Identifying differential exon splicing using linear models and correlation coefficients

With the availability of the Affymetrix exon arrays a number of tools have been developed to enable the analysis. These however can be expensive or have several pre-installation requirements. This led us to develop an analysis workflow for analysing differential splicing using freely available software packages that are already being widely used for gene expression analysis. The workflow uses the packages in the standard installation of R and Bioconductor (BiocLite) to identify differential splicing. We use the splice index method with the LIMMA framework. The main drawback with this approach is that it relies on accurate estimates of gene expression from the probe-level data. Methods such as RMA and PLIER may misestimate when a large proportion of exons are spliced. We therefore present the novel concept of a gene correlation coefficient calculated using only the probeset expression pattern within a gene. We show that genes with lower correlation coefficients are likely to be differentially spliced.

The LIMMA approach was used to identify several tissue-specific transcripts and splicing events that are supported by previous experimental studies. Filtering the data is necessary, particularly removing exons and genes that are not expressed in all samples and cross-hybridising probesets, in order to reduce the false positive rate. The LIMMA approach ranked genes containing single or few differentially spliced exons much higher than genes containing several differentially spliced exons. On the other hand we found the gene correlation coefficient approach better for identifying genes with a large number of differentially spliced exons.

We show that LIMMA can be used to identify differential exon splicing from Affymetrix exon array data. Though further work would be necessary to develop the use of correlation coefficients into a complete analysis approach, the preliminary results demonstrate their usefulness for identifying differentially spliced genes. The two approaches work complementary as they can potentially identify different subsets of genes (single/few spliced exons vs. large transcript structure differences).