U2AF65 plays an important role during splicing and is known to bind to the PY tract region located between the branchpoint sequence and the acceptor splice junction 38. Visual inspection of human introns reveals that, although the PY tract region is enriched in uridines in general, there is a great deal of sequence variation between introns. This degeneracy, at least in part, appears to reflect the low RNA site specificity that U2A65 displays compared to other RNA binding proteins that evolved to recognize highly specific targets. U2AF65 binds with high affinity to contiguous runs of uridines but appears to tolerate moderate interruptions of other nucleotides 27394041. Despite the ability of U2AF65 to bind to degenerate sites, an effective binding site must still be composed primarily of uridines 4041. However, many thousands of human introns contain PY tracts that do not contain any sequences that are likely to be effective binding sites (shown below). Many of these PY tracts either contain contiguous runs of cytidines or contain numerous purines, neither of which are likely to represent binding sites for U2AF65 4041. Therefore, it is likely that individual human intronic PY tracts possess a wide range of affinities towards U2AF65, and that many may possess only weak binding sites for it. It is possible that additional cis-sequence elements augment the role of the PY tract during splicing, and that such elements play crucial roles in splicing in the absence of a strong U2AF65 binding site.

Many human introns have been shown to be enriched in motifs containing GGG in the region upstream of the PY tract 4243 (Figure 1a). This observation demonstrates that this region is under compositional selection. G-triples located upstream of a weak PY tract have been shown to affect splice site usage 20. We hypothesized other cis-elements may also be located upstream of the PY tract and may compensate for PY tracts containing weak U2AF65 binding sites. To explore this possibility we performed a computational analysis to determine if the region upstream of the PY tract is enriched in specific motifs when the PY tract does not contain a strong U2AF65 binding site.

In order to carry out this analysis, we first needed to correlate the composition of the PY tract of introns with likely affinities towards U2AF65. Several theoretical models have been presented that describe the relationship between binding site composition and the ΔG of binding between nucleic acids and nucleic acid binding proteins 4445. These models require the use of a positional frequency model representing the preferred binding site. In vitro selection (SELEX) experiments using human U2AF65 did not reveal a well defined consensus motif shared by high affinity RNAs 2739. Several computational methods have been developed to define a degenerate consensus motif from a population of sequences that are thought to contain a common, but unknown, motif 4647. Though such methods have proven useful, each has its own weaknesses, and all such predictive methods introduce an added level of uncertainty. We decided to develop a computational method to predict the affinity between a short RNA sequence and U2AF65 that is independent of knowledge of a particular consensus binding motif. We refer to this score as an S₆₅ score. The S₆₅score, for a given intron, is the average degree to which all pentamers (using a sliding window) found in the PY tract region (-30 to -3 relative to the acceptor splice-junction) are themselves enriched within the SELEX derived sequences (see Materials and methods for a complete description).

For this analysis, the PY tract was defined as the region from -30 to -3 (relative to the acceptor splice junction). This region is highly enriched in the pentamers that are most abundant within the U2AF65 selected sequences (Figure 1a and data not shown). Although a small number of introns are thought to possess functional U2AF65 binding sites upstream of this region 48, the general enrichment for uridines in this region (Figure 1a) is consistent with the premise that the bulk of U2AF65 functional binding sites are located adjacent to the acceptor splice-junction.

The S₆₅ scores for the SELEX RNAs appear to be normally distributed with a mean of 1.5 (Figure 1b). In contrast, the S₆₅ scores for human PY tracts display a slightly skewed distribution with a mean of 0.877 and a median of 0.811. These are shifted significantly to the left (that is, weaker) relative to the scores for the U2AF65 selected RNAs, suggesting that a large portion of human PY tracts represent weaker than optimal U2AF65 binding sites.

We chose to classify PY tracts that score below the median of 0.811 as 'weak' PY tracts and those above 0.811 as 'strong' PY tracts or likely to have high affinity U2AF65 binding sites. Using this designation, only a single SELEX-derived sequence scores as 'weak'. We are therefore asking whether there are statistically significant differences in the composition of the -80 to -30 region of two types of introns: ones that contain a PY tract with affinities similar to those derived using SELEX, and those with PY tracts with lower affinities.

In order to asses the relationship between the S₆₅ score and observed U2AF65 binding affinities, we evaluated the binding of recombinant human U2AF65 to several human PY tracts of varying S₆₅ scores using gel-shift mobility assays (Figure 2). We chose one PY tract that had a very low score (MBNL1 intron 6, S₆₅ = 0.0750). This PY tract is interrupted by several purines that are expected to impair U2AF65 binding. We also evaluated three other low-scoring PY tracts with scores closer to the median, and, therefore, correspond to the more 'typical' human PY tract: BRUNOL4 intron 9 (S₆₅ = 0.3602), ITGB4 intron 31 (S₆₅ = 0.3608), and LCAT intron 4 (S₆₅ = 0.5068). All three of these are cytidine-enriched. In addition, we tested three high-scoring PY tracts that had scores spanning the higher range of the distribution: INSR intron 10 (S₆₅ = 0.9593), U2AF2 intron 6, (S₆₅ = 1.1787), and SR140 intron 9 (S₆₅ = 1.8434), and an altered version of the LCAT intron 4 in which the central region was modified to contain an eight nucleotide poly-uridine run (LCATmut with a S₆₅ of 1.2060). All four of these high-scoring sequences are uridine-enriched. Binding data were also obtained using two sequences derived from the PY tract of the adenovirus major late (ADML) pre-mRNA, similar to previously studied ADML PY tracts 3249.

