Clustering analysis is commonly used for interpreting microarray data. It provides both a visual representation of complex data and a method for measuring similarity between experiments (gene ratios). The widely used methods for clustering microarray data are: Hierarchical, K-means and Self-organizing map. In this article, the second in our series on Ambion's MessageAmp™ aRNA Amplification Kit, we present data and statistical analyses from experiments conducted by Drs. Philip Moos and Brian Dalley at the University of Utah, Huntsman Cancer Institute (HCI). The microarrays used in this study were manufactured at the HCI and contained 6912 cDNA clones deposited in duplicate using a Molecular Dynamics GEN III Array Spotter. Moos and Dalley compared the data generated from the HCI microarrays hybridized with mRNA and amplified antisense RNA (aRNA) generated with the MessageAmp Kit.
Experimental Design
Total RNA was isolated from HCT116 and RKO cells and aRNA was amplified from total RNA (2 µg). For non-amplified samples, mRNA was purified using Oligotex beads (Qiagen). To assess the reproducibility of RNA amplification and compare representation of messages in aRNA with those in mRNA, three samples each of RKO and HCT116 RNA were amplified independently. The quality and yields of RNA obtained from the various samples are presented in Figure 2. mRNA or aRNA (2 µg) samples were fluorescently labeled by incorporating Cy3-dCTP (RKO samples) or Cy5-dCTP (HCT 116 samples) during reverse transcription with random 9-mers. Each glass slide contained duplicate arrays and each labeled RNA sample was hybridized to two slides (4 replicates). Duplicate hybridizations were performed with each sample RNA and the quantified data was represented as slide averages resulting in 12 arrays for each aRNA and mRNA sample. Images were acquired using a Molecular Dynamics Array Scanner. Microarray data analysis was performed using Spotfire Software (Spotfire®, Somerville, MA).
Cluster Analysis
An example of the image data obtained from 4 of the 12 grids from a standard 6912 element HCI array is represented in Figure 1. Comparison of the signals obtained using mRNA vs. aRNA indicates that RNA amplification provides excellent signal to noise, even for genes with a low level of expression.
Hierarchical clustering analysis of the data obtained from 6912 elements was carried out using UPGMA (Unweighted Pair Group Method with Arithmetic Mean) analysis (see sidebar "Clustering Methods Used for Analyzing Microarray Data"), with an ordering function based on the input rank. This data is represented as a dendrogram (tree graph) with the closest branches of the tree representing arrays with similar gene expression patterns. Figure 3 depicts the hierarchical clustering data from all 6912 elements. The results indicate that there are broad similarities between arrays hybridized with aRNA or mRNA. Even though the overall signal patterns found on the aRNA and mRNA hybridized arrays are similar, a small subset of regions show differential expression (RKO/HCT116) signals between the aRNA and mRNA samples.
Figure 3. Hierarchical Clustering Analysis of All Array Elements. Hierarchical clustering data of all the elements in a HCl array. A dendrogram (tree graph) epicts the grouping of the genes based on the similarity between them. UPGMA analysis (unweighted average) was carried out using the "Euclidean Distance" to determine the similarity measure and the input rank as the ordering function. A subset of all the columns constituting the complete data is shown in this figure.
To obtain statistically significant data for the sub-regions that were distinct between the aRNA and mRNA (91 elements), a weighted average (WPGMA) analysis was carried out. The hierarchical clustering of these 91 elements is depicted in Figure 4. It is evident that there are very few genes that clearly segregate into either mRNA or aRNA groups. It is important to note, for those genes that do segregate, the gene expression differences (ratios) do not change direction (i.e. RKO>HCT to HCT>RKO), but show greater differences in the aRNA samples compared to the mRNA samples (as determined by the color shade).
Figure 4. Hierarchical Clustering Analysis of Selected Array Elements. Hierarchical clustering analysis of a few select genes (91 of 6912) that are very different between the aRNA and mRNA samples. This analysis was carried out using the weighted average (WPGMA) method.
