HITS-CLIP and integrative modeling define the Rbfox splicing-regulatory network linked to brain development and autism

Rbfox1, 2 and 3 HITS-CLIP in mouse brain

Considering the possibility that each Rbfox family member might differ in binding specificity despite their apparent functional redundancy, we performed HITS-CLIP experiments for all members of the family individually using mouse whole brain tissue. We firstconfirmed that the antibodies we used did not cross-react with different members and that they efficiently immunoprecipitated (IP) the targeted protein with minimal background under standard CLIP conditions (Figure S1; Supplementary Notes). Next, we used two different strategies to clone and amplify the isolated RNA fragments (Figure 1A). The first protocol, denoted as standard CLIP, was performed as described previously (Darnell, 2010; Licatalosi et al., 2008; Moore et al., 2014; Ule et al., 2005a). In this protocol, RNA linkers are ligated to the 5′ and 3′ ends of the RNA fragments (Figure 1A, left branch) and are later used for RT-PCR amplification. We and severalother groups have previously noted that after proteinase K digestion of the crosslinked protein-RNA complex, one or a few amino acids might remain attached to the RNA at the crosslink site, which cause informative errors at the crosslink site during reverse transcription (Granneman et al., 2009; Ule et al., 2005a). These crosslinking-induced mutation sites (CIMS) provide a footprint of protein-RNA crosslinking, and can be leveraged to determine protein-RNA interactions at a single nucleotide resolution (Moore et al., 2014; Zhang and Darnell, 2011). However, reverse transcription can abort prematurely at these sites, resulting in truncated cDNAs that lack the 5′ adaptor required for PCR (Konig et al., 2010; Sugimoto et al., 2012). To capture both truncated and non-truncated cDNAs, we developed a second CLIP protocol named BrdU-CLIP (Figure 1A, right branch). This protocol bears some conceptual similarity to individual nucleotide resolution CLIP or iCLIP (Konig et al., 2010). After ligation of the 3′ linker, purified RNA is reverse transcribed to introduce 5′ and 3′ PCR adaptor sequences separated by an apurinic/apyrimidinic endonuclease (APE) cleavage site. This is followed by the circularization of both read-through and truncated cDNAs and relinearization of cDNA via the cleavage site to place the 5′ and 3′ adaptor sequences in the correct orientation. One key difference between BrdU-CLIP and iCLIP is the incorporation of BrdUTP into the cDNA during reverse transcription so that the resulting cDNA can be purified in a stringent manner using an antibody that specifically recognizes BrdU (Core et al., 2008; Ingolia et al., 2009).

To evaluate the robustness of the Rbfox interaction sites, we prepared HITS-CLIP libraries for Rbfox1, Rbfox2, and Rbfox3 with 4, 4, and 5 biological replicates, respectively, which together resulted in about 870 million raw reads (CLIP tags). After stringent filtering, processing, and mapping (Moore et al., 2014; Zhang et al., 2010) (Experimental Procedures), we obtained a total of 4.6 million unique CLIP tags that represent independent captures of protein-RNA interactions, including 1,460,387 tags for Rbfox1, 868,366 tags for Rbfox2, and 2,308,632 tags for Rbfox3. Between 59-65% of these CLIP tags arelocated in introns, consistent with the known role of Rbfox proteins in regulating alternative splicing; an additional 23-28% unique CLIP tags are located in exons, mostly in the 3′ UTRs.

Rbfox1, 2, and 3 have similar protein-RNA interaction profiles

Initial inspection of the CLIP tag distribution suggests that the interaction profiles of the three Rbfox family members are very similar. For example, the Rbfox1 transcripts contain a cassette exon of 93 nucleotides (nt) (Figure 1B) encoding partof the RRM of the protein. Its skipping as a result of autoregulation generates a dominant negative form that lacks RNA-binding capability (Baraniak et al., 2006; Damianov and Black, 2010). Our CLIP data show that all three Rbfox family members bind to the upstream intronic sequences harboring a cluster of conserved UGCAUG elements, suggesting that this exon is under both auto-and cross-regulation by all family members. We also previously demonstrated that GABA receptor gamma 2 subunit (Gabrg2) exon 9 is under the synergistic regulation of Rbfox and Nova when they bind near the 5′ and 3′ splice sites of the downstream intron, respectively, based on detailed mutation analysis and splicing reporter assays in cell culture (Dredge and Darnell, 2003; Zhang et al., 2010). Our CLIP data now confirmed that Rbfox proteins indeed bind to the expected site in vivo inthe brain (Figure 1C).

