SNEP: Simultaneous detection of nucleotide and expression polymorphisms using Affymetrix GeneChip

Hironori Fujisawa, Youko Horiuchi, Yoshiaki Harushima, Toyoyuki Takada, Shinto Eguchi, Takako Mochizuki, Takayuki Sakaguchi, Toshihiko Shiroishi, Nori KurataBMC Bioinformatics 2009 · DOI

TL;DR

This paper introduces SNEP, a novel statistical method for simultaneously detecting both nucleotide (genetic variation) and expression polymorphisms from Affymetrix GeneChip data. This is crucial because traditional methods struggle to distinguish between these two sources of signal variation, and SNEP offers a robust solution that outperforms existing approaches.

What Problem Does This Paper Address?

High-density oligonucleotide microarrays, like Affymetrix GeneChips, are powerful tools for studying biological diversity. They can simultaneously measure two types of variations: nucleotide polymorphisms (differences in DNA sequence, often called Single Feature Polymorphisms or SFPs) and expression polymorphisms (differences in gene activity levels). However, a major challenge arises when comparing gene expression data between different strains or species. If these strains have different nucleotide sequences, it can lead to variations in how strongly their mRNA binds to the microarray probes. This makes it difficult to determine if a observed difference in signal intensity is due to a true change in gene expression or simply a consequence of the underlying genetic variation. Current methods typically analyze these two types of polymorphisms separately, which is inefficient and can lead to misinterpretations. This paper aims to bridge this gap by developing a method that can reliably detect both simultaneously.

Methods

The core of SNEP is a statistical model that treats nucleotide polymorphisms as "outliers" in the probe-level data. The authors hypothesize that if a probe set (a group of probes designed to measure a single gene) contains a nucleotide polymorphism, the signal intensity from that specific probe will deviate from the expected pattern across all probes for that gene. This deviation is attributed to the reduced hybridization efficiency caused by the genetic difference.

The model uses a robust statistical procedure to handle these potential outliers. Instead of relying on standard maximum likelihood estimation, which can be heavily influenced by outliers, SNEP employs a robust parameter estimation technique. This technique is designed to be less sensitive to extreme values, allowing for more reliable estimation of both nucleotide and expression differences.

Specifically, SNEP models the log-intensity of the probes using a basic statistical model that includes strain-specific means and probe effects. The key innovation is how it handles deviations from expected patterns. For nucleotide polymorphism detection, it tests hypotheses about the equality of probe effects within a probe set. For expression polymorphism detection, it tests the hypothesis that the overall gene expression level is the same between strains. The robustness is achieved through a divergence-based procedure that is less susceptible to the influence of SFPs.

Datasets

The authors evaluated SNEP using three distinct datasets:

  • Barley data: Publicly available data from different barley tissues. This dataset was used to compare SNEP with existing methods like Rostoks et al. [10] and Greenhall et al. [14].
  • Rice data: Newly generated data from two rice cultivars (japonica "Nipponbare" and indica "93-11"). This dataset was particularly valuable because the full genome sequences were available, allowing for more detailed analysis and validation. They also created a "canonical rice data" subset where probe sequences were perfectly matched to single copies in both genomes to isolate SFP effects.
  • Mouse data: Newly generated data from two inbred mouse strains (C57BL/6J and MSM/Ms). This dataset allowed for evaluation on a larger and more complex genome.

Key Insights and Ideas

  • Nucleotide polymorphisms as outliers: The central idea that SFPs manifest as deviations (outliers) in probe-level signal intensities is a clever way to frame the problem for robust statistical analysis.
  • Simultaneous detection is key: The paper convincingly demonstrates that trying to disentangle nucleotide and expression polymorphisms is difficult and error-prone if done separately. SNEP's integrated approach is a significant advantage.
  • Robustness matters: The authors highlight that standard statistical methods can be heavily skewed by SFPs when detecting expression differences. SNEP's robust nature protects against this bias.
  • Normalization is not always sufficient: For cross-strain or cross-species comparisons, normalization might not fully equalize hybridization affinities due to underlying genetic differences. SNEP's approach bypasses the need for normalization in this context.
  • Gene selection is critical: The performance of SNEP (and likely other methods) is sensitive to how genes are selected for analysis, particularly regarding signal intensity levels and the presence of multiple SFPs within a probe set.

Key Results

  • Superior SFP detection: SNEP consistently outperformed existing methods (LRT, Rostoks et al., Greenhall et al.) in detecting nucleotide polymorphisms (SFPs) across barley, rice, and mouse datasets, often showing significantly higher sensitivity and lower false positive rates.
  • Robust expression detection: SNEP demonstrated greater robustness in detecting gene expression differences when SFPs were present, compared to standard t-tests based on Tukey's biweight estimates. This was illustrated with rice data where SNEP's test statistic was less affected by SFPs than the standard method.
  • Performance across genomes: SNEP performed well across different genome sizes and levels of sequence information availability, indicating its broad applicability.
  • Impact of signal intensity: For rice data, SNEP's sensitivity and FPR were significantly affected by the median signal intensity of probe sets. Genes with higher median intensities (above 2.5) yielded more stable and reliable results.

Limitations and Future Work

  • Gene selection sensitivity: The authors acknowledge that the way genes are selected for analysis (e.g., based on signal intensity thresholds) is an important factor for stabilizing SFP detection. Further work could explore optimal gene selection strategies.
  • Multiple SFPs: While SNEP is designed to handle outliers, the paper notes that when multiple SFPs exist within a single probe set, the ratio of outliers can become significant, potentially impacting performance. The authors suggest that SNEP's divergence-based procedure is still robust in these cases, but it's an area for further investigation.
  • Software availability: The R-software 'SNEP' is provided, which is a positive step for adoption. However, further development and community support could enhance its usability and reach.
  • Comparison with next-generation sequencing: While acknowledging the power of next-generation sequencing for variation and expression analysis, the paper positions array technologies like Affymetrix GeneChips as a cost-effective option, suggesting that SNEP remains relevant for current research needs. Future work could explore integrating SNEP with data from newer sequencing technologies.

Summary updated

BibTeX
@article{fujisawa2009snep,
  author = {Hironori Fujisawa and Youko Horiuchi and Yoshiaki Harushima and Toyoyuki Takada and Shinto Eguchi and Takako Mochizuki and Takayuki Sakaguchi and Toshihiko Shiroishi and Nori Kurata},
  title = {{SNEP: Simultaneous detection of nucleotide and expression polymorphisms using Affymetrix GeneChip}},
  journal = {BMC Bioinformatics},
  year = {2009},
  doi = {10.1186/1471-2105-10-131},
}

Loading related content...

Figure