Microarray Analysis

Microarray Analysis

Microarray analysis is method to measure changes in gene expression (actually RNA content) over a large number of genes at the same time. DNA representing each gene (called a "probe") is placed on very small spots on a solid support. The mRNA (called a "target") is converted to fluorescently labeled cDNA and hybridized to the array. Analysis then reveals a number of genes that are differentially expressed at a higher or lower level in two different targets. Although the target is typically mRNA, microarray experiments are also used to quantify microRNAs, clarify splicing patterns, or directly on genomic DNA to measure amplification of genes in cancer cells. The underlying motive when measuring gene expression is that more mRNA usually means that more protein is made. The mechanism could either be a change in the rate of transcription, or a change in the rate of mRNA turnover. One always has to keep in mind that translational control, alteration of protein turnover, or regulation by protein modification may also be occurring, and will not be reflected by the amount of mRNA present.

The history of microarray analysis is summarized in:

• Fodor SP et al, Science 1991 251: 767-73
• Schena M et al., Science 270: 467-70.

Producing the targets

As with all RNA work, it is essential to avoid contamination with RNAse when preparing the mRNA. This requires RNAse inhibitors, specially cleaned reagents, fastidious technique, and an assay for RNA degradation. The classical assay for RNA degradation is a Northern blot. More recently, HPLC applications have been used.

This image from Agilent's Bioanalyzer web advertisement emphasizes the relative intensity of the 18S and 26S rRNA bands as an indicator of degradation. Note, however, that rRNA is much more resistant to RNAse than is mRNA. There would be no mRNA at all surviving in the sample to the right.

Typically, polyadenylated RNA is subjected to reverse transcription primed with oligo dT. The label can be produced as dye conjugated dUTP, although each label has a different efficiency of incorporation. It is also possible to incorporate amino allyl-dUTP, and conjugate the dye to the amino group after the fact of making the cDNA.

cDNA can be accumulated by linear amplification. For prokaryotic mRNA, random hexamer primers are used to prime conversion to cDNA. Cycling in this instance risks generating great variation in the representation of different mRNA species.

Because the method requires reproducible kinetics of hybridization, achieving a comparable concentration of cDNA in each of the samples is critical for producing usable results.

Probes

The probes are tightly arrayed on a chip. The different variations can be divided into short oligos (typically 25 nt), long oligos (typically 60 nt), and full cDNAs.

Different Microarray Instruments.

(Miller and Tang, [2009] Clin. Microbiol Rev. 22:611-633).

In situ synthesis

The best known microarray system is made by Affymetrix. ( The oligonucleotide probes are synthesized on the solid support by a photolithographic technique. This makes use of a blocking group on each added nucleotide that is removed by a photo-induced reaction allowing the next nucleotide to add. A "mask" is applied so that only the spots scheduled for the next base to be added at this position (e.g. a T below) are illuminated. Then the chip is washed with the reagent to add a T (a nucleotidyl phosphoramidite). The mask is then shifted to deprotect the spots to get another base (e.g. a C) and then that reagent is added. This is repeated 4 times at each position until an array of 25 nt long oligonucleotides is built up on the chip. These are typically called "short oligo" arrays.

Affymetrix tries to put ~11 oligos per gene all from the 3' UTR on each chip, and pairs them with an oligo that has a mismatched base in the center. The idea is that the signal from the mismatched probe can be used to subtract non-specific hybridization from the signal from the perfectly matched probe. This is called PM-MM scoring. However, it has been shown that correcting by the mismatched signal actually produces nosier data than ignoring the mismatched signal (Milenaar et al, BMC Bioinformatics 7:137 [2006]). So most users ignore the MM signal. The name of the software to reanalyze the data without doing PM-MM subtraction is RMA. The average fluorescence intensity after excluding outliers (probes whose signals do not change in the same pattern as the others) is called the "expression value".

Short oligo arrays have less sensitivity than long oligo arrays. For low concentration targets (e.g. microbiological samples) a preliminary PCR step may be required to amplify the target. Discrimination of a single mismatch becomes better at shorter lengths. The signal from a single mismatch is ~ 25% at a probe length of 19 nt. The lithographic method can produce very high density arrays with a million probes per chip.

Another manufacturer of high density in situ synthesized microarrays is Roche-NimbleGen. It gets around the Affymetrix patent on photolithographic synthesis by using micro-mirrors to focus light on the spots to be extended rather than a lithographic mask. Agilent uses high density in situ synthesized microarrays where the reagents to add each successive base are supplied in a focused way by an inkjet printing process. NimbleGen and Agilent equipment supports two color analysis, whereas Affymetrix equipment supports only one color analysis.

