Fig. 4

Many RBPs promote proximal or distal APA isoform abundance in hundreds of genes. (A) Correlation of all ψ values across HepG2 and K562 cell lines for all ENCODE RBP-knockdown RNAseq experiments. In gray, correlation coefficients for comparisons of different RBP knockdowns are shown (e.g. RBP X in HepG2 vs. RBP Y in K562). In yellow, correlation coefficients for comparisons of the same RBP knockdown are shown (e.g. RBP X in HepG2 vs RBP X in K562). In red, this comparison is restricted to only those genes whose ψ value significantly differed between the RBP knockdown and control knockdown samples (e.g. RBP X in HepG2 vs RBP X in K562, significant Δψ genes only). In identification of these significant genes, the cell line was included as a covariate. (B) Comparison of ψ values in RBP knockdown and control samples for genes whose ψ value was significantly different between these samples (FDR < 0.05). The number of genes with significant Δψ values in each comparison is indicated by the bar graph. A term, α, was defined as the fraction of these genes that displayed higher ψ values in the high RBP state (control knockdown) versus the low RBP state (RBP knockdown). (C) For each RBP knockdown, the number of genes with significant Δψ values (FDR < 0.05) is indicated on the y axis while the fraction of these genes with positive Δψ values (control knockdown - RBP knockdown) is indicated on the x axis. Knockdowns whose fraction of genes with positive Δψ values significantly differs from the expected 50% are indicated with red circles. (D) α values for each RBP knockdown in HepG2 cells were calculated using tandem UTR and ALE genes independently. These were then plotted and correlated. Each dot in this plot represents one RBP knockdown experiment. (E) Among 84 RBPs expressed in HepG2 cells, overlaps between the genes whose APA was sensitive to RBP knockdown and the genes whose 3′ UTRs were bound by the RBP in eCLIP experiments were calculated. The significance of this overlap was calculated using a binomial test. 21 RBPs bound the 3′ UTRs of their APA targets more often than expected (binomial p < 0.05). To assess whether this was more than the expected number of significant RBPs, relationships between RBPs and their lists of APA and eCLIP targets were shuffled 1000 times, and the analysis was repeated after each shuffle to create a null distribution (pink)