FDR-Corrected Sparse Canonical Correlation Analysis with Applications to Imaging Genomics

Alexej Gossmann, Pascal Zillé, Vince D. Calhoun, Yu-Ping Wang

Abstract

Reducing the number of false discoveries is presently one of the most pressing issues in the life sciences. It is of especially great importance for many applications in neuroimaging and genomics, where datasets are typically high-dimensional, which means that the number of explanatory variables exceeds the sample size. The false discovery rate (FDR) is a criterion that can be employed to address that issue. Thus it has gained great popularity as a tool for testing multiple hypotheses. Canonical correlation analysis (CCA) is a statistical technique that is used to make sense of the cross-correlation of two sets of measurements collected on the same set of samples (e.g., brain imaging and genomic data for the same mental illness patients), and sparse CCA extends the classical method to high-dimensional settings. Here we propose a way of applying the FDR concept to sparse CCA, and a method to control the FDR. The proposed FDR correction directly influences the sparsity of the solution, adapting it to the unknown true sparsity level. Theoretical derivation as well as simulation studies show that our procedure indeed keeps the FDR of the canonical vectors below a user-specified target level. We apply the proposed method to an imaging genomics dataset from the Philadelphia Neurodevelopmental Cohort. Our results link the brain connectivity profiles derived from brain activity during an emotion identification task, as measured by functional magnetic resonance imaging (fMRI), to the corresponding subjects' genomic data.

Introduction

In the era of high-dimensional data, fields such as imaging genomics face a significant challenge: identifying meaningful relationships between massive datasets where the number of features ($p$) vastly exceeds the number of samples ($n$). For instance, researchers may seek to correlate thousands of genetic markers (SNPs) with thousands of brain connectivity measures derived from fMRI. While Sparse Canonical Correlation Analysis (sCCA) has emerged as a standard tool for this task, it often operates as a "black box." Traditional sCCA implementations rely on heuristics or cross-validation to determine sparsity levels, which frequently leads to false discoveries—identifying features as significant when they are merely noise. This paper addresses the lack of theoretical guarantees in sCCA by introducing a statistically rigorous framework for False Discovery Rate (FDR) control.

Methodology

The authors propose an FDR-corrected sCCA procedure that shifts the focus from simple optimization to formal hypothesis testing. This framework is built upon three primary pillars:

  1. Asymptotic Normality: The mathematical foundation is a proof (Theorem 1) demonstrating that the distribution of the product of the datasets and a canonical vector ($X^T Y v$) converges to a multivariate normal distribution as the sample size increases.
  2. Three-Way Data Splitting: To prevent "double-dipping"—using the same data to both identify patterns and test their significance—the dataset is divided into three independent subsets. Subset 1 provides a preliminary, "liberal" estimate of the canonical vectors; Subset 2 is used to estimate the covariance structure; and Subset 3 is used to calculate p-values based on the asymptotic normality results.
  3. Statistical Correction: Finally, the standard Benjamini-Hochberg (BH) procedure is applied to these p-values, ensuring the FDR remains below a user-defined threshold ($q$).

While the three-way split reduces the raw statistical power by using only a fraction of the samples for the core CCA, it drastically enhances the reliability of the resulting discoveries.

Simulation Results

The method was rigorously tested using synthetic Gaussian data and "hybrid" synthetic data (pairing real SNP data with Gaussian noise). These simulations aimed to compare the proposed FDR-corrected sCCA against standard methods like 5-fold cross-validation and permutation-based sCCA.

FDR Control and Reliability

As shown in the simulation results, the proposed method consistently maintains the estimated FDR below the target level $q$. In many scenarios, the actual FDR reached only about $q/2$, demonstrating a conservative but highly effective control mechanism. In contrast, standard methods often exhibited FDR levels exceeding 0.80 in high-sparsity settings.

Estimated FDR for the proposed method vs. standard sCCA across varying sparsity levels.

Statistical Power (TPR)

Despite the conservative nature of the FDR control and the data-splitting requirement, the method maintains a competitive True Positive Rate (TPR). As the signal strength or sample size increases, the TPR of the FDR-corrected sCCA approaches that of less restrictive methods, while providing the necessary statistical guarantees that those methods lack.

Estimated TPR for the proposed method vs. standard sCCA across varying sparsity levels.

Application to Imaging Genomics

The practical utility of the framework was demonstrated using data from the Philadelphia Neurodevelopmental Cohort (PNC), involving 811 subjects. The analysis linked genomic data (reduced to gene-level features) with functional brain connectivity profiles from an emotion identification task.

The FDR-corrected sCCA identified 129 genes and 107 brain connectivity features. Notably, several identified genes, such as DAB1 and CNTNAP2, have well-documented associations with brain development, autism, and schizophrenia. This external biological validation suggests that the method is capable of uncovering true signal within complex, high-dimensional biological noise.

Conclusion and Future Work

This research provides a novel way to treat sparsity as a statistical decision rather than a tuning parameter. While the current implementation requires a "heavy" three-way data split that can limit power, it offers a level of reliability previously unavailable in sCCA. Future work will focus on more sample-efficient strategies, such as multi-split methods, and integrating direct control for confounding factors like population stratification in genomic data.

Summary updated

Abstract

Reducing the number of false discoveries is presently one of the most pressing issues in the life sciences. It is of especially great importance for many applications in neuroimaging and genomics, where datasets are typically high-dimensional, which means that the number of explanatory variables exceeds the sample size. The false discovery rate (FDR) is a criterion that can be employed to address that issue. Thus it has gained great popularity as a tool for testing multiple hypotheses. Canonical correlation analysis (CCA) is a statistical technique that is used to make sense of the cross-correlation of two sets of measurements collected on the same set of samples (e.g., brain imaging and genomic data for the same mental illness patients), and sparse CCA extends the classical method to high-dimensional settings. Here we propose a way of applying the FDR concept to sparse CCA, and a method to control the FDR. The proposed FDR correction directly influences the sparsity of the solution, adapting it to the unknown true sparsity level. Theoretical derivation as well as simulation studies show that our procedure indeed keeps the FDR of the canonical vectors below a user-specified target level. We apply the proposed method to an imaging genomics dataset from the Philadelphia Neurodevelopmental Cohort. Our results link the brain connectivity profiles derived from brain activity during an emotion identification task, as measured by functional magnetic resonance imaging (fMRI), to the corresponding subjects' genomic data.
BibTeX
@article{gossmann2018fdr,
  author = {Alexej Gossmann and Pascal Zillé and Vince D. Calhoun and Yu-Ping Wang},
  title = {{FDR-Corrected Sparse Canonical Correlation Analysis with Applications to Imaging Genomics}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2018},
  url = {https://openreview.net/forum?id=mAqt6cjkYi},
}
Figure