Abstract
We have designed the SWTP and the SWPT-Bi which are using the stationary wavelet packet transform with the hard thresholding and the new bivariate shrinkage estimator respectively to smooth the array CGH data. We demonstrate the effectiveness of our approach through theoretical and experimental exploration of a set of array CGH data, including both synthetic data and real data. The comparison results show that our method outperforms the previous approaches.
Introduction
In cancer research, array-based comparative genomic hybridization (array CGH) is a vital tool for identifying DNA copy number variations (CNVs), such as deletions and duplications. These variations serve as critical biomarkers for diagnosis and understanding tumor progression. However, array CGH data is inherently noisy, often presenting a signal-to-noise ratio that makes identifying clear "step" signals—representing genomic gains or losses—extremely difficult.
Traditional smoothing techniques, such as Lowess or standard wavelets, often face a trade-off: they either fail to remove enough noise or "over-smooth" the data, blurring the sharp boundaries of genomic aberrations. Because these aberrations typically manifest as step functions, preserving sharp transitions is essential for accurate downstream analysis.
Methodology
This paper proposes two advanced denoising approaches: SWPT (Stationary Wavelet Packet Transform with hard thresholding) and SWPT-Bi (SWPT combined with a novel bivariate shrinkage model).
Stationary Wavelet Packet Transform (SWPT)
Unlike the standard Discrete Wavelet Transform (DWT), which primarily decomposes low-frequency components, the SWPT analyzes the signal across the entire frequency spectrum equally. This is crucial for CGH data because the sharp "steps" of genomic changes contain significant high-frequency information that standard wavelets might discard as noise. Furthermore, SWPT is shift-invariant, which prevents the creation of artificial ripples or artifacts around discontinuities.
Bivariate Shrinkage Estimator (SWPT-Bi)
Most denoising algorithms assume that wavelet coefficients are independent. The authors argue that in packet transforms, "cousin" coefficients (coefficients at the same level but in different sub-bands) provide powerful context. The SWPT-Bi model exploits these dependencies to make more accurate distinctions between true signal and random noise.
Experimental Evaluation
The researchers validated their approach using both synthetic and biological datasets. Synthetic data consisted of 1,000 artificial chromosomes with varying noise levels ($\sigma_n$ from 0.1 to 0.275), while real-world validation utilized BAC array data from fibroblast cell lines (GM13330).
Key Results
The proposed SWPT-Bi method consistently outperformed all competitors across all noise levels:
- RMSE Reduction: SWPT-Bi reduced the Root Mean Squared Error (RMSE) by 43.4% – 55% compared to Lowess and by 50.3% – 57.9% compared to Quantreg.
- Superiority over Wavelets: It outperformed previous state-of-the-art wavelet methods (like SWTi) by 26.8% – 35.3%.
- Visual Clarity: In real-world tests, SWPT-Bi produced the smoothest horizontal segments while maintaining the sharpest vertical transitions at genomic break points.
Conclusion
By utilizing the Stationary Wavelet Packet Transform and accounting for coefficient dependencies through a bivariate shrinkage estimator, the authors have developed a more precise method for cleaning noisy genomic data. While the current work focuses on smoothing, future research aims to integrate these denoising capabilities into automated segmentation algorithms for clinical aberration detection.
Summary updated