To quantitatively compare the RNA-binding profiles of different Rbfox family members, we defined a non-redundant set of Rbfox-RNA interaction sites using all unique CLIP tags pooled together (Figure 1D; Experimental Procedures). A stringent set of 41,182genic CLIP tag clusters with at least one statistically significant peak (p<0.01) was obtained (Table S1), 70% of which are located in introns and the other 30% are in exons (Figure 1E). We then counted the number of CLIP tags per cluster for each protein. CLIP tags for different members are very well correlated in each pairwise comparison, especially between Rbfox1 and Rbfox3 (Pearson correlation R=0.97); the correlation between Rbfox2 and the other two members is somewhat lower (R=0.76-0.80; Figure 1F). In addition, we confirmed that the two CLIP protocols gave very reproducible results inthe global profiles (R=0.97; Figure 1G). Based on these observations, we conclude that the three Rbfox family members have similar RNA-interaction profiles on a genome-wide scale, consistent with the notion that their binding specificity is largely determined by their very similar RRMs. Although it remains possible that a small proportion of the binding sites could be preferentially recognized by a specific member, for this work, CLIPtags of all three members were pooled together for further analysis.

A single-nucleotide resolution map of Rbfox binding sites by CIMS and CITS analysis

Using two CLIP protocols in parallel allowed us to employ different strategies to pinpoint the exact Rbfox-RNA crosslink and interaction sites (Figures 2 and S2). For CLIP tags obtained by the standard protocol, we performed CIMS analysis using our established method (Figure 2A) (Moore et al., 2014; Zhang and Darnell, 2011). Nucleotide deletions were observed in 14% of standard CLIP tags, from which 1,424 reproducible CIMS were identified (FDR<0.001). A substantial enrichment of the Rbfox binding motif GCAUG was observed in the immediate vicinity of the reproducible deletion sites (Figure 2B). In contrast, when we analyzed substitutions and insertions using the same method, we did not observe elevated motif enrichment near the mutation sites (data not shown), suggesting that crosslinking predominantly, if not exclusively, introduces deletions rather than insertions or substitutions in Rbfox CLIP.

We then examined the enrichment of UGCAUG or VGCAUG (V=non-U) relative to the crosslink sites with reproducible deletions in more detail (Figure 2B inset). UGCAUG isenriched 28-41 fold at positions −5, −4, and −1 relative to the crosslink site, corresponding to crosslinking at G2, U5, and G6 of the UGCAUG element. Interestingly, enrichment of VGCAUG is most predominant at position −1 relative to the crosslink site (64 fold), corresponding to crosslinking of G2 (Figure 2C). We also examined the base composition of the sequences around CIMS regardless of the presence of (U)GCAUG, and observed a slight bias toward uridine compared to the flanking sequences (Figure S2A; see Discussion below). De novo motif analysis using sequences [−10,10] around CIMS uncovered (U)GCAUG as the only motif with strong enrichment (36% of 1,158 non-repetitive CIMS in [−10,10], E<3.7×10⁻³²⁰; Figure S2B). The first position ofthe motif is the most variable, which is consistent with previous findings that Rbfox binds to both UGCAUG and VGCAUG with high affinity (Jin et al., 2003; Ponthier et al., 2006). Additional deviations from the consensus appear to be tolerated to some extent (e.g., in positions 3 and 4), providing a partial explanation for why (U)GCAUG is not present at all crosslink sites (Figure S2 E-G). Finally, we observed deletions in 6.2% of BrdU-CLIP tags, and analysis combining standard and BrdU-CLIP tags defined 2,298 CIMS (FDR<0.001; Table S2).

The sensitivity of crosslink site identification by CIMS analysis is limited by therelatively low deletion rate among tags that are read through. We therefore looked for reproducible cross-linking induced truncation sites (CITS) in BrdU-CLIP data (Figure 2D; Experimental Procedures). Overall, 6,606 robust CITS were identified (p<0.001; Table S3). Among these, the UGCAUG element is enriched 319-fold at a single position (−5) relative to CITS, corresponding to predominant crosslinking at G6 of the motif (Figure 2E). The same position was crosslinked in the VGCAUG element, although there is less enrichment of the motif (88 fold; Figure 2F). Analysis of the base composition [−10,10] around CITS revealed the UGCAUG motif directly (Figure S2C), and this was confirmed by de novo motif analysis (60% of 1000 randomly sampled non-repetitive sites; E<1.9×10⁻⁶³¹; Figure S2D). As a control, we repeated the same analysis in the standard CLIP data, which presumably lacked truncated tags, and did not observe enrichment of the (U)GCAUG motif in these specific positions (Figure S2 H,I).

