arrayQualityMetrics report for obj

	array	sampleNames	*1	*2	*3
	1	GSM2800849_SG14124393_257236310089_S002_GE1_1105_Oct12_1_1.txt
	2	GSM2800850_SG14124393_257236310089_S002_GE1_1105_Oct12_1_2.txt	x
	3	GSM2800851_SG14124393_257236310089_S002_GE1_1105_Oct12_1_3.txt
	4	GSM2800852_SG14124393_257236310089_S002_GE1_1105_Oct12_1_4.txt
	5	GSM2800853_SG14124393_257236310089_S002_GE1_1105_Oct12_2_1.txt
	6	GSM2800854_SG14124393_257236310089_S002_GE1_1105_Oct12_2_2.txt	x
	7	GSM2800855_SG14124393_257236310089_S002_GE1_1105_Oct12_2_3.txt
	8	GSM2800856_SG14124393_257236310089_S002_GE1_1105_Oct12_2_4.txt
	9	GSM2800873_SG14124393_257236310095_S001_GE1_1105_Oct12_1_1.txt
	10	GSM2800874_SG14124393_257236310095_S001_GE1_1105_Oct12_1_2.txt
	11	GSM2800875_SG14124393_257236310095_S001_GE1_1105_Oct12_1_3.txt
	12	GSM2800876_SG14124393_257236310095_S001_GE1_1105_Oct12_1_4.txt
	13	GSM2800877_SG14124393_257236310095_S001_GE1_1105_Oct12_2_1.txt
	14	GSM2800878_SG14124393_257236310095_S001_GE1_1105_Oct12_2_2.txt
	15	GSM2800879_SG14124393_257236310095_S001_GE1_1105_Oct12_2_3.txt	x
	16	GSM2800880_SG14124393_257236310095_S001_GE1_1105_Oct12_2_4.txt	x	x	x
	17	GSM2800913_SG14124393_257236310133_S001_GE1_1105_Oct12_1_1.txt
	18	GSM2800914_SG14124393_257236310133_S001_GE1_1105_Oct12_1_2.txt
	19	GSM2800915_SG14124393_257236310133_S001_GE1_1105_Oct12_1_3.txt
	20	GSM2800916_SG14124393_257236310133_S001_GE1_1105_Oct12_1_4.txt
	21	GSM2800917_SG14124393_257236310133_S001_GE1_1105_Oct12_2_1.txt
	22	GSM2800918_SG14124393_257236310133_S001_GE1_1105_Oct12_2_2.txt
	23	GSM2800919_SG14124393_257236310133_S001_GE1_1105_Oct12_2_3.txt
	24	GSM2800920_SG14124393_257236310133_S001_GE1_1105_Oct12_2_4.txt

The columns named *1, *2, ... indicate the calls from the different outlier detection methods:

1. outlier detection by Distances between arrays
2. outlier detection by Boxplots
3. outlier detection by MA plots

The outlier detection criteria are explained below in the respective sections. Arrays that were called outliers by at least one criterion are marked by checkbox selection in this table, and are indicated by highlighted lines or points in some of the plots below. By clicking the checkboxes in the table, or on the corresponding points/lines in the plots, you can modify the selection. To reset the selection, reload the HTML page in your browser.

At the scope covered by this software, outlier detection is a poorly defined question, and there is no 'right' or 'wrong' answer. These are hints which are intended to be followed up manually. If you want to automate outlier detection, you need to limit the scope to a particular platform and experimental design, and then choose and calibrate the metrics used.

Figure 1 (PDF file) shows a false color heatmap of the distances between arrays. The color scale is chosen to cover the range of distances encountered in the dataset. Patterns in this plot can indicate clustering of the arrays either because of intended biological or unintended experimental factors (batch effects). The distance d_ab between two arrays a and b is computed as the mean absolute difference (L₁-distance) between the data of the arrays (using the data from all probes without filtering). In formula, d_ab = mean | M_ai - M_bi |, where M_ai is the value of the i-th probe on the a-th array. Outlier detection was performed by looking for arrays for which the sum of the distances to all other arrays, S_a = Σ_b d_ab was exceptionally large. 4 such arrays were detected, and they are marked by an asterisk, *.

