The recognition of alternative splicing and alternative isoform expression as an important component in gene expression analysis has prompted the introduction of isoform sensitive microarray platforms. By targeting individual exons, exon junctions, and annotated isoform variants, such platforms possess the ability to profile not only the expression levels of the entire transcript, but also variations in the types of expressed isoforms. The Affymetrix Exon Array 1.0 ST is one of such commercially available platforms. To date, it has been shown that the Exon Array produces gene expression measurements that are comparable with the previous generation 3' targeted arrays. However, little is known about the in-depth level of similarities and particularly differences among WT and 3' based technologies.
This comparison utilizes the well studied brain and reference samples previously used in the MAQC study to determine sources of variability in profiling gene expression using microarrays. These samples are particularly valuable for the purposes of benchmarking the performance of the Exon Array for two reasons: 1) they allow easy comparison of gene expression level measurements with other platforms that have already been tested, and 2) they allow detection of alternative splicing and isoform difference, since neural tissues are known to be particularly prone to alternative splicing.
Our first conclusions concern the utility of the Exon Array as an expression profiling tool. We note that although the Exon Array results are very consistent with 3' profiling methods, the level of agreement between the Exon Array and 3' targeted platforms (Illumina and Affymetrix U133) is slightly lower than the agreement between the 3' platforms. There are at least two reasons for the decreased concordance.
Firstly, the Exon Array uses a whole transcript, randomly primed amplification protocol, while the two other platforms rely on polyA tail priming. As a result, the two approaches amplify a slightly different RNA pool. This is illustrated very well by the example of several histone genes (known to lack a polyA tail), which the Exon Array indicates are expressed at a much lower level in the brain than in the reference, while the other two platforms indicate a uniform very low level of expression of histone transcripts. As far as we know, differences in expression of histone genes across tissues and treatments have not previously been detected by microarray analysis, and this result is only detectable using the WT approach.
Secondly, many of the outliers in the correlation plot (Figure ) are due to the presence of real variations in the expression of specific isoforms. This is illustrated using a previously noted example of the ELAVL1
gene, which showed discordance across platforms in the original MAQC study, as well as in additional new examples (Additional file 1
). The detected expression differences of transcript variants may have important biological significance. For example the longer 3' UTR in the dominant ELAVL1
transcript in brain has a different set of putative micro RNA binding sites than the shorter 3' UTR in the reference RNA.
It should also be noted that discordant results will often be obtained because of differences in the annotation provided by microarray manufacturers. We circumvented most of such problems here by re-mapping the probes and selecting only a subset of genes that we were confident were correctly targeted by all three platforms, but researchers should keep in mind that the annotations and gene assignments provided by manufacturers contain numerous errors [31
]. In the case of the Exon Array, we found that the most common annotation error resulted from joining together distinct transcripts into single meta-probesets, particularly in the case of transcripts that partially overlap. Thus, we recommend that lists of candidates from individual experiments should be carefully curated.
We also outline how the Exon Array can be used to detect alternative splicing and alternative mRNA processing events. Although our analysis methods are not in themselves novel, and most of them have been briefly described elsewhere [12
], our goal is to convey to the potential users their intuitive appeal and potential pitfalls. The most challenging step remains the decoupling of whole transcript expression, and individual probeset inclusion. The simplest solution to this problem is to consider only the genes that do not change overall expression levels, but contain probesets that exhibit individual variations. Although this approach produces a highly confident set of alternative events, it can result in a huge reduction of the dataset, particularly in case of comparisons across samples with highly heterogeneous gene expression levels. In the case of MAQC dataset, which has been chosen for the exact reason of it's extreme gene expression variability, imposing the restriction of expression fold change of less than 2 reduces the total number of genes considered by 31% (from 17665 to 12198).
A more inclusive approach is to attempt to correct for gene expression differences that may occur concurrently to splicing differences. We discuss two such approaches: 1) the splicing index, which compares probeset inclusion across samples after normalizing by gene expression levels, and 2) two-way ANOVA, where the interaction term between sample type and probeset can be used to indicate differential inclusion of probesets within transcripts. Both approaches suffer from similar systematic biases; they assume a uniform (linear or log-linear) response of each probeset within a meta-probeset. This assumption is violated in many cases, particularly for probesets that hybridize at very high levels (saturated response) or probesets with hybridization levels close to background (poorly or non-responsive). As a result, in the presence of significant gene expression changes, such analyses predominantly indicate three types of events: dead probesets, saturated probesets, and probesets that may be predominantly skipped (alternative), but not necessarily differentially included across samples. All three types of results constitute false positives, and contribute to the high false positive rates of such analyses.
We also point out two major systematic errors. First, we show that hybridization intensity decreases for probesets close to the 3' mRNA ends, an effect that we believe stems from the random amplification protocol used by the Exon Array. We argue that this is not an annotation artefact, but most likely results from the end of template and reduced random priming potential in the first strand synthesis step amplification. As a result, 3' regions of genes are detected at near background levels, and frequently indicate alternative isoform presence using the SI or ANOVA approaches. A similar problem exists at the 5' end of transcripts, where we hypothesize that the reduction in DABG levels is caused by the elevated GC content of the probesets. These problems are particularly troubling, since many cases of alternative polyadenylation and promoter usage may in fact be associated with changes in transcript expression. This may be due to different promoter strength, or microRNA mediated regulation in 3' UTR (as is likely to be the case in the ELAVL1 example shown in Figures and ). Such real and potentially extremely interesting cases may be difficult to distinguish from differences in probe hybridization potential.
Many of the above systematic errors can be avoided by filtering out potentially troublesome subsets of the data: probesets with extremely low variability (saturated), probeset with low inclusion levels (close to background), and genes with extremely high differences in expression levels across samples. However, such filtering decreases the false positive rates at the cost of reduced genomic coverage.
In our earlier studies, we have also pointed out that in many experimental designs, particularly when samples originate from different genetic backgrounds (e.g. different individuals), the presence of sequence variants within probe target sequences may be a very significant source of errors [8
]. This effect can be especially prominent in eQTL association studies, where we have shown that it can be responsible for a false positive rate > 80% in alternative splicing analysis [32
]. Thus, unless all tested samples are isogenic, we highly recommend additionally "masking" all probes containing known polymorphisms before performing the analysis.