to BioTechniques free email alert service to receive content updates.
Median filter algorithm for estimating the threshold of detection on custom protein arrays
David D. Smith1, Susan Kovats2, Terry D. Lee1, and Leticia Cano3
1City of Hope National Medical Center and Beckman Research Institute, Duarte, CA
2Oklahoma Medical Research Foundation, Oklahoma City, OK
3NHLBI, NIH, Bethesda, MD, USA
BioTechniques, Vol. 41, No. 1, July 2006, pp. 74–78
Full Text (PDF)
Supplementary Material
Smith411Supl (.pdf)

We constructed protein arrays according to a titration design to estimate the assay sensitivities over varying concentrations of flu vaccine and human immuno globulin G (IgG). After imaging, we considered the problem of appropriately distinguishing background noise from foreground signal. We applied the median filter smoothing technique and estimated the differences of the observed signal compared to the smoothed signal. If the absolute value of the difference was large, the feature was easily detectable, indicating that the spot did not blend with its surrounding neighbors. After estimating the residuals, we applied thresholding algorithms to estimate the limits of detection for each assay. At sufficiently large smoothing spans, our median filter approach performed as well or better than visual inspection and two other competing analysis methods. This suggests that a median filter approach has utility in high-throughput arrays where visual inspection is impractical.


Protein arrays have proven to be valuable tools in a wide variety of settings. The high-throughput features of array technology allow for the screening of thousands of proteins at a time (1,2,3,4,5). However, there are some occasions in immunology when experiments do not easily lend themselves to parallel processing. For example, it is often difficult to isolate an interesting set of proteins for further study when the set of input proteins is unknown. Many effective statistical approaches have been proposed for commercial protein array feature detection, but custom experiments remain in which there are too many unknown quantities for a conventional feature detection analysis.

In our protein expression studies, we were faced with analyzing a series of custom, hand-spotted arrays. One of our early objectives was to estimate the material concentration at which features were detectable. Our design was a titration design. Our approach for detecting features was to find the concentration where foreground signal was indistinguishable from background noise. We present our results using a two-dimensional (2-D) median filter with varying neighborhoods to separate features from noise, and we compare these results with results from alternative methods of signal detection.

Materials and Methods


Influenza virus vaccine was purchased from Sanofi-Aventis (Bridgewater, NJ, USA). Immune Globulin Gammagard® (Baxter International, Deerfield, IL, USA) was used as a source of immunoglobulin G (IgG). Nitrocellulose membrane with a 0.2 µm pore size was purchased from Schleicher & Schuell (Keene, NH, USA). A multiblot replicator, a multiprint, and a library copier (V&P Scientific, San Diego, CA, USA) were used to spot the protein arrays. We will refer to these as macroarrays, since the hand-stamped spots were 2.25 mm apart. The replicator contains 96 pin tools designed to hold 0.1 µL fluid by the hanging drop method. A high-density multiprint holds the nitrocellulose membrane in place and contains 16 alignment holes. The alignment holes allow 16 microtiter plates to be spotted onto one membrane. Cy™3-donkey anti-human antibody and horseradish peroxidase (HRP)-donkey anti-human antibody were obtained from Jackson ImmunoResearch (West Grove, PA, USA). ECL Plus™ with HRP-labeled secondary antibody was obtained from Amersham Biosciences (Piscataway, NJ, USA).

Protein Macroarray Construction

Different concentrations of flu vaccine and human IgG were spotted onto arrays to determine the limits of detection of two visualization techniques: Cy3 directly coupled secondary antibody versus ECL Plus with an HRP-coupled secondary antibody. The two arrays consisted of 16 identical spots of different concentrations of flu vaccine and human IgG. For this study, the 16 replicates will be referred to as a well. Different concentrations of flu vaccine and IgG were used to create a mother plate. Flu concentrations ranged sequentially from 2.5 to 0.000005 ng; IgG concentrations ranged sequentially from 25 to 0.000005 ng. A figure showing the macroarray layout is included in the Supplementary Material (see Supplementary Figure S1 available online at Thirty-six wells contained buffer only as a control set. Samples were diluted in 100 µL start buffer then 100 µL Ponceau S in water. Start buffer consisted of 6 M urea, 0.2% CHAPS, 100 mM ammonium bicarbonate, pH 8.0. The pink-colored Ponceau S was used as a visual check to ensure that each pin tool delivered sample to the membrane. The replicator was dipped into the mother plate and 0.1 µL sample was transferred onto a precut nitrocellulose membrane. Macroarrays used in this study were spotted on the same day. After spotting, the arrays were dried under a hood. Manual stamping created an imprint onto the nitrocellulose membrane, which indicated the spot position. These imprints were later used after scanning to assist with gridding.

Macroarray Processing

To prevent cross-contamination, each array was processed in a separate Petri dish. The macroarrays were blocked in 1% bovine serum albumin (BSA) in Tris-buffered saline (TBS) overnight at 4°C. After blocking, the arrays were washed twice in TBS and probed with a 1:200 dilution of human sera for 1 h at room temperature. The macroarrays were then washed two more times in TBS (20 mM Tris-HCl, pH 7.5, 500 mM NaCl, pH 7.5).

Protein arrays were visualized using either Cy3-labeled secondary antibody or ECL Plus with HRP-labeled secondary antibody. For Cy3 comparison: a 1:800 dilution of Cy3-donkey anti-human was used for detection. For ECL Plus, a 1:100,000 dilution of HRP-donkey anti-human antibody was used for detection, which was developed using 4 mL of the ECL Plus solution. Blots were imaged on a Typhoon™ 9410 variable mode imager (Amersham Biosciences) using 50 µm resolution. The Cy3 blots were imaged using a 532 nm excitation (green), 580BP30 filter, at both PMT 400 and 500 V, respectively. The ECL Plus™ blots were imaged with a 457 nm excitation (blue), 520BP40 filter, PMT at both 450 and 500 V, respectively. This gave us four images from two arrays, one Cy3 array imaged at 400 and 500 V, and one ECL Plus array imaged at 450 and 500 V.

  1    2    3