Together, our data suggest that the two guanines G2 and G6 in the (U)GCAUG motif are particularly prone to crosslinking with the Rbfox protein. Consistent with this finding, examination of a previously determined NMR structure of the Rbfox1 RRM in complex with UGCAUGU RNA revealed that these two guanines are buried in two pockets of the RRM, and are stabilized by multiple hydrogen bonds and stacking interactions (Figure 2G); mutations in each of these two nucleotides resulted in the largest increases in the free energy of binding (Auweter et al., 2006). Based on the single-nucleotide-resolution map of in vivo Rbfox interaction sites and on the characterized specificity of the proteins, we developed the motif enrichment and conservation score (MECS) of (U)GCAUG elements by comparing CLIP tag clusters and regions without CLIP tags (Figure S2 J,K and Supplementary Notes). A motif site with higher conservation receives a higher score, especially if it is located in an intronic region. A UGCAUG element receives a higher score than a VGCAUG element with the same level of conservation, reflecting greater enrichment of the former in CLIP tag clusters.

Identifying Rbfox-dependent exons using RNA-Seq

We previously used HeLa cells with perturbed Rbfox1 or Rbfox2 expression to validate over half (55-59%) of bioinformatically predicted Rbfox target alternative exons tested with RT-PCR (Zhang et al., 2008). We therefore used this established experimental system to expand the list of Rbfox-dependent exons by RNA-Seq, which provides information complementary to the CLIP data. As we described previously (Zhang et al., 2008), HeLa cells were treated with a short hairpin RNA (shRNA) targeting Rbfox2 (shRbfox2) to generate stable knock-down (KD) of the protein, which is endogenously expressed at a level that is low but sufficient for splicing regulation; HeLa cells expressing the empty vector were used for comparison (control; Figure 3A, left panel). In total, RNA-Seq of the control and shRbfox2 samples resulted in 60 million and 48 million paired-end reads, respectively, of which 62-65% were mapped unambiguously to the genome or to the exon-junction database.Examination of the gene-expression level confirmed that Rbfox2 was specifically knocked down 3.3-fold (Figure 3A, right panel), consistent with the protein level changes observed from immunoblot analysis. Accordingly, we were able to identify 126 cassette exons, 17 tandem cassette exon events, and 4 mutually exclusive exon events showing Rbfox2-dependent inclusion or exclusion [FDR ≤ 0.1 and proportional change of exon inclusion |ΔI|≥0.1 (Ule et al., 2005b)] (Figure 3B and Table S4). Among the 22 cases of alternative exons we tested by RT-PCR (which includes 3 cases with a read coverage slightly below the threshold we used), 21 showed Rbfox2-dependent splicing (Figure 3C,D; Tables S4 and S5), giving a validation rate of 95%. The direction of Rbfox2-dependent splicing of these exons can be predicted by the position-dependent RNA map based on either the CLIP data derived from mouse brain or the bioinformatically predicted motif sites (data not shown), indicating that these Rbfox2-dependent exons are enriched in direct Rbfox targets in the brain.

Integrative modeling defines the Rbfox target splicing-regulatory network

To comprehensively define the functional target network directly regulated by the Rbfox proteins, we took an integrative modeling approach which we recently developed and successfully applied to study the Nova target network (Zhang et al., 2010). This method uses a Bayesian network to probabilistically weigh and combine multiple types of data complementary to each other: bioinformatically predicted motif sites represented by MECS scores, protein-RNA interaction sites mapped by Rbfox HITS-CLIP, Rbfox1-dependent splicing identified by comparison of WT with Rbfox1 KO mouse brain using exon-junction microarrays (Gehman et al., 2011), Rbfox2-dependent splicing in HeLa cells as described above, tissue-specific splicing as measured by RNA-Seq [only for training; ref. (Brawand et al., 2011)], and evolutionary signatures including preservation of reading frame and conservation of alternative splicing pattern (Figure 4A and Experimental Procedures).