We expected the MBNL1 intron 6 PY tract to represent the weakest U2AF65 binding target and observed no detectable levels of U2AF65 binding at the protein concentrations tested (Figure 2). Meanwhile, all three of the cytidine-rich sequences with moderate S₆₅ scores demonstrated moderate affinities in the binding assay. In contrast, three of the uridine-rich sequences (with high S₆₅ scores) bound with high affinity. An interesting exception was the INSR-derived sequence, which bound U2AF65 more weakly than the more cytidine-rich LCAT-derived sequence. Importantly, for both LCAT and ADML, the binding of the mutant versions correlates well with the predicted affinities based upon the S₆₅ score.

Overall, there is a good agreement between the observed binding affinities for U2AF65 and the predicted affinities based upon the S₆₅ score. Plotting the observed K_d values versus the predicted S₆₅ score revealed that the ln of the K_d appears to be linearly related to the S₆₅ score (Figure 2c). Since ΔG is related to K_d according to the equation ΔG° = -RTln(K_d), this is consistent with the supposition that S₆₅ is linearly related to ΔG. Linear regression of the observed affinities and S₆₅ scores demonstrates that these values are strongly correlated (R² = 0.77; Figure 2c). Some of the observed deviations may be due to influences of RNA secondary structures present in some of the templates. Such secondary structure could greatly influence U2AF65 interactions, but this parameter is not addressed in the S₆₅ score. Since U2AF65 is known to have a strong preference for uridines, it is possible that the observed binding affinities simply reflect overall uridine content. However, linear regression analysis of the uridine content versus binding affinities demonstrates that these values are not well correlated (R² = 0.27, data not shown). Therefore, the S₆₅ score is a better predictor of binding affinity than uridine content alone and suggests that U2AF65 is recognizing sequence features more complex than the simple presence or absence of contiguous runs of uridines.

It is possible that introns containing weak U2AF65 binding sites might be enriched in specific sequences that can compensate for the lack of a well-defined PY tract. In order to identify such motifs, we first characterized the relative enrichment of all 4-7 nucleotide n-mers in the 50 nucleotide region from -80 to -30 (relative to the splice-junction) for introns with PY tracts categorized as 'weak' relative to the set of all introns (S₆₅ scores less than 0.811; see Materials and methods). We were specifically interested in identifying sequences located in the region upstream of the branchpoint itself. Since most branchpoints are located between -17 and -30 (Figure 1a), the region evaluated would exclude the majority of branchpoint-like sequences.

Human introns have been shown to fall into two classes based upon GC or AT content 50. In order to be sure that we were not merely measuring compositional biases between AT-rich and GC-rich introns, we classified introns according to the GC content of the last 100 bases. Introns with greater than 50% GC content were categorized as GC-rich while those with less than 50% GC were categorized as AT-rich. As measured using our criteria, 37% of AT-rich introns were found to have 'weak' PY tracts, and 72% of GC-rich introns were determined to have 'weak' PY tracts.

Enrichment of n-mers in the -80 to -30 region for introns with weak PY tracts versus all GC or AT-rich introns was determined (see Materials and methods). The entire list of enriched n-mers used in this study is available in Additional data files 2 and 3. According to this analysis, 99 n-mers were determined to be significantly enriched (P < 0.01) in the AT-rich class, and 349 n-mers were determined to be significantly enriched in the GC-rich class. For comparison, we drew random samples of the same size as the corresponding weak PY tract class for both the AT-rich and GC-rich introns, and determined enrichment using the same method as above. The average number of n-mers (for to seven nucleotides) that were determined to be significantly enriched in the randomly drawn samples was ten for the AT-rich and zero for the GC-rich class. Therefore, the enrichment measured appears to be strongly correlated with the composition of the PY tract as measured by the S₆₅ score.

It has been proposed that signals that govern splicing of shorter (<200 nucleotides) introns may differ from those governing splicing of longer introns 51. Therefore, we also evaluated short (<200 nucleotides) and long (≥ 200 nucleotides) AT-rich and GC-rich introns as independent classes. We found that enrichment was similar for both short and long GC-rich introns as evidenced by the observation that the enrichment score for n-mers correlated between these groups (Additional data file 6a). Meanwhile, little correlation was seen between the enrichment scores for long versus short AT-rich introns (Additional data file 6b). This is likely due to the fact that few n-mers were actually determined to be significantly enriched in the short AT-rich population (Additional data file 6b, and data not shown). Together, these data suggest that the compositional biases seen in the region upstream of the PY tract correlate with the potential for U2AF65 binding, especially for GC-rich introns, and that the bias is similar for both long and short introns.