An alternate methodology used to understand the clustering of microarray data is k-means clustering. This method does not suffer from some of the problems associated with hierarchical clustering such as irrelevance of gene expression data as clustering progresses or spurious results due to errors in assigning clusters initially in the analysis (2). K-means clustering of all the elements of the HCI arrays with 6 clusters was determined (Figure 5). After 45 iterations, a total score of 1.082e+004 was computed. The most similar "similarity value" was 0 and the least similar "similarity value" was 1.798e+308. This grouping of genes to identify sets of genes that appear to be differentially expressed between aRNA and mRNA resulted in two clusters (91 elements among clusters 5 and 6) that have the largest difference between aRNA and mRNA (Figure 6).
Figure 5. K- means cluster analysis of the 6912 elements using a userdefined cluster number of 6. 45 iterations were carried out to group the genes within a given cluster using a data centroid based search. The total score was calculated to be 1.082e+004. The most closely clustering genes had a similarity value of 0 and the least similar gene had a similarity value of 1.798e+308.
Figure 6. An analysis of variance calculation of K-mean clusters 5 and 6. This plot indicates the confidence limit of the data. A Œp-value‚ of less than 0.0001 was used, indicating that the genes represented in this plot are unique at the 99.9999% cut-off value.
Analysis of variance (ANOVA) of genes in clusters 5 and 6 indicated that the clusters contained genes that behave distinctly between the mRNA and aRNA samples at a confidence limit of 99.99999% (p<0.00001). ANOVA measurements processed the gene-by-gene fluctuations from the mean value and accounted for variance across samples.
A scatter plot analysis of the raw Cy3 and Cy5 values of all the 6912 elements within the 5 gene clusters is shown in Figure 7 (4 plots). The top two plots represent all the elements and the bottom two depict the genes that show the largest differences in signal. Most of the genes that are distinct between the samples are expressed at lower levels (low fluorescent signal). These differences were more exaggerated in the aRNA than the mRNA sample because the signal-to-noise ratio was typically much greater in the aRNA sample. The distinct genes in the aRNA panel might be elements that were not clearly discernible in the mRNA sample due to ribosomal contamination (27% in the mRNA used for this analysis). The presence of ribosomal RNA can increase background in mRNA samples, resulting in variations in mRNA concentration between samples and decreasing the efficiency of cDNA probe synthesis. Thus the presence of ribosomal RNA could have cumulatively skewed the detection and quantification of genes that were expressed in very low amounts when mRNA or total RNA was used.
Figure 7. Scatter Plot Analysis of all Array Elements. A scatter plot of the Cy™5 vs. Cy™3 values obtained for an aRNA and a mRNA array is shown. The top two panels depict the 6 clusters (obtained after K-Mean Clustering Analysis) containing all 6912 elements. A subset of elements that are distinct between the two arrays and which deviate the most in signal intensity are depicted in the lower panels.
Amplification of RNA thus provides a means of measuring expression from genes transcribed at very low levels. In many cases the RNA concentration of an experimental sample is under the optimal required amount for synthesizing labeled cDNA for microarray analysis. MessageAmp is a viable technology for increasing the yield of useful probe and can greatly lower the starting amount of RNA required to produce biologically relevant signals.
Figure 1. Array Analysis of 6912 Element Arrays. Image Data from 4 of the 12 grids of a standard 6912 element Hunstman Cancer Institute cDNA microarray. mRNA or aRNA (2 µg) was reverse transcribed in the presence of random ninemers and Cy™ dyes. HCT 116 samples were labeled with Cy5-dCTP while RKO cells were labeled with Cy3-dCTP. Analysis of the microarray data was carried out using Spotfire Software (Spotfire®, Somerville, MA).
| mRNA Sample | A260/280 | aRNA Yield |
|---|---|---|
| RKO-1 | 2.12 | 37.0 µg |
| RKO-2 | 2.15 | 28.7 µg |
| RKO-2 | 2.12 | 26.8 µg |
| HCT116-1 | 2.12 | 33.5 µg |
| HCT116-2 | 2.12 | 36.2 µg |
| HCT116-3 | 2.12 | 29.5 µg |
Figure 2. Yields of Amplified RNA. Yields obtained after amplification of total RNA (2 µg) from HCT116 and RKO cells using the MessageAmp™ Kit.
