A computational approach for genome-wide mapping of splicing factor binding sites

Abstract

Alternative splicing is regulated by splicing factors that serve as positive or negative effectors, interacting with regulatory elements along exons and introns. Here we present a novel computational method for genome-wide mapping of splicing factor binding sites that considers both the genomic environment and the evolutionary conservation of the regulatory elements. The method was applied to study the regulation of different alternative splicing events, uncovering an interesting network of interactions among splicing factors.

Figure 1

Schematic representation of the COS(WR) function. (a) A candidate human sequence is queried with a regulatory motif. (b) The weighted rank (WR) is computed only for significant positions by combining all scores above the suboptimal threshold in a sequence window of size w. (c, d) We calculate WR scores for the candidate's homologous region in mouse that aligns to the human sequence flanking the significant hits. (e) WR scores of the candidate sequence and its homologue are combined by calculating the Conservation Of Score (COS).

Figure 2

Sensitivity of multiplicity estimators. The average true positive rate (TPR) at a fixed false positive rate of 0.01 when training the data with four different multiplicity estimators: weighted rank (WR), weighted average (WA), median (M) and sum of scores (SS), compared to Single Scores (S). For each estimator the TPR was calculated when considering (dark columns) or not considering (light columns) the Conservation Of Score (COS).

Figure 3

Specificity calculated by the COS(WR) method. The percent of accurate predictions derived from a screening of experimentally validated sequences with 30 different SFBS queries. The x-axis shows the rank of the true positive hits (that is, experimentally validated SFBSs) among the list of predictions derived from the screening. The top curve displays the percent of predictions higher than the COS(WR) threshold and the bottom curve shows the percent of predictions below the threshold.

Figure 4

Specificity of the COS(WR) algorithm compared to ESEfinder. A pie chart representing prediction results for four SFs - SF2/ASF, SRp40, SRp55, and SC35 - obtained from screening experimentally validated sequences using (a) ESEfinder and (b) COS(WR). The different slices represent the percent of true SFBS predictions in the first, second, third, and fourth ranks (color scale is shown on the right). As shown, using the COS(WR) approach, 50% of predictions were ranked at the top rank, while only 9% were top ranked using ESEfinder. nf, not found.

Figure 5

Enrichment of SFBSs in alternative exons. A heat map representing the -log₁₀(P-value) of a series of Wilcoxon tests, comparing the normalized density of SFBS predictions in cassette exons (CE), alternative acceptors (AA), and alternative donors (AD) to a background of constitutive exons. The tests were carried out for the full exonic sequences (E), for 100-nucleotide intronic sequences (5' and 3') flanking the alternative exon and for extended regions 'exons and/or introns' (E/I). The P-values were corrected with the Westfall-Young procedure.

Figure 6

An induced subgraph of SF inter-regulation. The network represents AS regulation among SFs as predicted with the COS(WR) function. Arrows indicate that at least one of the alternative exons (and/or flanking introns) was predicted to be regulated by another factor. Light blue nodes stand for SFs that undergo AS and are thus part of the core network. SFs without AS support (the small gray nodes) are part of the extended network. The network is drawn in three layers: the upper layer displays SFs that have only out-edges (sources), the middle layer shows SFs that have both out-edges and in-edges (mixed), and the bottom layer includes SFs that have only in-edges (sinks). Graphs were drawn using Cytoscape [80].

Figure 7

Tissue specificity of the SFs. The TSI of SFs grouped according to their positions in the network: 'extended', 'source', 'mixed', 'sink', and 'self-regulatory'. As shown, low tissue specificity is observed for the top layers while higher tissue specificity is characteristic of the bottom layers.