To determine motifs, the enriched n-mers were clustered using the graph clustering method and software presented by Voelker and Berglund 52. Clustering of the n-mers derived from the GC-rich introns yielded 25 clusters (Additional data file 4). These were manually separated into eight groups of compositionally similar motifs (Figure 3a). The n-mers derived from the AT-rich introns yielded eight clusters, of which the three most significant are shown in Figure 3b.

Motifs containing three to four contiguous guanidines are greatly enriched upstream of weak PY tracts for both AT-rich and GC-rich introns (Figure 3, motifs GC2-GC8 and AT1-AT2). Similar G-rich motifs have been previously shown to be enriched in this region 4243. G-rich intronic tracts have been shown to play important roles as splicing signals 53545556, and several heterogeneous nuclear ribonucleoproteins (hnRNPs), including hnRNPs A1, A2, F, and H, have been shown to bind G-rich RNA motifs 54575859. The majority of the G-rich motifs appear to contain a common substring of three to four contiguous Gs separated by one to two nucleotides, and the preferred di-nucleotide spacers appear to be CT, CC, and CA.

In addition, we observed that C-rich motifs (containing three to four contiguous cytidines) are enriched upstream of weak GC-rich PY tracts (Figure 3, motif GC1). Using different computational methods, similar C-rich motifs have been predicted to be ISEs 60. Our analysis provides additional evidence suggesting that C-rich motifs, located upstream of the PY tract, may play important roles in splicing.

We also observed that AT-rich introns with weak PY tracts were enriched in motifs similar to a motif recognized by the protein CUG-BP1 (Figure 3, motif AT3) 61. It is interesting that these motifs did not appear in the GC-rich class. This may be due to compositional biases in the GC-rich class that preclude their identification using the computational methods that we employed, or it may imply that these motifs are, in fact, more abundantly represented in the AT-rich class.

These analyses demonstrate that certain motifs are statistically over-represented upstream of human introns containing weak PY tracts. We also wanted to assess how prevalent these motifs are among introns in general, and also determine the relative level of enrichment between introns with strong versus weak U2AF65 binding sites. Therefore, for each intron, we determined the percentage of the region from -80 to -30 that matched one or more of the n-mers determined to be enriched in introns with weak PY tracts relative to those with strong PY tracts (see above). We refer to this value as the percent coverage. As an example, 80% coverage indicates that 80% of the -80 to -30 region (or 40 of the 50 nucleotides) matches one or more of the enriched n-mers. This analysis (Additional data file 7) revealed that most introns have at least one match to an enriched n-mer. This is not surprising considering that the n-mers are only four to seven nucleotides in length, and, therefore, are expected to occur by chance with fairly high frequency. However, this analysis also revealed that introns with weak PY tracts are likely to have a greater coverage than introns with strong PY tracts. This is especially true of the GC-rich class of introns. For instance, while only 10% of GC-rich introns with strong PY tracts have 80-100% coverage, 23% of introns with weak PY tracts have this level of coverage (Additional data file 7). A smaller difference in coverage is seen between AT-rich introns with strong and weak PY tracts; however, the overall trend is the same (Additional data file 7). In both cases, the enriched n-mers tend to make up a greater portion of the -80 to -30 region for introns with weak PY tracts. Together, these observations indicate that the sequences represented by the enriched n-mers are rather common but they tend to cluster in introns with weak PY tracts.

LCAT intron 4 contains both C-rich and G-rich motifs upstream of the PY tract similar to those we identified computationally that are also highly conserved. The PY tract of LCAT intron 4 is a low-scoring PY tract and is not well conserved. To investigate the role of C-rich and G-rich motifs present in LCAT intron 4, we used a mini-gene system. We created a mini-gene that contains the last 50 nucleotides of LCAT intron 3, LCAT exon 4, LCAT intron 4, LCAT exon 5 and the first 50 nucleotides of LCAT intron 5. We included the downstream and upstream flanking introns in order to allow exon definition to occur, although short introns are often observed to function by intron definition 19.

We examined the role of two G-rich motifs (G-rich motif (GRM)1 and GRM2) present upstream of the PY tract of LCAT intron 4 (Figure 4a). The wild-type (WT) LCAT intron 4 mini-gene splices such that 5 ± 1% pre-mRNA is observed (Figure 4b, lane 1, and 4c). Mutation of GRM1 to AAA (MUT 3, Figure 4a) had a strong effect, and increased the unspliced product to 19 ± 5% (Figure 4b, lane 2, and 4c). Mutation of GRM2 to AAA (MUT 4, Figure 4a) had slightly less of an effect than MUT 3, resulting in 14 ± 3% pre-mRNA (Figure 4b, lane 3, and 4c). Mutation of both GRM1 and GRM2 (MUT 7, Figure 4a) had a similar effect as mutation of GRM1 alone (Figure 4b, lane 4, and 4c), suggesting that the two GRMs do not function additively towards recognition of LCAT intron 4. We also mutated a region that was neither a G-rich motif nor C-rich motif (MUT 5, Figure 4a) to be sure that the AAA motif we were inserting was not acting as an ISS. MUT 5 spliced similarly to WT (Figure 4b, compare lanes 1 and 5; Figure 4c), suggesting that the presence of the mutant AAA sequence in that region of LCAT intron 4 does not act as an ISS. These results suggest that GRM1 and GRM2 are ISEs important for the splicing of LCAT intron 4.