Focusing initially on cassette exons, we found that the estimated model parameters confirmed, quantified, and extended our understanding of Rbfox splicing regulation (Figure 4B-E, Figure S3A-E and Supplementary Notes). For example, strongermotif sites are more likely to be bound by the protein (Figure 4B), and regions inferred to be bound by Rbfox have more CLIP tags than those inferred not to be bound (Figure 4C). In addition, the model was able to quantify the position-dependent RNA map: binding of Rbfox in the downstream intron is predicted to result in Rbfox-dependent inclusion with a probability of 0.99, while binding of Rbfox in the upstream intron or exon is predicted to result in repression with a probability of 0.75 and 0.61, respectively. Binding of Rbfox in both exon and upstream intron is expected to increase the probability of repression to 0.84 (Figure 4D).

The model was then applied to each annotated cassette exon in the mouse genome to predict the probability of its activation or repression by Rbfox through direct protein-RNA interactions. After using 10-fold cross validation to ensure the model was not overfit (Figure S3F), we predicted 772 cassette exons as direct Rbfox targets (FDR<0.05; Table S6). Among these targets, Rbfox was predicted to activate 421 exons (probability of activation>0.7) and repress 113 exons (probability of repression>0.7), respectively. For the remaining 238 exons predicted as Rbfox targets, the Bayesian network was unable to assign the direction of regulation unambiguously (Figure 4F, left panel). This uncertainty is presumably due to a lack of observed Rbfox-dependent splicing in the current experimental settings and to binding of Rbfox in both upstream and downstream introns simultaneously (Figure 4F, right panel). Based on comparison of the predicted exons with previously validated Rbfox-regulated exons compiled from the literature, we estimated that our Bayesian network analysis has a sensitivity of 73-79% (Supplementary Notes, Figure S4 and Table S7). We also compared the results of the Bayesian network analysis to our previous motif-based bioinformatic predictions and to another recent study that predicted Rbfox target exons based on the presence of the Rbfox motif sites and the correlation of exon splicing with Rbfox expression (Ray et al., 2013). These comparisons gave substantial overlap between exons predicted by different methods, but also highlighted that the Bayesian network analysis effectively integrated features known to be consistent with regulated alternative splicing events, such as preservation of the reading frame and conservation of the alternative splicing pattern (Figure S5 and Supplementary Notes).

After we confirmed the performance of the Bayesian network, we applied the model toother types of alternative splicing events, and predicted 212 events of tandem cassette exons (300 exons, Table S6) and 75 events of mutually exclusive exons (107 exons, Table S6) as direct Rbfox targets. Altogether, 587 genes have one or more alternative splicing events directly regulated by Rbfox. To understand the molecular function of these genes, we performed gene ontology (GO) analysis, and found very significant enrichment of genes with annotated function in “cytoskeleton” (Benjamini FDR < 3.8×10⁻¹⁷) and “neuron projection” (Benjamini FDR < 3.8×10⁻⁸), compared to all brain-expressing genes (Table S8). In addition, proteins encoded by Rbfox target transcripts are enriched in PDZ domains that are known to be important for anchoring transmembrane proteins to the cytoskeleton and for functioning as scaffolds for signaling complexes (Benjamini FDR < 7.2×10⁻⁸) (Ranganathan and Ross, 1997).

Rbfox regulates global dynamic splicing changes during brain development

It has previously been shown that all three Rbfox family members undergo increased expression in the mouse brain at prenatal stages between embryonic day E12 and E18 (Tang et al., 2009), and that the increase of Rbfox1 expression further extends into postnatal stages (Hammock and Levitt, 2011). In mouse and chicken, the differential expression of Rbfox proteins is correlated with splicing changes in several exons during CNS development(Kim et al., 2013; Tang et al., 2009), but how Rbfox proteins affect the global switch ofthe developmental splicing program is unclear. We found that Rbfox family members undergodynamic expression change between E17 and adult mouse cortex, as evaluated from a published RNA-Seq dataset (Dillman et al., 2013). Rbfox1 and Rbfox3 show 1.6-fold (p<0.02; t-test) and 3.1-fold (p<10⁻⁶; t-test) increases in the adult cortex compared to E17 cortex, respectively, while the expression of Rbfox2 is reduced 3.1-fold (p<10⁻⁴; t-test; Figure 5A). The expression changes of the Rbfox proteins parallel the splicing changes of Rbfox target exons: 55% of Rbfox target exons with sufficient read coverage to quantify splicing show splicing changes between the two developmentalstages, as compared to 32% for exons not regulated by Rbfox (odds ratio=2.4, p<7.6×10⁻¹⁶; Fisher’s exact test; Figure 5B). In addition, a majority (77%) of the exons activated by Rbfox have increased exon inclusion in the adult, while over half (57%) of the exons repressed by Rbfox have decreased exon inclusion (odds ratio=4.4, p<0.003; Fisher’s exact test; Figure 5C). This asymmetry indicates that increased expression of Rbfox1 and Rbfox3 predominates the developmental splicing change of their targets, and in general they promote the switch to the adult splicing program through direct regulation.

