Background Correction Methods Implemented by Illumina
As discussed by Dunning et al. (2006), background correction is a critical step in microarray data analysis. However, there are a lot of confusions of how exactly the background correction was done under the hood. I investigated the problem by digging the manual published by Illumina and verified the findings with Dr. Jason Harger of Illumina technical support.
The confusion stemmed from the two-step data processing used by Illumina:
- Image processing by BeadScan
- Probe-summary by BeadStudio
- BeadScan corrects the intensity of each bead by subtracting the local background. Details are discussed in Dunning et al. (2006). The goal is to remove optical background using image-processing techniques.
- BeadStudio further corrects the nonspecific binding by subtracting the intensities of negative control (nonspecific) beads. The goal is to remove nonspecific binding using analytical chemistry. Dunning et al. (2006) did not mention this process in the data analysis.
- The BeadScan process (imaging correction of local subtraction) happens prior to the data analysis in BeadStudio (correction of nonspecific binding).
their signal represents an very good approximation of signal due to nonspecific hybridization. The number of negative control beadtypes is generally 700 - 1500 per array. Unlikely the mismatch probes (MMs) on Affymetrix arrays, we observed a more homogeneous behavior of the negative control beads of Illumina.
There is an option in BeadStudio to turn off nonspecific background correction. However, one cannot manipulate the process in BeadScan.
BeadStudio UserGuide (2007):
"Background is defined as the average signal intensity estimated
from the negative control bead types. Outliers are removed
using the median absolute deviation method.
Background is capable of minimizing the amount of variation in
background signals between arrays. Background subtraction can
be performed to remove signal due to nonspecific hybridization.
As a result, the expected signal for unexpressed targets is equal
to zero. Half of the unexpressed targets are expected to have
Dunning et al. (2006), Quality Control and Low-Level Statistical Analysis of Illumina BeadArrays. Revstat 4:1-. 30.