between the numerous pairwise tests. This is a novel application of existing statistical techniques for high-dimensional covariance testing.
  2. Clique-Based Selection: After identifying pairs of partitions with statistically indistinguishable covariance matrices, this stage uses a graph-based approach to select the largest "quasi-clique." A quasi-clique is a set of vertices (partitions) in a graph where a high proportion of possible edges (accepted covariance equality tests) exist. This method aims to find the largest possible subset of partitions that are likely to share a common underlying covariance structure. A key aspect here is the development of an algorithm that exhibits a "monotone property," ensuring that as more null hypotheses are accepted (indicating greater homogeneity), the number of selected partitions does not decrease.

Datasets

The primary dataset used is the BrainSpan dataset, a large microarray expression dataset containing gene expression data from human brain tissues across various regions and developmental periods. The authors also utilize TADA scores (Transmission and De novo association) from two sources:

  • De Rubeis et al. (2014): Older TADA scores used for initial gene screening and for estimating HMRF model parameters.
  • Satterstrom et al. (2020): Newer TADA scores used to assess the quality and enrichment of the identified risk genes.

Key Insights and Ideas

  • Heterogeneity is a major hurdle: The paper strongly emphasizes that gene expression data, especially from biological tissues, is inherently heterogeneous. Ignoring this heterogeneity by selecting only a small subset of data can significantly limit the power of downstream analyses.
  • Covariance as a homogeneity marker: The authors propose that the population covariance matrix is a robust indicator of sample homogeneity. Samples sharing a similar covariance structure are more likely to come from the same underlying biological process.
  • The power of "guilt by association" with better data: By improving the quality of the input data (gene expression subsets) for methods like DAWN, COBS can significantly enhance their ability to discover meaningful biological insights, such as identifying autism risk genes.
  • Monotonicity is crucial for selection algorithms: The development of a clique-based selection algorithm that preserves monotonicity is a key contribution, ensuring that the selection process is stable and interpretable as the stringency of the statistical tests changes.

Key Results

  • Improved gene detection: COBS, when integrated with the DAWN framework, demonstrates an improved ability to predict risk genes. Specifically, it helps cluster newer TADA-identified risk genes closer to older ones in the gene network, suggesting a more robust and informative network.
  • Enrichment of risk genes: The risk genes implicated by COBS analysis show better enrichment for genes detected by newer TADA scores compared to the traditional Window 1B analysis (18.8% vs. 14.6%).
  • Robustness: The method shows robustness to the parameter $\gamma$ (controlling quasi-clique density), consistently implicating a substantial set of risk genes.
  • Performance comparison: Simulations show that COBS outperforms other methods, including spectral clustering and other quasi-clique finding algorithms, in identifying homogeneous partitions and maintaining the desired monotone property.

Limitations and Future Work

  • Theoretical understanding of FWER level $\alpha$: While the paper provides a theoretical guarantee for the set of accepted null hypotheses, the authors note that the theoretical role of the FWER level $\alpha$ for the selected set of partitions $\hat{P}$ is not fully understood and is a direction for future work. They suspect that modifications to COBS and "closed testing" might offer a theoretical guarantee for $\hat{P}$.
  • Computational cost of the original Stepdown method: The original Stepdown method can be computationally intensive. While an "Accelerated Stepdown" procedure is proposed to address this, it was found to lead to a loss of power in simulations and is not used in the main BrainSpan analysis. However, the authors suggest this accelerated approach might be useful in other settings.
  • Generalizability of the "Accelerated Stepdown" procedure: The authors suggest that the computational speedup idea from the Accelerated Stepdown procedure could be utilized more effectively in other covariance testing scenarios.
  • Broader implications: The authors highlight that COBS has broader implications beyond autism research, offering a statistical direction for selecting relevant samples from large, heterogeneous gene expression datasets in various scientific tasks.

Summary updated

Abstract

Risk for autism can be influenced by genetic mutations in hundreds of genes. Based on findings showing that genes with highly correlated gene expressions are functionally interrelated, "guilt by association" methods such as DAWN have been developed to identify these autism risk genes. Previous research analyze the BrainSpan dataset, which contains gene expression of brain tissues from varying regions and developmental periods. Since the spatiotemporal properties of brain tissue is known to affect the gene expression's covariance, previous research have focused only on a specific subset of samples to avoid the issue of heterogeneity. This analysis leads to a potential loss of power when detecting risk genes. In this article, we develop a new method called COBS (COvariance-Based sample Selection) to find a larger and more homogeneous subset of samples that share the same population covariance matrix for the downstream DAWN analysis. To demonstrate COBS's effectiveness, we use genetic risk scores from two sequential data freezes obtained in 2014 and 2020. We show COBS improves DAWN's ability to predict risk genes detected in the newer data freeze when using the risk scores of the older data freeze as input.
BibTeX
@article{lin2020covariance,
  author = {Kevin Lin and Han Liu and Kathryn Roeder},
  title = {{Covariance-Based Sample Selection for Heterogeneous Data: Applications to Gene Expression and Autism Risk Gene Detection}},
  journal = {Journal of the American Statistical Association},
  year = {2020},
  doi = {10.1080/01621459.2020.1738234},
}

Loading related content...

Figure