Spotted (printed) arrays

Before microarrays, people made hybridization arrays by simply spotting DNA on a glass slide. This strategy has been scaled down through the use of robotic spotting machines that dip an array of needles into a microplate with solutions of different probes and then touch them to a glass slide or other solid support. Spotted arrays can be made directly from denatured cDNAs (or more commonly PCR amplicons from cDNAs), in which case the DNA sticks to the glass support by electrostatic interaction. However, in order to distinguish between closely related members of gene families, they are usually made from synthetic oligonucleotides which are typically 60 nt long and correspond to 3' untranslated regions. Oligos are bound to the support through a modified 5' or 3' end. To distinguish from the lithographically synthesized arrays, printed oligo arrays are often called "long oligo" arrays, although they can be made with oligos of any length. Printed arrays usually have a density of only 10,000 - 30,000 spots, and have less redundancy per gene.

The most well known vender of printed microarray equipment is Agilent. UTHSCSA has Agilent equipment used under the supervision of Dr. Yidong Chen at CCRI. Core facilities can provide substantial savings by buying a set of oligos, printing the arrays locally, and distributing the cost over multiple users.

Most spotted array equipment allows two dye measurements. In two dye (or two channel) measurements, two different targets are labeled with dyes of different colors. The most commonly used dyes are named cy3 (green) and cy5 (red). Both targets are hybridized at once to the chip, and the colors are analyzed separately. This has the advantage of automatically normalizing for variation in the amount of probe from spot to spot on the array. For two color analysis the raw readout is M = log2 Red/Green = log2R - log2G. A plot of M for each gene against 1/2(log2R + log2G) is called an MA plot.

From Wikipedia:

The MA plots will typically be subjected to low level normalization based on the assumption that the average gene is not differentially expressed.

Bead Arrays

Illumina mounts their oligos on 3 micron beads, that fit into wells in a microplate such that fiber optic sensors attach to each bead. Their two platforms are called Sentrix Array Matrixes (SAM) or Sentrix BeadChips. The beads have to go through a process called "decoding" to figure out what oligo sequence became attached to each sensor. Basically, when the probe oligos are initially synthesized they are concatenated to a 29 nt sequence designed to be easily identifiable by a series of hybridizations to short oligonucleotides. Hybridizations to these oligos is carried out first, identifying the address attached to each optical fiber, and hence the associated probe sequence. The Illumina technology is related to their next generation sequencing technology, and they sell a platform that will do both kinds of analysis. Bead Arrays are configured for fewer probes and many more target samples.

Sources of noise in microarray experiments.

Technical noise.

Probe efficiency variation: There can be variation in the amount of probe per spot, or the efficiency of hybridization due to formation of hairpins within probes, or due to the Tm being out of range. The first step of analysis is typically an imaging of the hybridized spots and the exclusion of spots that are misshapen or compromised by dust or other inappropriate fluorescent signals in the image.

Image of a section of a microarray from Howlader & Chaubey, IEEE Trans Image Process 19:1953-1967 (2010).

Nonlinear hybridization kinetics:

To understand the capabilities and limitations of a microarray experiment, a comparison to its forerunner, the Northern Blot, is given below.

Northern Blot

In a Northern Blot, the RNA from a cellular preparation is run on a denaturing agarose gel and then the pattern is transferred to a filter to which the RNA becomes permanently affixed. For a given gene, a cDNA is prepared and either radiolabeled or fluorescently labeled. The cDNA is called the "probe" in this case. The probe is hybridized to the filter over a long period of time until all complementary RNA on the filter is duplexed. After a wash to eliminate probe that is not duplexed, the filter is imaged. Since the hybridization was to completion, the amount of signal is proportional to the amount of the cognate RNA on the filter.

Image from Wikipedia Commons.

The strengths of the Northern Blot are that quantification is accurate, degradation or lack of it of the RNA is apparent, and it has a large dynamic range. Its weakness is that for each gene, one has to prepare a separate labeled probe, strip the filter, and then hybridize again. Hence Northern blots are not applicable to analyzing any appreciable number of different genes. For analyzing large numbers of genes (up to 23,000) a microarray experiment is used. For analyzing smaller numbers (up to 100) qPCR is now the preferred method.

Kinetics of microarray hybridization

In microarray analysis, a probe for each gene is placed on a spot on a supporting material, such that thousands of genes may be represented on a small area (called a "chip"). Fluorescently labeled cDNA corresponding to total RNA is prepared and hybridized to the chip. The RNA is called the "target" or sometimes the "treatment". Typically one will be comparing the amount of RNA present under two different circumstances (e.g., with and without application of a drug to cells in culture), so the usual outcome is to identify genes which are more or less heavily represented in the RNA of cells for treatment A vs. treatment B. In this case, if the target cDNA were hybridized to completion, then all quantitative information would be lost. The signal intensity would represent the amount of probe on the chip, not the amount of RNA in the target. Instead of hybridizing to completion, the target is hybridized for a fixed amount of time so that each spot is partially duplexed. Since the on rate for hybridization is related to the concentration of the hybridizing species, the signal produced in a fixed hybridization time reflects the concentration of the complementary cDNA in the target.

