Reliability and reproducibility issues in DNA microarray measurements

Abstract

DNA microarrays enable researchers to monitor the expression of thousands of genes simultaneously. However, the current technology has several limitations. Here we discuss problems related to the sensitivity, accuracy, specificity and reproducibility of microarray results. The existing data suggest that for relatively abundant transcripts the existence and direction (but not the magnitude) of expression changes can be reliably detected. However, accurate measurements of absolute expression levels and the reliable detection of low abundance genes are difficult to achieve. The main problems seem to be the sub-optimal design or choice of probes and some incorrect probe annotations. Well-designed data-analysis approaches can rectify some of these problems.

Figure 1

Different probes meant to represent the same transcript can yield widely different signals. The left panels show the perfect match (PM, in blue) and mismatch (MM, in green) values across the probes in the same probe set. (a) For the gene shown, probe 10 (from left to right) is close to saturation level, whereas probes 5, 9, 13, 15 and 16 are close to the background level. Programs such as MAS 5.0 call this gene present (P) and calculate the expression level based on the average difference between PM and MM probes. This illustrates how sensitive the actual numbers are to the choice of the specific probes. (b) Most probes corresponding to this gene are expressed at the background level; therefore, this gene is absent (A). However, probes 2 and 3 produce high levels of signal intensity. In some cases, the PM–MM difference for some of the probes of an absent gene can be so large that the average difference, often used to represent the expression level of the gene, can be higher for some absent genes than some of those declared present (reproduced with permission from Ref. [23]).

Figure 2

Probe intensity ratios are more consistent (have a smaller variance) than absolute probe intensities. Probe-level intensities from two Affymetrix HG-U133A arrays were log₂ transformed and the standard deviation was calculated for the 11 probes in each probe set. The smoothed distribution of these standard deviations are plotted for brain (blue), pancreas (green) and the probe-level ratios of brain to pancreas (red). Most variances of the ratios are smaller than the variances of either tissue. Data are from the GNF expression atlas [66].