Figure 2 (PDF file) shows a bar chart of the sum of distances to other arrays S_a, the outlier detection criterion from the previous figure. The bars are shown in the original order of the arrays. Based on the distribution of the values across all arrays, a threshold of 6610 was determined, which is indicated by the vertical line. 4 arrays exceeded the threshold and were considered outliers.

array sampleNames

Figure 3 (PDF file) shows a scatterplot of the arrays along the first two principal components. You can use this plot to explore if the arrays cluster, and whether this is according to an intended experimental factor (you can indicate such a factor by color using the 'intgroup' argument), or according to unintended causes such as batch effects. Move the mouse over the points to see the sample names.
Principal component analysis is a dimension reduction and visualisation technique that is here used to project the multivariate data vector of each array into a two-dimensional plot, such that the spatial arrangement of the points in the plot reflects the overall data (dis)similarity between the arrays.

Figure 4 (PDF file) shows boxplots representing summaries of the signal intensity distributions of the arrays. Each box corresponds to one array. Typically, one expects the boxes to have similar positions and widths. If the distribution of an array is very different from the others, this may indicate an experimental problem. Outlier detection was performed by computing the Kolmogorov-Smirnov statistic K_a between each array's distribution and the distribution of the pooled data.

Figure 5 (PDF file) shows a bar chart of the Kolmogorov-Smirnov statistic K_a, the outlier detection criterion from the previous figure. The bars are shown in the original order of the arrays. Based on the distribution of the values across all arrays, a threshold of 0.0421 was determined, which is indicated by the vertical line. One array exceeded the threshold and was considered an outlier.

array sampleNames

Figure 6 (PDF file) shows density estimates (smoothed histograms) of the data. Typically, the distributions of the arrays should have similar shapes and ranges. Arrays whose distributions are very different from the others should be considered for possible problems. Various features of the distributions can be indicative of quality related phenomena. For instance, high levels of background will shift an array's distribution to the right. Lack of signal diminishes its right right tail. A bulge at the upper end of the intensity range often indicates signal saturation.

Figure 7 (PDF file) shows a density plot of the standard deviation of the intensities across arrays on the y-axis versus the rank of their mean on the x-axis. The red dots, connected by lines, show the running median of the standard deviation. After normalisation and transformation to a logarithm(-like) scale, one typically expects the red line to be approximately horizontal, that is, show no substantial trend. In some cases, a hump on the right hand of the x-axis can be observed and is symptomatic of a saturation of the intensities.

Figure 8 (PDF file) shows MA plots. M and A are defined as:
M = log₂(I₁) - log₂(I₂)
A = 1/2 (log₂(I₁)+log₂(I₂)),
where I₁ is the intensity of the array studied,and I₂ is the intensity of a "pseudo"-array that consists of the median across arrays. Typically, we expect the mass of the distribution in an MA plot to be concentrated along the M = 0 axis, and there should be no trend in M as a function of A. If there is a trend in the lower range of A, this often indicates that the arrays have different background intensities; this may be addressed by background correction. A trend in the upper range of A can indicate saturation of the measurements; in mild cases, this may be addressed by non-linear normalisation (e.g. quantile normalisation).
Outlier detection was performed by computing Hoeffding's statistic D_a on the joint distribution of A and M for each array. Shown are first the 4 arrays with the highest values of D_a, then the 4 arrays with the lowest values. The value of D_a is shown in the panel headings. One array had D_a>0.15 and was marked as outlier. For more information on Hoeffing's D-statistic, please see the manual page of the function hoeffd in the Hmisc package.

Figure 9 (PDF file) shows a bar chart of the D_a, the outlier detection criterion from the previous figure. The bars are shown in the original order of the arrays. A threshold of 0.15 was used, which is indicated by the vertical line. One array exceeded the threshold and was considered an outlier.