TL;DR
This paper introduces a novel semiparametric regression model for analyzing oligonucleotide microarray data, specifically addressing limitations in existing methods by employing spline smoothing to capture non-linear relationships between perfect match (PM) and mismatch (MM) probe intensities. The proposed method, LWS, demonstrates improved accuracy and stability in estimating gene expression indexes compared to the widely used Li-Wong reduced (LWR) model.
What Problem Does This Paper Address?
DNA microarrays, particularly high-density oligonucleotide arrays, are powerful tools for simultaneously monitoring gene expression levels across thousands of genes. A key challenge in analyzing this data is accurately estimating gene expression indexes, which are crucial for downstream analyses like identifying important genes or classifying disease types. Existing methods, such as the Li-Wong reduced model (LWR), often assume a linear relationship between perfect match (PM) and mismatch (MM) probe intensities. However, empirical evidence, as shown in Figure 1 of the paper, reveals that this relationship can be non-linear, with curvature patterns that are not adequately captured by linear models. This can lead to suboptimal model fitting and potentially miss important biological information. Furthermore, some existing methods employ "brutal truncation" of data points where MM intensities exceed PM, which might discard valuable biological signals.
Methods
The authors propose a semiparametric regression model, termed LWS (Li-Wong-Spline), that builds upon the Li-Wong reduced model but incorporates spline smoothing. The core idea is to model the relationship between PM and MM intensities using a nonparametric cubic spline function, denoted as $g(\cdot)$. This allows for a flexible functional form for the MM intensities, avoiding the restrictive assumption of linearity. The model is expressed as:
$PM_{ij} = g(MM_{ij}) + \theta_j \phi_i + \epsilon_{ij}$
where $\theta_j$ represents the gene expression index for gene $j$, $\phi_i$ is a probe-specific effect, and $\epsilon_{ij}$ is the error term. The function $g(\cdot)$ is estimated using cubic spline smoothing, which minimizes a penalized residual sum of squares. This approach allows for capturing non-linearities without specifying a rigid parametric form. The estimation procedure involves a three-stage iterative algorithm:
- Stage 1: Obtain initial estimates of $\theta_j$ and $\phi_i$ using the existing Li-Wong reduced (LWR) model.
- Stage 2: Fit a nonparametric cubic spline model to estimate $g(MM_{ij})$ using the current estimates of $\theta_j$ and $\phi_i$.
- Stage 3: Update the PM values using the estimated $g(MM_{ij})$ and re-estimate $\theta_j$ and $\phi_i$ using a regression approach. This process is repeated until convergence.
Datasets
The study utilizes a dataset of 18 HuGeneFL arrays from an experiment conducted at the Division of Human Cancer Genetics at the Ohio State University. This dataset includes control genes (BioB, BioC, BioD, Cre) and spiked-in genes (Lys, Phe, Dap, Thr) under different experimental conditions (stimulated, starved, and 50:50 mixture). The experimental design is described in detail in the Appendix, including the specific concentrations of spiked-in RNAs.
Key Insights and Ideas
- Non-linearity is the norm, not the exception: The paper strongly demonstrates that the relationship between PM and MM probe intensities is often non-linear, a crucial observation that existing linear models overlook. This non-linearity can be a source of significant information about gene expression.
- Spline smoothing offers flexibility: By employing cubic spline smoothing, the LWS model gains the flexibility to adapt to various non-linear patterns in the data, leading to a more accurate representation of the underlying biological processes.
- Improved estimation stability: The LWS model consistently produces more stable estimates of gene expression indexes, as evidenced by smaller standard errors and variances across different experimental conditions. This increased stability is vital for reliable downstream analyses.
- Data-driven approach: The semiparametric nature of LWS means it makes fewer assumptions about the functional form of the MM intensities, making it more robust to model mis-specification and more data-driven.
Key Results
- Reduced variance in expression indexes: The LWS model consistently yields smaller sample variances of gene expression indexes among replicated arrays compared to the LWR model, particularly in the "Stimulated" and "50:50" conditions. For instance, the mean sample variance for LWS was 369,334.5 in the stimulated condition, while LWR was 384,779.9.
- Lower standard errors: The standard errors of estimated gene expression indexes from LWS are generally much smaller than those from LWR. Histograms in Figure 2 show that most LWS standard errors fall within (0, 500), whereas LWR's can exceed 1000.
- Higher correlations: The correlations between observed and predicted PM intensities are higher for LWS, indicating a better model fit. Most correlations for LWS are between 0.92 and 1, while LWR's can drop below 0.90 (Figure 4).
- More accurate ranking of unexpressed genes: When assessing the ranks of unexpressed control genes, LWS provides much lower median ranks and sums of ranks compared to LWR, suggesting it better identifies genes that should have minimal expression (Table 2 and 3).
- Lower coefficient of variation (C.V.): A comparison of coefficients of variation shows that LWS has a smaller average median C.V. (0.088) compared to LWR (0.094), indicating more consistent expression index estimates relative to their means.
Limitations and Future Work
- Incorporating correlation structures: The authors note that current model-based methods, including their own, largely assume independent variation, which is biologically unrealistic for co-regulated genes. They suggest that a new methodology to incorporate correlation structures is needed.
- Goodness-of-fit criteria: The paper highlights the lack of standard criteria for justifying and comparing models for gene expression data, indicating an area for future research.
- Alternative nonparametric techniques: While the proposed model uses cubic spline smoothing, the authors suggest that other nonparametric techniques like kernel smoothing could also be applied to fit MM intensities.
- Adaptive approach: They propose an adaptive approach where LWR is first tested for goodness-of-fit. If LWR is not sufficient, then LWS (spline) is used. This suggests a potential pathway for automated model selection.
Summary updated