To determine whether the C-rich motifs function as ISEs, we mutated two C-rich motifs: C-rich motif (CRM)1 and CRM2 (Figure 5a), which are present upstream of the PY tract in LCAT intron 4. Mutation of CRM1 to AAA (MUT 1, Figure 5a) did not have a significant effect on splicing (Figure 5b, lane 2, and 5c). We also created a CRM1 mutant where we mutated CCC to AUA (MUT 1b, Figure 5a) and observed the same level of splicing as the AAA mutant (Figure 5b, compare lanes 2 and 3; Figure 5c). Similarly, mutation of CRM2 to AAA (MUT 2, Figure 5a) did not have a significant effect on splicing (Figure 5b, lane 4, and 5c). However, mutation of both CRM1 and CRM2 (MUT 6, Figure 5a) resulted in a decrease in splicing to 19 ± 3% pre-mRNA (Figure 5b, lane 5). These results suggest that while CRM1 and CRM2 do not individually contribute significantly to the splicing of LCAT intron 4, mutation of multiple C-rich motifs has a combinatorial effect.

We hypothesized that the G-rich motifs and C-rich motifs could be functioning together in the recognition of LCAT intron 4. We have observed that there are many examples of introns where the G-rich and C-rich motifs are both present (data not shown). Mutation of both GRM1 and CRM1 (MUT 24, Figure 6a) resulted in a greater decrease in splicing (shown as an increase in percent pre-mRNA) than mutation of either motif alone (Figure 6b, compare MUT 24, lane 5, to MUT 1, lane 2, or MUT 3, lane 3; Figure 6c). An even greater decrease in splicing was observed for the combined mutation of GRM1, CRM1 and CRM2 (MUT 25, Figure 6b, compare MUT 25, lane 6, to MUT 3, lane 3 or MUT 6, lane 4; Figure 6c). These results suggest that the G-rich motifs and C-rich motifs function in combination to promote the splicing of LCAT intron 4.

We examined whether the C-rich motifs could function in the place of the G-rich motifs. Mutation of GRM1 to CCC (MUT 27, Figure 7a) resulted in a smaller decrease in splicing compared to that observed for mutation of GRM1 to AAA (Figure 7b, compare MUT 27, lane 5, to MUT 3, lane 2; Figure 7c). Mutation of GRM1 and GRM2 to C-rich motifs (MUT 28, Figure 7a) also resulted in a smaller decrease in splicing compared to mutating GRM1 and GRM2 to AAA (Figure 7b, compare MUT 28, lane 6, to MUT 7, lane 3). We observed that both the single and double GRM to CRM mutations resulted in similar effects on splicing (Figure 7b, compare MUT 27, lane 5, to MUT 28, lane 6). These results suggest that a C-rich motif can partially compensate for a G-rich motif in this location. Furthermore, it appears that a C-rich motif followed by a G-rich motif (MUT 27) functions as effectively as two C-rich motifs (MUT 28). Mutation of CRM1 and CRM2 to G-rich motifs (MUT 29, Figure 7a) resulted in splicing similar to WT (Figure 7b, compare MUT 29, lane 7, to WT, lane 1; Figure 7c). We conclude that G-rich motifs can fully compensate for, and function in the place of, C-rich motifs, while C-rich motifs can only partially compensate for G-rich motifs.

We next investigated the role of the PY tract in LCAT intron 4 splicing. We mutated the PY tract to determine whether the C-rich sequences in the PY tract were also being recognized. Mutation of a C-rich sequence in the PY tract (CRM3, MUT 16B, Figure 8a) resulted in a minor decrease in splicing (MUT 16B, Figure 8b, lane 9, and 8c), indicating that CRM3 is not singly making a major contribution to the recognition of LCAT intron 4. However, the minor decrease in splicing does suggest that the PY tract may be playing a role. Strengthening the PY tract by mutating the sequence to include a run of eight uridines (MUT 17, Figure 8a) resulted in similar splicing to WT (Figure 8b, compare WT, lane 1, to MUT 17, lane 5). However, in the context of this strengthened PY tract, mutation of CRM1 and CRM2 (MUT 20, Figure 8a) did not result in decreased splicing (Figure 8b, compare MUT 20, lane 6, to MUT 6, lane 2; Figure 8c). Furthermore, the cumulative mutation of GRM1 and CRM1 (MUT 48, Figure 8a) or GRM1, CRM1 and CRM2 (MUT 49, Figure 8a) did not affect splicing in the presence of the strengthened PY tract (Figure 8b, compare MUT 48 to MUT 24 and MUT 49 to MUT 25). This result suggests that, in the context of a strengthened PY tract, the C-rich motifs and G-rich motifs are no longer necessary for recognition, while in the WT context the C-rich motifs and G-rich motifs function as ISEs to compensate for the weak LCAT intron 4 PY tract.

