Microarray Denoising

Document Sample
Microarray Denoising Powered By Docstoc
					A Novel Approach to Improve
Reliability of Gene Expression
  Values from Microarrays
Promises of Microarray Technology
  Use of gene expression values as a basis for:

      Disease Diagnosis
      Individualized Treatment Planning

      Drug Discovery




A Problem…..

Microarray data set contains significant noise that
prevents accurate estimation of gene expression values
Types of Microarray Noise
Sample Preparation Noise
        Random Nature of RNA Amplification
Hybridization Noise
        Probabilistic Nature of Hybridization
Scanning Noise
      Variation in laser intensity during chip
       reading
      Leak of external light during chip reading

      Presence of dust on glass slide
 Current Denoising Approaches and their
             Inadequacies
Approaches to reduce noise induced during hybridization
      Limitations: Ignores the noise added during scanning process
            Vikaloet. al., IEEE trans on signal processing, 2006.
            Hassibi et. al.,Journal of Applied Physics, 2005.



Approaches that develop noise models for microarrays used in
identifying differentially expressed genes
      Limitations: Noise model not suitable for other applications

            Lukac  et. al., IEEE trans. on Nanobioscience, 2004.
            Aris et. al., BMC Bioinformatics, 2004.




Denoising approach using standard 2-D multiresolution analysis
       Limitations: Suitable only for natural images

            Wang     et. al., IEEE trans. on Nanobioscience, 2003
  A Novel Multiresolution Approach

We develop a novel denoising approach
using 2-D multiresolution analysis that
removes noise from both hybridization and
scanning processes
     Considers a special characteristic of microarray
      images that is different from natural images
     Incorporates probabilistic properties of different
      noise components
      Wavelet Multiresolution Representation of Signals
One-Dimensional Signal Decomposition
                                         Approximation
        Decompose                         Coefficients                     Reconstruct




                                             Detail
                                           Coefficients

Two-Dimensional Signal Decomposition
                                                                                         (Approximation)

                                                                                         (Horz. Detail)
              Along Rows                       Along Columns
                                                                                         (Vert. Detail)

                                                                                         (Diag. Detail)




               N.Kingsbury,”Image Processing with Complex Wavelets, 1997
Representing Signal at Different Frequencies
        Microarray Noise Comprises of
      Components at Different Frequencies
Original Image
  A Novel Multiresolution Approach

We develop a novel denoising approach
using 2-D multiresolution analysis that
removes noise from both hybridization and
scanning processes
     Considers   special characteristic of
        microarray images that is different from
        natural images
       Incorporates probabilistic properties of different
        noise components
                                   Microarray Image

   Microarray Chip :
   1.28cm X 1.28cm
   ~1,000,000 squares


                 Probes for one gene                                                 8000
                                                                                     7000
                                                                                     6000
                                                                                     5000
                                                                                     4000
                                                                                     3000
                                                                                     2000
                                                                                     1000
                                                                                        0



                                                                                            Probe Intensities
Microarray images contain a large number of edges
http://www.affymetrix.com/corporate/media/image_library/low_res/single_feature.jpg
Natural Image




       Contains Few Edges
        Creating Sub-images
  All probes assigned to a gene




Microarray Chip
        Creating Sub-Images
 Replace original probes
Take data set of all probes
 with denoised probes
                                          Create a Dyadic Square
                                                Sub-Image




                                                          Decompose using
                                                             Wavelets
                          Thresholding
                               &
                         reconstruction
                                                  Detail coefficients
                                                     at different
    Denoised Image                                  frequencies


   Repeat for all Genes
  A Novel Multiresolution Approach

We develop a novel denoising approach
using 2-D multiresolution analysis that
removes noise from both hybridization and
scanning processes
       Considers special characteristic of microarray
        images that is different from natural images
     Incorporates    probabilistic properties of
        different noise components
                       Noise Properties
Bivariate Shrinkage thresholding given by Sendur et. al., IEEE Transactions on
Signal Processing, 2002, considers noise to be Normal(0,σ)


                            Microarray Noise


Poisson Noise                                      Normal Noise
•Random nature of Hybridization
                                                   •Noise from Scanning Process
 (Tu et. al., PNAS, 2002)




                              Normal Transformation
Normal Noise
                                   of Poisson
        +                                                   Normal Noise
                            Poisson (λ) ~ Normal (λ, √λ)
Poisson Noise
 A Novel Multiresolution Approach
                                    Steps
Step 1: Create sub-images (putting together all probes of each gene) of the microarray.
Step 2: Take one sub-image a time for denoising and follow the steps below
Step 3: For each probe square obtain estimate of noise
        Intensity of each pixel of probe= Constant expression value E+ Noise
        Taking median as the estimator of E take difference of median and pixel intensity to obtain
         noise estimate
Step 4: Taking normal approximation of Poisson the estimated noise components consist
   of two normal distributions N(0,σ) and N(λ, √λ).
Step 5: As per Maximum Likelihood Estimator λ= Sample mean
    1.   Here sample consists of both N(0,σ) and N(λ, √λ). However, we know that:
    2.   Mean (sum of two independent normal distributions) =sum of individual means
    3.   Implies λ = mean of estimated noise
Step 6: Using the value of λ apply standard normal conversion on the noise estimate to
   get a total noise component as N(0,σt)
Step 7: Add back the noise estimate to the median using appropriate sign
Step 8: Apply wavelet decomposition to sub-image to obtain components at different
   frequencies
Step 9: Apply Bivariate Shrinkage thresholding for identification and elimination of noise
Step 10: Repeat above steps for each sub-image
Step 11: Reconstruct the microarray image by sending back the probe squares to the
   respective locations
Step 12: This image undergoes further analysis by using current software to get final
   gene values after analysis of perfect match and mis-match probe squares
       Evaluation of Methodology [On-going]
                                                                            Poisson+       Poisson+       Poisson+
                                                        Noise Added         Normal(0,10)   Normal(0,20)   Normal(0,30)

                                                                                  18.57          16.57            15.8
    Phase I:
                                                        Noisy Data
                Testing the methodology on a
    fabricated sub-image
                                                        Denoising without
                                                        modification for
     The Table shows the values of PSNR                poisson noise             22.45          18.49          18.07
     before and after denosing                          Denoising with
                                                        modification for
                                                        poisson noise             33.32          32.11          31.14




   Phase II:     Applying methodology on Affymetrix HG-U133 Plus 2.0 arrays on HCT-116 cell
    line and performing statistical tests to evaluate performance

Test Strategy 1: To test reduction of scanning noise         Test Strategy 2: To test reduction of hybridization
                                                             noise
Multiple scans of a microarray chip were obtained
                                                             Yet to be done: Hybridize and scan multiple arrays
and denoised seperately
                                                             from same sample (divide sample into multiple
                                                             portions after preparation). Denoise individually and
Coefficient of Variation (CV) between multiple scans        use CV to test reduction in CV across arrays
was used to analyze the reduction in scanning noise.

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:14
posted:9/10/2012
language:English
pages:17