The accumulation of target cDNA on the probe is linear up to a point and then the spot become saturated. cDNAs present at low levels may not exceed the amount of probe. Those (red line in the fig. above) will begin to saturate at lower levels because the off rate becomes significant as the amount of free cDNA approaches zero. The presence of non-specific opportunities for low level cDNAs to interact with the chip may differ from spot to spot on the chip and produce spot to spot variation in the signal due to localized exhaustion of cDNA. The end result is that the signal tends to have poor linearity, expression changes may be attenuated by saturation, and random variation can be introduced into the results. Significant expression differences can be lost in the random background from nonspecific interaction with the chip. This problem is discussed by Chudin et al., (2002) Genome Biol. 3(1), RESEARCH0005, and Ono et al., Bioinformatics. (2008) 24:1278-85.

Noise with bioinformatics sources.

Probes are usually placed in the 3' untranslated regions of genes for two reasons: 1) Many genes fall in families with sufficiently closely related paralogs that hybridization to a coding region oligonucleotide would not distinguish between the family members. 2) The cDNA is usually primed with oligo dT, and the amount of cDNA produced falls off with distance from the polyA tail. Hence it is necessary to accurately predict the polyadenylation sites of each gene. Two computer programs are in use for this purpose: 1) polyadq, and 2) Genescan. Neither is 100% reliable, and sometimes gene expression is missed because the probe was placed downstream of the polyA addition site. Different placement of the probe from one chip design to the next, coupled to different potentials for cross hybridization to related sequences and the generally nonlinear response of the method have historically led to extensive discrepancies in expression profiles from one experiment to another. These properties are exacerbated when working from a genomic sequence that is still in early stages of finishing.

Noise due to specifics of different oligos.

Variation in signal intensity among different oligos for the same gene. (modified from Li & Wong. (2001) PNAS 98, 31-36. Each point is a different oligo in the same gene. The variation in oligo response is generally greater than the differences from one treatment to another. An oligo's response can be low because of hairpin formation or because its Tm is out of range. Oligo response can be high due to cross hybridization to other species in the cDNA.

Many of the technical sources of noise should not affect a relative change in expression level, and two color measurements should help

In principle, replicate microarray hybridizations (for both treatments) could be used to drive down the noise. However, microarray experiments typically cost $1000 - $3000 per chip, so most experimenters do one chip per treatment, accept that the results are noisy, and then proceed to use qPCR to filter out some genes with actual expression changes. Some microarray setups allow measuring two colors. If treatment A is labeled in one color and treatment B in another, then the ratio of A/B is measured over each spot should be free of variation based on the amount of probe washing over the spot.

Another problem is that the total amount of RNA for treatment A and B may not have been the same, or the cDNA synthesis may not have occurred comparably. Typically there will be a normalization intended to enforce the assumption that most RNAs did not change intensity between the two treatments.

Noise from Biological Variation

Deciding that there is confidence in a finding of differential gene expression requires a statistical analysis that apportions the variation observed between differences in gene expression and sources of random variation. Consider for example that RNA preparations were made from the livers of several mice, some receiving no treatment (A)and some treated with a drug (B). Consider the following experiment that would require 8 mice, four chips, a two color instrument, and 8 labeling reactions:

Here, if we had only done one chip with one treated and one untreated mouse (replicate 1), we might have concluded that gene 1 expression was suppressed by the drug while gene 2 expression was unaffected. Upon repeating four replicates, we would realize that expression of gene 1 was subject to extensive biological variation, whereas gene two is relatively steadily expressed. The statistics to determine if the average change between A and B for each mouse is significant given the variation from replicate to replicate is beyond the scope of this course. But clearly without enough replicates there will be genes identified as being affected by the drug treatment that are, in fact, not.

At first we may wonder if the "biological" variation was really variation of expression among the livers of these different mice, or technical variation in how much total mRNA was recovered from each mouse, or how efficiently each total mRNA sample was converted to probe. However, if there are a large number of genes like B whose expression appears very steady across replicates, that would suggest that our procedure for labeled target production was consistent. More often, the RNA preparations won't be as consistent as we'd like, but it would be possible to notice that some large majority of gene varied consistently in expression from replicate to replicate, and then we would normalize all the signals based on the assumption most genes are expressed the same from mouse to mouse. Similarly, the steadily expressed set of genes may show a consistent difference between treatment A and B. Commonly a normalization is imposed to enforce the assumption that the expression of most genes is not different between treatment A and B.

Alternatively, it can be done to devote two chips to the same A and B samples, only with the probe labeling (red vs. green) reversed.

The proper statistics for dealing with data of this type is beyond the scope of this (or any introductory statistics course), but exists in a number of varieties that go by names like "linear model fitting", and "empirical Bayesian analysis". The mainline software package for finding confidence levels that expression has changed is limma, which is freeware that runs within the R programming language and comes as part of the bioconductor package (