GNPTG intron 2 is an alternatively spliced (intron retention) short intron containing multiple C-rich motifs upstream of a low scoring PY tract (Figure 9a, S₆₅ score = 0.536). In order to test the function of the three C-rich motifs, we created a mini-gene containing exon 2, intron 2 and exon 3. The WT GNPTG intron 2 mini-gene splices such that 29 ± 6% pre-mRNA is observed (Figure 9b,c). Mutation of the three C-rich motifs upstream of the PY tract (Figure 9a) had a significant effect on splicing, resulting in 63 ± 5% pre-mRNA (Figure 9b,c). This result provides an additional example of C-rich motifs functioning as ISEs in an intron containing a weak PY tract.

Our U2AF65 binding studies using various human intron PY tracts (Figure 2) confirm that the computational prediction can generally delineate strong and weak U2AF65 binding sites. Two caveats to our scoring system are: it is based solely on the U2AF65 SELEX data and, therefore, does not take into account nucleotide substitutions that are particularly deleterious for U2AF65 binding; and it cannot account for RNA secondary structure. Each of these parameters can contribute to lower than predicted binding affinities and may partially explain the deviations observed between predicted and observed binding strengths. Nevertheless, the S₆₅ score is generally able to distinguish between sequences displaying strong and weak interactions with U2AF65, and it is more accurate than using simple uridine content alone.

For this analysis we also assume that the PY tract is located in the last 30 nucleotides of the intron. While this is a fair assumption for the vast majority of human introns, there are examples of introns where the PY tract and branchpoint sequence are located a further distance from the 3'ss AG 48626364. Some of the human introns that score as having low scoring PY tracts may actually have high scoring PY tracts that are distally located. Although there are caveats to our scoring system, the S₆₅ score generally distinguishes low and high affinity U2AF65 binding sites, allowing us to ask questions about the population of human introns with low affinity U2AF65 binding sites.

We have identified families of motifs that are over-represented upstream of weak PY tracts but not upstream of strong PY tracts (Figure 3). Our evidence, combined with previous observations, suggests that these motifs function as ISEs that appear to compensate for weakened U2AF65-PY tract interactions. While we chose to focus our attention on the G-rich and C-rich triplet motifs, our study identified at least one additional motif that may represent binding sites for members of the CELF family of proteins. However, additional experimental evidence will need to be obtained to verify the functional significance of the other motifs identified by our study.

The experimental work presented here has focused on two relatively short introns, but our computational analysis found that the same families of motifs were over-represented in both short and long human introns (Additional data file 6). Although LCAT intron 4 is constitutively spliced, expressed sequence tag data suggest that GNPTG intron 2 is alternatively spliced, with some expressed sequence tags containing a retained intron 2. We expect to find examples where these motifs may play important roles in both constitutive and alternative splicing for both short and long introns.

G-rich motifs have been shown to be enriched in short mammalian introns 2065. The G-rich motif GRM1 is the strongest ISE we have observed in LCAT intron 4 (Figure 4). Double mutation of the two sequential G-rich motifs does not result in an additive effect on splicing. G-rich motifs have been shown to function in a combinatorial manner to promote splicing 2056, although the spacing between G-rich motifs was greater (for example, 8-10 nucleotides 56), than in LCAT intron 4, where only a single nucleotide separates the two G-rich motifs. Our studies confirm that G-rich sequences play an important role in promoting the recognition of GC-rich introns with weak PY tracts as previously observed 20.

Our results also show that C-rich motifs can act as ISEs like the G-rich motifs, but that the C-rich motifs may play more of an ancillary role to the G-rich motifs, at least in the case of LCAT intron 4 (Figure 5). C-rich motifs have been shown to function as an ISE in a chicken intron near the 5'ss 66, and as an ISS in a human intron near the 3'ss 67. The single C-rich motif mutational studies presented here suggest that the C-rich motifs present in LCAT intron 4 have little individual effect on LCAT intron 4 splicing. However, we have observed that the C-rich motifs function additively, and that mutating both C-rich motifs (CRM1 and CRM2) is equivalent to mutating the one stronger G-rich motif (GRM1). Furthermore, we show that the mutation of multiple C-rich motifs in GNPTG intron 2 has a significant effect on splicing. This provides an additional example of C-rich motifs functioning as ISEs in an intron with a weak PY tract. C-rich motifs have not been previously shown to function as ISEs in human, nor to function in tandem to produce an additive effect on splicing.