Rbfox target genes are linked to autism

We previously demonstrated that Rbfox target transcripts predicted bioinformatically based on conserved motif sites have significant overlap with genes implicated in autism, supporting the notion that disruption of either Rbfox1 itself or of its targets observed in autism patients is likely pathogenic (Zhang et al., 2010). This hypothesis was further supported by recent findings that RBFOX1 lies in a hub in gene co-expression networks based on microarray profiling of autistic and control postmortem human brains and its reduced expression in a subset of autism patients is correlated with altered splicingof predicted Rbfox target exons (Voineagu et al., 2011). The comprehensive Rbfox target network defined by integrative modeling now allows us to examine the link between RBFOX1 and autism in more detail.

Among the 235 Rbfox target cassette exons that are conserved in human and that havesufficient RNA-Seq read coverage to evaluate splicing change in autistic versus control brains (Voineagu et al., 2011), 97 (41%) show alteration of splicing in autistic brains (|ΔI|≥0.1, and FDR≤0.05), a very significant overlap compared to random chance (odds ratio=3.4, p<1.4×10⁻¹⁶; Fisher’s exact test; Figure 6A). The autistic brains compared here were selected to have low expression level of Rbfox1 (Voineagu et al., 2011) (4.1-fold down-regulation compared to control brains as measured by RNA-Seq; p<0.05, t-test; Figure 6B). However, the massive splicing change of Rbfox targets detected in autistic vs. control brains is somewhat surprising given the redundant role of the other Rbfox family members, as observed in Rbfox1 KO mice(Gehman et al., 2012; Gehman et al., 2011). Interestingly, we found that the expression of the other two family members, Rbfox2 and Rbfox3, was also down-regulated 3.3 fold (p<0.05; t-test) and 3.2 fold (p<0.09; t-test), respectively. Therefore, simultaneous down-regulation of all Rbfox family members might explain the massive splicing misregulation of Rbfox target exons observed in these autism patients. On the other hand, many of thesplicing changes observed in autism patients may not be regulated by Rbfox proteins directly.

To focus on Rbfox target genes that are likely genetic risk factors of autism, we examined candidate autism-susceptibility genes in the SFARI autism gene database (Basu et al., 2009). Among the 519 candidate autism-susceptibility genes with mouse orthologs, 48 were identified as Rbfox targets by Bayesian network analysis (odds ratio=2.8, p<9.6×10⁻¹⁸, Fisher’s exact test; Table 1 and Table S6). The list includes three genes that are currently regarded as causal in syndromic autism spectrum disorders (ASD): Shank3 (Phelan-McDermid Syndrome), Cacna1c (Timothy syndrome), and Tsc2 (tuberous sclerosis complex). For a specific example, Rbfox is predicted to activate the inclusion of alternative exon 25 in the Tsc2 gene, which isconserved between human and mouse (Figure 6C). Although the function of this alternative exon has not been characterized, its inclusion was recently shown to be dependent on Rbfox2 using cell culture models of epithelial-to-mesenchymal transition (Braeutigam et al., 2013).

Inclusion or exclusion of Rbfox target exons mostly introduces alteration in local amino acid sequences. However, we identified several cases (Fat1, St7, Scn2a1, and Scn8a) in which alternative splicing is potentially coupled with non-sensemediated mRNA decay [NMD; ref. (Maquat, 2004)]. In Scn2a1 for instance, Rbfox is predicted to activate a cryptic exon harboring an in-frame premature stop codon via extensive binding to the downstream intron (Figure 6D). Inclusion of this exon is undetectable in the brain from RNA-Seq data, presumably due to NMD of the inclusion isoform. Interestingly, while this alternative exon is conserved in vertebrates, compensatory mutations in the stop codon have accumulated. This has resulted in different stop codons in different species, suggesting an evolutionary selection pressure to preserve a stop codon. Scn2a1 is one of the few genes with recurrent de novo gene-disrupting mutations according to exome-sequencing of ASD patients (Sanders et al., 2012), and our analysis suggests that the same gene can potentially be disrupted in autism by different mechanisms.

