A DNA microarray (DNA chip or biochip) is a technology used to identify and measure the expression level of specific mRNA molecules in order to ascertain transcriptional profiles in response to differing conditions. The most commonly used microarray is the Affymetrix® GeneChip® family of arrays. Each GeneChip® consists of a silicon chip with fixed locations called cells, spots or features [1]. Each spot contains millions of identical 25 base oligonucleotides (probes) which are selected to be complementary to various transcript regions of a gene [2]. In order to determine transcript expression, which directly infers gene expression, groups of 11-20 probes matching the same gene/transcript are arranged in a probe set. Given a particular Affymetrix® GeneChip® platform, the design of the probes is fixed based on earlier genome assemblies and annotation available at that time. Since the design of the first Affymetrix® GeneChip®, rapid progress has been made in genome sequencing resulting in more accurate databases of annotated coding and non-coding genes.

In addition, updating links between probe sets and their corresponding genes/transcripts does not provide a solution for problems caused by individual probes such as single nucleotide polymorphisms (SNPs) [4, 5], probes that target genes other than the designated gene of a probe set, and probes that no longer align to a genomic location. For example in the Affymetrix® GeneChip® HG-133 Plus 2 array, a total of 40,680 probes out of 603,158 (excluding quality control probes) do not have a perfect match to the most recent human genome assembly (hg38).

Several research groups have reassigned probes into new probe sets by creating their own custom Chip Description Files (CDF) [7–13], which are specially formatted files used to store the layout information for an Affymetrix® GeneChip® array. Given a CDF, the intensity values of probes located in the CEL file can be extracted and summarized as a defined probe set to detect the expression level of genes or transcripts.

These approaches have a similar workflow of mapping probes but differ in terms of the groupings of probe sets, including: data source used, the selected target level (gene or transcript), whether to create probe sets from scratch or redesign the existing groups and sharing probes between probe sets.

In terms of annotations used, most approaches have mapped the probe sequences to the transcripts obtained from one or more databases such as GenBank, NCBI RefSeq and Ensembl. Unlike other approaches, Harbig et al. [14] mapped to the target sequences of probes obtained from Affymetrix® rather than the actual mRNA sequences themselves, where the target sequence is an exemplar region of a specific transcript ≤600 bases in length. After mapping, they grouped probes to unique transcripts or genes based on the mapping results. Some approaches update the original probe set groups by removing select probes and changing the link between probe set and gene/transcript. The most comprehensive study for probe annotation remapping was achieved by Dai et al. (brainarray CDFs) [8]. Rather than focusing on one reference database or combining multiple sources to create one custom CDF, they mapped probes to different annotation databases and created a specific custom CDF for each database.

While microarrays have been successfully utilized for understanding differential expression at the gene or probe set level, less attention has been given to the potential analysis at the individual exon, alternative transcript, and untranslated region (UTR) level. While the selection bias of probes on the 3′ ends of genes for earlier iterations of Affymetrix® GeneChip® designs presents limitations on the completeness of transcript information, more recent designs allow for a more complete coverage of exons and exon junctions. However, information concerning individual exons can still be extracted from earlier GeneChip® designs, particularly in the 3′ UTR regions that have been shown to play important roles in cancer [15–17], development [18–22], and localization in the nervous system [23–27]. In fact, over 40% of genes have been shown to generate multiple mRNAs with variable 3′ UTR lengths [28]. These 3′ UTRs harbor binding sites for molecules including microRNAs (miRNAs) and RNA-binding proteins. Thus, mRNA isoforms with lengthened 3′ UTRs have increased numbers of sites for these cis-interacting factors. The diversity of 3′ UTRs is predominantly regulated by alternative polyadenylation (APA), which employs alternative mRNA cleavage sites that lie progressively distal to the stop codon. APA-driven mRNA diversity is required for normal physiology, and misregulation of this process is associated with diverse disease states [29]. We therefore have developed a framework for analysis of Affymetrix® GeneChip® data by regrouping probes into probe sets based on Ensembl annotations at the gene, transcript, individual exon, and UTR levels in order to detect changes in gene expression that may occur within specific regions of the transcript.

One of the limitations of microarray technologies is the design of probes based on available sequence and annotation data at the time of design. Based on our analysis, the percentage of uniquely mapping probes varies from 84% (rat) to 87% (human), indicating that changing knowledge about the genome itself plays a role in probe utilization. In terms of annotation, the rat genome is known to have more incomplete information when compared to mouse and human, which is reflected in the fact that only 47% of the rat probes lie in region-based locales (exons and UTRs) compared to 65% for mouse, and 69% for human. Since this can potentially lead to a small number of probes in each annotated region (and thus increased false positive rates), we have further required at least three probes be present in each probe set for our analysis. Both unrestricted (1 or more) and restricted (3 or more) probe groupings are available as CDFs.

To further illustrate the importance of region-based CDFs, using the subset of 4024 genes with probesets in both the CDS and 3′ UTR regions, we were able to identify 203 differential expression events at the 3′ UTR level that do not show differential expression within the CDS. In addition, these events are not detected using the standard Affymetrix® CDF. Further analysis of these 203 genes yields some genes of particular interest. For instance, the 3′ UTR of GRIK4 (Glutamate Ionotropic Receptor Kainate Type Subunit 4) was up-regulated (p-value 0.0450) while the CDS was not significantly regulated (FC = 1.07; p-value 0.4525), suggesting the 3′ UTR of this gene was lengthened (Fig. 4). GRIK4 regulates kainite-receptor signaling and neuroplasticity [38] and its missregulation is associated with neurological diseases including Alzheimer’s [39], bipolar disorder [40], and others. Interestingly, a deletion variant specific to the 3′ UTR of GRIK4 is protective of bipolar disorder [40]. Alongside our observation, this suggests that regulation of this plasticity-associated gene occurs though its 3′ UTR. We also observed that the 3′ UTR of VEGFA (vascular endothelial growth factor-A) was downregulated (−1.17 FC; p = 0.0102) and expression of its CDS was unchanged (1.01 FC; p = 0.8334) (Fig. 5). The 3’UTR of VEGFA, a potent neuromodulator, undergoes a well-described binary switch to regulate its expression [41]. Our observations suggest the VEGFA 3′ UTR undergoes an additional layer of regulation by shortening during collateral sprouting.

As our analysis with the GSE48611 and GSE72551 datasets show, reanalysis of publicly available datasets using updated annotations can yield additional information when compared to the use of the original CDFs. In our case, the region-based CDFs allow for a better understanding of 3′ UTR dynamics through the reanalysis of publicly available data. While current high-throughput sequencing technologies may allow for a more complete picture, this custom CDF approach will allow for deeper insight with only minimal computational cost, taking advantage of the high volume of publicly available GeneChip® data.

We proposed a framework for reannotating and reassigning probe groups for Affymetrix® GeneChip® technology based on functional regions of interest. Our work differs from others in that we annotated probes in UTR and exon levels in addition to gene and transcript (isoform) levels. We illustrated how this framework affects the detection of differentially expressed genes, particularly when focusing on functional regions of interest. Removing probes that no longer align to the genome without mismatches or align to multiple locations can help to reduce false-positive differential expression, as can removal of probes in regions overlapping multiple genes.

The main motivation of our work was profiling the contribution of UTR and exon regions to the gene expression levels globally. Our results indicate that features differentially expressed in either the gene-based or region-based CDF show high concordance and separating out into functional regions allows for the detection of subtle changes in transcript formation.