Interestingly, the C-rich and G-rich motifs together function additively (Figure 6). This suggests a model where the combinatorial recognition of two separate ISEs promotes LCAT intron 4 splicing. It is intriguing to consider that interactions between the two ISEs could exist and represent an opportunity for protein-protein interactions between the C-rich and G-rich trans-factors, which could enhance intron recognition to a greater extent than either ISE (or trans-factor) alone. Combinatorial recognition of G-rich repeats with each other has been reported 205556, as has the combinatorial function of a G-rich ISS with upstream ESSs 54. Here we show that G-rich motifs can function in conjunction with C-rich ISEs to promote splicing, showing the flexibility of the G-rich motifs to function in different contexts.

We have also observed that the C-rich motifs can only partially compensate in the place of G-rich motifs in LCAT intron 4 splicing, while G-rich motifs appear to fully compensate in the place of C-rich motifs (Figure 7). It may be that the C-rich motifs have a positional dependence while the G-rich motifs do not. G-rich motifs appear to be the dominant ISE in LCAT intron 4 splicing. G-rich motifs appear to be capable of alleviating the need for C-rich motifs. The observation that C-rich motifs only partially rescue splicing reinforces the model that the C-rich motifs do not have the same enhancer strength as the G-rich motifs.

An examination of the primary sequence and predicted secondary structure (using mfold 68) of LCAT intron 4 suggests that the intron could be folding into a stem-loop structure with the G-rich and C-rich sequences base-pairing (data not shown). While this is an intriguing model, when the C-rich motifs are replaced with G-rich motifs (a mutation that would abolish stem-loop formation; MUT 29, Figure 7), we observe splicing similar to WT levels, suggesting that a stem-loop structure is not contributing to the splicing of LCAT intron 4.

There are multiple candidate proteins that could be recognizing the G-rich and C-rich motifs present in LCAT intron 4. A G-rich motif trans-factor, hnRNP H, has been identified and shown to bind G-rich sequences and regulate splicing both positively and negatively 545556. Several additional hnRNP proteins, including hnRNPs A1, A2, and F, have also been shown to bind G-rich RNA sequences 54575859. An alternative model for G-rich sequence recognition involves RNA-RNA interactions to promote U1 snRNP binding. G-triplets near the 5'ss have been shown to bind the U1 snRNP by interacting with the U1 snRNA and this interaction was shown to be important for human alpha globin splicing in vivo 69.

hnRNP K and the α-CP proteins are the major poly-C-binding proteins identified in mammalian cells 7071. Both hnRNP K and several α-CP isoforms have been implicated in post-transcriptional control 70. There have also been two studies that have implicated these proteins in splicing. hnRNP K was shown to enhance the splicing of a chicken b-Tropomyosin intron by binding a C-rich motif near the 5'ss 66. A recent study has shown that α-CP2 binds a C-rich patch upstream of a weak PY tract in the human α-globin intron 1 transcript and inhibits splicing of this intron in vitro 67. This is in contrast to our results with LCAT intron 4 where the C-rich motifs function as splicing enhancers, not silencers. Several cis-elements, including G-rich ISEs, have been shown to act as both splicing enhancers and silencers 54555672. The C-rich motifs and their trans-factor may also possess the flexibility to function as silencers and enhancers.

Our results suggest a model where ISEs present upstream of a weak PY tract compensate for a weakened U2AF65-RNA interaction (Figure 10). In the case of LCAT intron 4, the G-rich and C-rich motifs and the branchpoint sequence are highly conserved and yet the PY tract is not well conserved. It is possible that the presence of strong enhancers upstream of the PY tract has allowed for greater degeneracy in the PY tract region. In support of this model we have observed that when the PY tract is strengthened to include a run of eight uridines, mutation of both C-rich motifs or the cumulative mutation of the G-rich and C-rich motifs no longer have an effect on LCAT intron 4 splicing (Figure 8). The G-rich and C-rich motifs appear dispensable in the presence of a strong PY tract. G-rich motifs have previously been shown to be dispensable for maximal splicing in the presence of a strengthened PY tract 20. These results suggest that strong U2AF65-PY tract interactions alleviate the role of upstream ISEs.

An alternative model to weakened U2AF65-RNA interactions is that a splicing factor other than U2AF65 binds the weakened PY tracts. Recent work has shown that PUF60 plays a role in splicing by interacting with the PY tract 29. Our observation that many of the weak PY tracts are particularly C-rich leads us to propose that a C-rich binding protein may function in this region. When we mutated a C-rich motif in the LCAT intron 4 PY tract we observed a small effect on splicing, suggesting that the C-rich motifs in the PY tract itself are recognized by a trans-factor (Figure 8). It is possible that such a splicing factor, be it U2AF65 or another protein, could be functioning in conjunction with the factor(s) that recognize the G-rich and C-rich ISEs upstream of the PY tract.