We tested a select set of exons in potentially autism-related genes for regulation by Rbfox1 in the mouse brain. The splicing level of these exons was measured in CNS-specific Rbfox1-KO mouse brains compared to WT controls by semi-quantitative RT-PCR. Among eight exons for which both the inclusion and exclusion isoforms can be detected, three exons showed altered splicing upon Rbfox1 depletion (ΔI≥0.1; Figure S6 and Table S5), and these were not identified as Rbfox1 targets in the previous genome-wide analysis (Gehman et al., 2011). The relatively small number of validated exons is likely due to the redundancy of Rbfox proteins and the compensatory increase of Rbfox2 protein expression in Rbfox1 KO animals (Gehman et al., 2011), although it is possible that some of the predicted targets might represent false positives.

HITS-CLIP Experiments using Mouse Brain

Antibodies used in HITS-CLIP experiments for individual Rbfox members were tested to ensure that there is no or minimal cross-reactivity between family members and to optimize the signal-to-noise ratios of IP in the CLIP condition. Rbfox1, 2, and 3 CLIP experiments were performed with whole brain tissue lysate of P15 CD1 mice using two different protocols. The standard CLIP libraries were prepared as described previously (Darnell, 2010; Licatalosi et al., 2008; Moore et al., 2014). Details of the BrdU-CLIP protocol are described in the Supplemental Protocols online.

Bioinformatic analysis of CLIP data was performed essentially as described previously to obtain unique CLIP tags and define CLIP tag clusters and peaks (Charizanis et al., 2012; Moore et al., 2014; Zhang and Darnell, 2011). Mutations in unique CLIP tags were recorded for CIMS analysis as described previously (Moore et al., 2014; Zhang and Darnell, 2011). CITS analysis was performed using a computational method similar to that used to analyze iCLIP data with several modifications (Konig et al., 2010; Wang et al., 2010). We removed CLIP tags with deletions, which presumably had read through the crosslink sites, and clustered the remaining tags based on their potential truncation sites, then used a stringent statistical test to evaluate the significance of the observed truncation frequency.

RNA-Seq in HeLa Cells

HeLa cells were infected with MSCV expressing shRbfox2 or the empty retroviral vector (control) (Zhang et al., 2008). To reduce splicing precursors and intermediates, cells were fractionated to enrich cytoplasmic RNA before sequencing. The downstream analysis was performed as described previously (Charizanis et al., 2012) except that the proportional change of exon inclusion was estimated using the ASPIRE algorithm (Ule et al., 2005b).

RT-PCR Validation

Semi-quantitative RT-PCR in HeLa cells was performed as described previously (Zhang et al., 2008). For RT-PCR validation, total RNA was extracted from brains of WT and homozygous Rbfox1 KO mice. Candidate exons for validation were selected to cover the whole range of prediction confidence, avoiding exons with complex splicing patterns orvery high or low inclusion levels.

Integrative Modeling using Bayesian Network Analysis

We used our established Bayesian network framework with some modifications (Zhang et al., 2010). We considered differential splicing between WT and Rbfox1 KO mousebrain, differential splicing between shRbfox2 and control HeLa cells, tissue-specific splicing of exons in cortex and cerebellum compared to other tissues (liver, kidney and testis), and tissue-specific splicing of exons in heart compared to other tissues (liver, kidney and testis). For motif sites, we considered both UGCAUG and VGCAUG elements, and measured their strength by their MECS score.

To train the model, we used 121 cassette exons compiled from the literature that have been previously validated as Rbfox targets in humans or mice and assigned them a class label (activation or repression). We also included unlabeled exons that showed relatively large or no changes in inclusion in exon-junction microarray or RNA-Seq data. Thisresulted in 551 non-redundant exons, representing both targets and non-targets, to estimate model parameters. The trained Bayesian network model was applied to 16,034 cassette exons, as well as 4,074 events of tandem cassette exons and 960 events of mutually exclusive exons, to predict direct Rbfox targets.

Details of all experimental protocols and computational analyses are described in the Extended Experimental Procedures.