TL;DR
This paper introduces novel nonparametric methods to estimate the genewise variance of gene expression in microarray data. These methods are crucial for accurately identifying differentially expressed genes and validating normalization procedures, especially when dealing with correlated measurements and a large number of genes relative to samples.
What Problem Does This Paper Address?
Microarray experiments allow scientists to measure the expression levels of thousands of genes simultaneously. A key challenge in analyzing this data is accurately estimating the "genewise variance" – the variability of a specific gene's expression. This estimation is critical for two main tasks:
- Identifying Differentially Expressed Genes: When comparing gene expression between different conditions (e.g., healthy vs. diseased), we need to know if the observed differences are statistically significant or just due to random variation. Accurate variance estimation improves the power of statistical tests (like the t-test) used for this purpose.
- Validating Normalization: Microarray data often requires normalization to remove systematic biases. Estimating genewise variance helps in assessing whether these biases have been effectively removed or in selecting the best normalization method.
The problem is particularly challenging because:
- High Dimensionality: We have many genes (thousands) but often a limited number of experimental replicates (samples).
- Nuisance Parameters: Each gene has its own mean expression level, which acts as a nuisance parameter that needs to be accounted for.
- Correlated Measurements: Replicates within an array or across different arrays can be correlated, making standard variance estimation techniques unreliable.
Existing methods often struggle with these issues, leading to biased estimates or reduced statistical power.
Methods
The authors propose a "two-way nonparametric model" that extends the classic Neyman-Scott model. The core of their approach involves developing novel nonparametric estimators for the genewise variance function and semiparametric estimators for measurement correlation.
Key methodological contributions include:
- Two-Way Nonparametric Model: They model the log-ratio of gene expression ($Y_{gi}$) as a function of the gene's average expression level ($X_{gi}$) and an error term ($\epsilon_{gi}$), where the variance of the error term is a smooth function of the average expression level: $Y_{gi} = \alpha_g + \sigma(X_{gi})\epsilon_{gi}$. This allows the variance to vary with the intensity level.
- Nonparametric Variance Estimation: They develop estimators for the variance function $\sigma^2(x)$ by solving systems of nonlinear equations derived from the expected values of squared differences in expression levels.
- Correlation Estimation: For correlated replications, they propose an iterative algorithm to simultaneously estimate the variance function, the mean parameters ($\sigma_1, \sigma_2$), and the correlation coefficient ($\rho$).
- Local Linear Regression: They leverage local linear regression techniques to estimate the smooth variance function $\sigma^2(x)$.
- Aggregated Estimator: When within-array replications are not available, they propose an aggregated estimator that treats all available measurements as replications.
Datasets
The methodology is illustrated and evaluated using:
- Simulation Studies: Extensive simulations are conducted to assess the finite sample performance of their proposed estimators under various scenarios, including different correlation structures and gene effect functions.
- MicroArray Quality Control (MAQC) Project Data: Real data from the MAQC project, specifically using Agilent arrays, is used for validation tests and gene selection applications. This dataset involves two-color microarrays with multiple sites and dye-swapped experiments.
Key Insights and Ideas
- Variance is Intensity-Dependent: A crucial insight is that genewise variance is not constant but systematically depends on the average intensity level of the gene. This non-trivial relationship needs to be modeled nonparametrically.
- Correlation Matters: Ignoring correlation among replicates can lead to biased variance estimates and reduced power in downstream analyses. The proposed methods explicitly account for this correlation.
- Nuisance Parameters are Manageable: The paper demonstrates that even with a large number of nuisance parameters (one for each gene's mean expression), it's possible to consistently estimate the variance function without making strong assumptions like sparsity or smoothness on these nuisance parameters.
- Improved Statistical Power: The new variance estimators lead to more powerful tests for detecting differentially expressed genes, meaning they can identify more true positives and potentially reduce false negatives.
- Unified Framework: The proposed model and estimators are flexible enough to handle both within-array and across-array replications, unifying different experimental designs under a single theoretical framework.
Key Results
- Asymptotic Normality: The paper establishes the asymptotic normality of the proposed nonparametric estimators, providing theoretical guarantees for their performance.
- Simulation Performance: Simulation studies show that the proposed estimators significantly reduce bias compared to naive nonparametric approaches, especially when the gene effect is not smooth. The estimators also exhibit good performance in terms of variance and mean squared error.
- Real Data Application (MAQC):
- Validation Tests: Applying the method to MAQC data showed that, in most cases, arrays did not require further normalization, aligning with previous findings. However, some arrays were identified as needing normalization, particularly when dye-swap effects reduced the estimated correlation.
- Gene Selection: The z-test using the new variance estimator identified a larger number of differentially expressed genes compared to the standard t-test, demonstrating improved power. The empirical power increase was found to be close to the expected theoretical power increase.
- Computational Efficiency: The proposed algorithm is computationally efficient, taking only a few minutes, compared to other methods like SIMEX which can take over an hour.
Limitations and Future Work
- Assumptions: The theoretical results rely on certain regularity conditions for the variance function (bounded and continuous second derivative) and the kernel function (bounded symmetric density with bounded support).
- Future Work: The authors suggest possible extensions, including:
- Applying SIMEX-like ideas to account for measurement error.
- Developing methods for cases with prior correlation structures among replications that are not identical.
- Extending the model to handle more complex experimental designs.
Summary updated