The pool of introns containing low-scoring U2AF65 binding sites represents a significant class of human introns lacking a canonical splicing element. The ISEs identified and validated here suggest that novel mechanisms exist in the cell for coping with weakened U2AF65-RNA interactions. Specifically, we have observed that the interplay of multiple cis-elements, in this case the G-rich and C-rich motifs, appears to be crucial for the recognition of non-canonical introns. In the future we plan to explore additional ISEs identified by this study to gain a broader picture of how the splicing machinery functions in the recognition of introns with weak PY tracts. While we have focused our attention here on a single key canonical splicing element, the PY tract, we plan on extending our analyses and expect to find that similar strategies exist for the recognition of other classes of non-canonical pre-mRNAs.

The March 2006 human reference sequence (NCBI Build 36.1) in conjunction with the UCSC KnownGenes (hg18) annotation database (Release 8 April 2007) 73 was used to create a non-redundant database of human intronic sequences. After excluding annotated introns that did not begin with G [T/C] [A/G] and end with AG, and were less than 60 bases in length, we were left with 171,475 unique acceptor ends.

In order to score PY tracts according to their likely affinity towards U2AF65, we developed a score that reflects the level of similarity of the PY tract sequence to the sequences that were enriched in RNAs derived from in vitro SELEX experiments using human U2AF65. In particular, if we let the frequency of occurrence of an n-mer n of length k within the SELEX sequences be represented by f_n, then for a subject sequence of length L the S₆₅ score is determined according to:

S 65 = ∑ i = 0 i = L − k + 1 ln ⁡ ( f n i 1 4 k ) L − k + 1 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciGacaGaaeqabaqabeGadaaakeaacaWGtbWaaSbaaSqaaiaaiAdacaaI1aaabeaakiabg2da9KqbaoaalaaabaWaaabCaeaaciGGSbGaaiOBamaabmaabaWaaSaaaeaacaWGMbWaaSbaaeaacaWGUbWaaSbaaeaacaWGPbaabeaaaeqaaaqaamaaliaabaGaaGymaaqaaiaaisdadaahaaqabeaacaWGRbaaaaaaaaaacaGLOaGaayzkaaaabaGaamyAaiabg2da9iaaicdaaeaacaWGPbGaeyypa0JaamitaiabgkHiTiaadUgacqGHRaWkcaaIXaaacqGHris5aaqaaiaadYeacqGHsislcaWGRbGaey4kaSIaaGymaaaaaaa@4D00@

where f_n represents the frequency (within the SELEX population) of the n-mer found at position i in the subject sequence.

The term ln⁡(fni14k) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8viVeY=Nipec8Eeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaciiBaiaac6gadaqadaqcfayaamaalaaabaGaamOzamaaBaaabaGaamOBamaaBaaabaGaamyAaaqabaaabeaaaeaadaWccaqaaiaaigdaaeaacaaI0aWaaWbaaeqabaGaam4AaaaaaaaaaaGccaGLOaGaayzkaaaaaa@3963@ is a log-odds representation of the degree to which the particular n-mer was enriched within the SELEX sequences. Since the SELEX experiment began with uniformly random sequences, the denominator is simply the expectation for random occurrence of an n-mer of length k. For this study we chose the n-mer length to be five and the SELEX data were those reported in Singh et al. 27 and both SELEX experiments reported in Banerjee et al. 39. The frequency of occurrence for all pentamers within these sequences is shown in Additional data file 1. Introns with 'strong' PY tracts (that is, expected to have high affinity for U2AF65) were defined to be those that are above the median value for all introns (0.811). All but one of the RNAs derived from in vitro SELEX had S₆₅ scores above this value.

In order to avoid biases due to long interspersed repetitive elements (LINEs) and short interspersed repetitive elements (SINEs), repetitive elements in the intronic sequence database (obtained as described above) were masked using the masking coordinates associated with the UCSC hg18 annotation database (Release 8 April 2007) 73. However, simple repeats (many of which resemble known hnRNP binding sites) were not masked. The intronic acceptor sequences were then separated according to their GC content within the last 100 bases (or last half if the intron was less than 200 bases in length). AT-rich introns were defined to be introns containing less than 50% GC content. GC-rich introns were defined to be those containing greater than or equal to 50% GC content.

For each of these data sets, the occurrence of all n-mers (4-7 nucleotides) in the 50 nucleotide region from -80 to -30 (relative to the acceptor splice-junction) were determined using a sliding window. These counts were used to determine the background expectations for each n-mer. The occurrence of each 4-7 nucleotide n-mer within the equivalent region for all introns possessing 'weak' PY tracts (defined as above) was determined using a sliding window. From these values, n-mers that are enriched upstream of the branchpoint region for introns possessing weak PY tracts was determined using the binomial confidence interval method described in Voelker and Berglund 52. For the AT-rich class, 99 n-mers were determined to be significantly enriched (P < 0.01), and 349 n-mers were determined to be significantly enriched for the GC-rich class. Enriched n-mers and corresponding counts and statistics are available in Additional data files 2 and 3. Enriched n-mers were used to construct motifs as in Voelker and Berglund 52. All of the derived motifs and the identities and occurrences of all n-mers that were used to construct the motifs are available in Additional data files 4 and 5.

