Tiling array data analysis: a multiscale approach using wavelets

Alexander Karpikov, Joel S. Rozowsky, Mark GersteinBMC Bioinformatics 2011 · DOI · OpenReview

Abstract

ChIP-chip tiling arrays are essential for mapping protein-DNA interactions but suffer from significant noise due to non-specific hybridization. We present a multiscale denoising approach using the Discrete Wavelet Transform (DWT). By employing "Coiflet" wavelets that match biological peak shapes and enforcing cross-scale consistency, this method effectively separates signal from noise. Validation on ENCODE datasets demonstrates superior performance in identifying both narrow and broad binding sites compared to existing tools.

Introduction

ChIP-chip experiments are a cornerstone of modern genomics, enabling the identification of transcription factor binding sites and histone modifications across the genome. However, the resulting data is notoriously noisy, primarily due to non-specific cross-hybridization of probes. Existing analysis methods, such as moving averages or fixed-thresholding, often struggle to distinguish true biological signals from random background noise, particularly when signals vary in scale from narrow peaks to broad domains.

Methods: Multiscale Wavelet Analysis

To address these challenges, we developed a denoising framework based on the Discrete Wavelet Transform (DWT). Unlike Fourier transforms, wavelets provide localization in both frequency and position, making them ideal for identifying genomic peaks.

Shape Matching and Decomposition

We specifically utilized "Coif1" wavelets. The mathematical shape of a Coiflet is nearly symmetrical and triangular, closely mimicking the physical profile of a ChIP-chip binding peak. The algorithm decomposes the input signal into multiple resolution scales. As illustrated in Figure 4, the signal is broken down into approximation coefficients (A) and detail coefficients (D) across several levels.

Multiscale decomposition of a noisy signal using wavelets.

Consistency and Thresholding

A key innovation is the "consistency constraint": a biological signal is only considered valid if it appears consistently across multiple resolution scales. Noise, which typically resides in the fine-grained detail scales, is filtered out if it lacks this cross-scale persistence. Additionally, we determined that noise from non-specific binding follows a log-normal distribution. This allows for the application of statistically grounded thresholds rather than arbitrary cutoffs.

Results

The method was validated using ENCODE spike-in benchmarks and real-world data for Pol II and various histone modifications.

Performance Benchmarking

In comparisons using ROC curves, the wavelet-based approach consistently achieved the highest scores among seven popular analysis tools. The multiscale nature of the DWT allows the algorithm to handle both narrow transcription factor peaks and broad histone modification regions with high precision.

Statistical Validation of Noise

Analysis of the wavelet coefficients confirmed the log-normal nature of the background noise. Figure 6 demonstrates the distribution of these coefficients across different scales, supporting the use of log-normal thresholding to separate signal from noise effectively. The consistency check alone was found to eliminate approximately 84% of the noise typically found in tiling array datasets.

Distribution of wavelet coefficients and log-normal thresholding across scales.

Discussion

The multiscale wavelet approach offers a robust and computationally efficient ($O(N)$) solution for tiling array analysis. By incorporating biological "shape" into the mathematical model and leveraging cross-scale consistency, the method significantly improves signal detection. While currently focused on Nimblegen arrays, the statistical principles of this approach are potentially applicable to other platforms and emerging technologies like ChIP-seq. Future work will aim to automate the selection of optimal scale ranges for different experimental designs.

Summary updated

Abstract

Background Tiling array data is hard to interpret due to noise. The wavelet transformation is a widely used technique in signal processing for elucidating the true signal from noisy data. Consequently, we attempted to denoise representative tiling array datasets for ChIP-chip experiments using wavelets. In doing this, we used specific wavelet basis functions, Coiflets, since their triangular shape closely resembles the expected profiles of true ChIP-chip peaks. Results In our wavelet-transformed data, we observed that noise tends to be confined to small scales while the useful signal-of-interest spans multiple large scales. We were also able to show that wavelet coefficients due to non-specific cross-hybridization follow a log-normal distribution, and we used this fact in developing a thresholding procedure. In particular, wavelets allow one to set an unambiguous, absolute threshold, which has been hard to define in ChIP-chip experiments. One can set this threshold by requiring a similar confidence level at different length-scales of the transformed signal. We applied our algorithm to a number of representative ChIP-chip data sets, including those of Pol II and histone modifications, which have a diverse distribution of length-scales of biochemical activity, including some broad peaks. Conclusions Finally, we benchmarked our method in comparison to other approaches for scoring ChIP-chip data using spike-ins on the ENCODE Nimblegen tiling array. This comparison demonstrated excellent performance, with wavelets getting the best overall score.
BibTeX
@article{karpikov2011tiling,
  author = {Alexander Karpikov and Joel S. Rozowsky and Mark Gerstein},
  title = {{Tiling array data analysis: a multiscale approach using wavelets}},
  journal = {BMC Bioinformatics},
  year = {2011},
  doi = {10.1186/1471-2105-12-57},
}
Figure