[Paper Review] Correlation-sharing for detection of differential gene expression

Robert Tibshirani, Larry Wasserman|ArXiv.org|Aug 2, 2006

Gene expression and cancer classification10 references22 citations

TL;DR

This paper proposes correlation-sharing, a method that improves differential gene expression detection by averaging t-statistics of genes within correlation neighborhoods, leveraging biological pathway co-expression. It reduces false discovery rates compared to standard t-statistic thresholding, especially when differentially expressed genes are correlated, and extends to other data types like protein mass spectrometry.

ABSTRACT

We propose a method for detecting differential gene expression that exploits the correlation between genes. Our proposal averages the univariate scores of each feature with the scores in correlation neighborhoods. In a number of real and simulated examples, the new method often exhibits lower false discovery rates than simple t-statistic thresholding. We also provide some analysis of the asymptotic behavior of our proposal. The general idea of correlation-sharing can be applied to other prediction problems involving a large number of correlated features. We give an example in protein mass spectrometry.

Motivation & Objective

• To improve detection of differentially expressed genes in microarray data by exploiting correlations among genes.
• To reduce false discovery rates (FDR) in gene selection compared to conventional t-statistic thresholding.
• To develop a robust, simple method that adapts to local correlation structure without requiring arbitrary clustering parameters.
• To extend the approach to non-two-class outcomes, such as survival data and quantitative responses.
• To validate the method on real and simulated microarray datasets and a protein mass spectrometry dataset.

Proposed method

• Compute standard two-sample t-statistics for each gene.
• For each gene, define a correlation neighborhood $ C_\rho(i) $ as genes with correlation $ \geq \rho $ to gene $ i $.
• For each gene, compute $ u_i = \max_{\rho \in [0,1]} \text{ave}_{j \in C_\rho(i)} |T_j| $, selecting the $ \rho $ that maximizes the average absolute t-statistic.
• Define the correlation-shared score as $ r_i = \text{sign}(T_i) \cdot u_i $, preserving directionality.
• Identify significant genes as those with $ |r_i| > c $, using permutation-based FDR estimation.
• Apply the method to two-class and survival outcomes by replacing the t-statistic with appropriate score statistics (e.g., partial likelihood for survival).

Experimental results

Research questions

• RQ1 Can leveraging gene correlation improve the detection of differentially expressed genes compared to univariate t-statistic thresholding?
• RQ2 How does the performance of correlation-sharing depend on the residual correlation structure between non-null and null genes?
• RQ3 What is the optimal neighborhood size and correlation threshold $ \rho $ for maximizing detection power?
• RQ4 Does correlation-sharing maintain low false discovery rates when the true differential expression is sparse and clustered?
• RQ5 Can the method be generalized to non-two-class outcomes such as survival or continuous responses?

Key findings

• In simulated data with correlated non-null genes, correlation-sharing reduced both false positives and false negatives compared to standard t-statistic thresholding.
• When non-null genes were uncorrelated before the group effect, the advantage of correlation-sharing disappeared, confirming that correlation among non-nulls is essential.
• On four real cancer datasets (skin, breast, BRCA, lymphoma), correlation-sharing consistently reduced FDR and improved power across varying thresholds.
• In the skin cancer dataset, 100 genes were selected with correlation-sharing, and 18 of them were not detected by the t-statistic alone, indicating improved sensitivity.
• For protein mass spectrometry data, correlation-sharing identified more relevant peaks by averaging signals from neighboring peaks with high correlation.
• Asymptotic analysis shows that correlation-sharing can improve non-centrality in the test statistic when the neighborhood size is chosen adaptively, especially in typical correlation structures.