RNA oligonucleotides (listed in Figure 2b, IDT, Integrated DNA Technologies, San Diego, CA, USA) for U2AF65 binding assays were 5' end-labeled with γ-³²P ATP using T4 polynucleotide kinase (NEB, Ipswich, MA, USA) for 30 minutes at 37°C. The RNAs were then gel purified using an 8% denaturing gel, eluted from the gel in 0.3M Na acetate and ethanol precipitated. The resulting pellet was resuspended in nanopure water and purified with a Bio-spin 6 column (BioRad, Hercules, CA, USA) equilibrated with nanopure water. The radioactivity level of the purified RNA solution was determined by scintillation. Gel-shift binding assays were performed using varying concentrations of recombinant human U2AF65 with constant amounts of radiolabeled RNA oligonucleotides as previously described 49. The Ensembl gene accession numbers for the genes addressed in this study are: BRUNOL4 [ENSEMBL: ENSG00000101489], INSR [ENSEMBL: ENSG00000171105], LCAT [ENSEMBL: ENSG00000124067], MBNL1 [ENSEMBL: ENSG00000152601], SR140 [ENSEMBL: ENSG00000163714], and U2AF2 [ENSEMBL: ENSG00000063244].

WT LCAT intron 4 mini-gene was cloned from HeLa genomic DNA using primers to amplify the region between the last 50 nucleotides of LCAT intron 3 to the first 50 nucleotides of LCAT intron 5 (502 nucleotides). The forward primer included a BamH1 site and the reverse primer included an EcoR1 site. The amplified genomic DNA was cut with BamH1 and EcoR1, inserted into pcDNA3 and sequenced. LCAT intron 4 mutants were made by PCR using the WT LCAT 4 mini-gene as template and primers containing the mutation of interest. LCAT intron 4 mutants were also cloned into pcDNA3 using BamH1 and EcoR1 and sequenced. The WT and mutant GNPTG [ENSEMBL: ENSG00000090581] intron 2 mini-genes were cloned using overlapping primers to create a sequence containing exon 2, intron 2 and exon 3. This sequence was flanked by cut sites HindIII and Not1, cloned into pcDNA3 and sequenced.

HeLa cells were grown in monolayers in DMEM with GLUTAMAX (Invitrogen, Carlsbad, CA, USA) and supplemented with 10% fetal bovine serum (GIBCO). For the LCAT splicing experiments 1.5 (± 0.2) × 10⁵ cells were plated in 6-well plates and transfected 18-20 h later at approximately 70% confluency. Plasmid (1 μg) was transfected into each well of cells using 5 μl of Lipofectin (Invitrogen, Carlsbad, CA, USA) and 10 μl of Plus reagent (Invitrogen) according to the manufacturer's protocols. For the GNPTG splicing experiments, 2 × 10⁵ cells were plated in 6-well plates and transfected with 1 μg plasmid 18-20 h later using 5 μl of Lipofectamine 2000. Cells were harvested 24 h (LCAT experiments) or 16 h (GNPTG experiments) after transfection using TriplE (GIBCO) and then pelleted by centrifugation. RNA was isolated from the cell pellets using an RNeasy kit (QIAGEN, Valencia, CA, USA).

Isolated RNA (500 ng) was incubated with 1 unit of RQI DNase (Promega, Madison, WI, USA) in a 10 μl reaction for 2 h (LCAT experiments) or 1 h (GNPTG experiments) according to the manufacturer's protocol. DNAsed RNA (2 μl (100 ng)) was reverse transcribed in a 10 μl reaction (1:5 dilution) using Superscript II and an LCAT-specific reverse primer or a reverse primer to the pCDNA3 SP6 sequence for the GNPTG experiments, according to manufacturer's protocols with the exception that we used half the recommended amount of Superscript II (Invitrogen, Carlsbad, CA, USA). For the LCAT splicing experiments, 2 μl of the reverse transcription reaction was subjected to 20 rounds of PCR amplification in a 20 μl reaction (1:10 dilution) using LCAT specific primers spiked with a kinased LCAT forward primer (0.4 nM). Twenty rounds of PCR were found to be within the linear range for this PCR experiment (data not shown). The resulting PCR products were run on an 8% (19:1) polyacrylamide native gel. For the GNPTG splicing experiments, 2 μl of the reverse transcription reaction was subjected to 27 rounds of PCR amplification in a 20 μl reaction (1:10 dilution) using primers specific to the T7 (forward) and SP6 (reverse) sequences of the pcDNA3 plasmid spiked with kinased T7 forward primer. Twenty-seven rounds of PCR were found to be within the linear range for this PCR experiment (data not shown). The resulting PCR products were run on a 10% (19:1) polyacrylamide gel. The gels were dried and exposed overnight to a phosphorimager screen. Quantification of the radioactive bands was performed using ImageQuant software (GE Healthcare, London, UK). The percent pre-mRNA was calculated by dividing the amount of the pre-mRNA band by the total amount of the pre-mRNA and mRNA bands and multiplying by 100%.