Hierarchical clustering analysis of the data obtained from 6912 elements was carried out using UPGMA (Unweighted Pair Group Method with Arithmetic Mean) analysis (see sidebar "Clustering Methods Used for Analyzing Microarray Data"), with an ordering function based on the input rank. This data is represented as a dendrogram (tree graph) with the closest branches of the tree representing arrays with similar gene expression patterns. Figure 3 depicts the hierarchical clustering data from all 6912 elements. The results indicate that there are broad similarities between arrays hybridized with aRNA or mRNA. Even though the overall signal patterns found on the aRNA and mRNA hybridized arrays are similar, a small subset of regions show differential expression (RKO/HCT116) signals between the aRNA and mRNA samples.
To obtain statistically significant data for the sub-regions that were distinct between the aRNA and mRNA (91 elements), a weighted average (WPGMA) analysis was carried out. The hierarchical clustering of these 91 elements is depicted in Figure 4. It is evident that there are very few genes that clearly segregate into either mRNA or aRNA groups. It is important to note, for those genes that do segregate, the gene expression differences (ratios) do not change direction (i.e. RKO>HCT to HCT>RKO), but show greater differences in the aRNA samples compared to the mRNA samples (as determined by the color shade).
An alternate methodology used to understand the clustering of microarray data is k-means clustering. This method does not suffer from some of the problems associated with hierarchical clustering such as irrelevance of gene expression data as clustering progresses or spurious results due to errors in assigning clusters initially in the analysis (2). K-means clustering of all the elements of the HCI arrays with 6 clusters was determined (Figure 5). After 45 iterations, a total score of 1.082e+004 was computed. The most similar "similarity value" was 0 and the least similar "similarity value" was 1.798e+308. This grouping of genes to identify sets of genes that appear to be differentially expressed between aRNA and mRNA resulted in two clusters (91 elements among clusters 5 and 6) that have the largest difference between aRNA and mRNA (Figure 6).
Figure 5. K- means cluster analysis of the 6912 elements using a userdefined cluster number of 6. 45 iterations were carried out to group the genes within a given cluster using a data centroid based search. The total score was calculated to be 1.082e+004. The most closely clustering genes had a similarity value of 0 and the least similar gene had a similarity value of 1.798e+308.
Figure 6. An analysis of variance calculation of K-mean clusters 5 and 6. This plot indicates the confidence limit of the data. A Œp-value‚ of less than 0.0001 was used, indicating that the genes represented in this plot are unique at the 99.9999% cut-off value.
Analysis of variance (ANOVA) of genes in clusters 5 and 6 indicated that the clusters contained genes that behave distinctly between the mRNA and aRNA samples at a confidence limit of 99.99999% (p<0.00001). ANOVA measurements processed the gene-by-gene fluctuations from the mean value and accounted for variance across samples.
A scatter plot analysis of the raw Cy3 and Cy5 values of all the 6912 elements within the 5 gene clusters is shown in Figure 7 (4 plots). The top two plots represent all the elements and the bottom two depict the genes that show the largest differences in signal. Most of the genes that are distinct between the samples are expressed at lower levels (low fluorescent signal). These differences were more exaggerated in the aRNA than the mRNA sample because the signal-to-noise ratio was typically much greater in the aRNA sample. The distinct genes in the aRNA panel might be elements that were not clearly discernible in the mRNA sample due to ribosomal contamination (27% in the mRNA used for this analysis). The presence of ribosomal RNA can increase background in mRNA samples, resulting in variations in mRNA concentration between samples and decreasing the efficiency of cDNA probe synthesis. Thus the presence of ribosomal RNA could have cumulatively skewed the detection and quantification of genes that were expressed in very low amounts when mRNA or total RNA was used.
Figure 7. Scatter Plot Analysis of all Array Elements. A scatter plot of the Cy™5 vs. Cy™3 values obtained for an aRNA and a mRNA array is shown. The top two panels depict the 6 clusters (obtained after K-Mean Clustering Analysis) containing all 6912 elements. A subset of elements that are distinct between the two arrays and which deviate the most in signal intensity are depicted in the lower panels.
Amplification of RNA thus provides a means of measuring expression from genes transcribed at very low levels. In many cases the RNA concentration of an experimental sample is under the optimal required amount for synthesizing labeled cDNA for microarray analysis. MessageAmp is a viable technology for increasing the yield of useful probe and can greatly lower the starting amount of RNA required to produce biologically relevant signals.
In future columns we will continue to report results from MessageAmp microarray studies from Ambion researchers, our collaborators, and our customers.