In vitro and in silico analysis reveals an efficient algorithm to predict the splicing consequences of mutations at the 5′ splice sites

Abstract

We have found that two previously reported exonic mutations in the PINK1 and PARK7 genes affect pre-mRNA splicing. To develop an algorithm to predict underestimated splicing consequences of exonic mutations at the 5′ splice site, we constructed and analyzed 31 minigenes carrying exonic splicing mutations and their derivatives. We also examined 189 249 U2-dependent 5′ splice sites of the entire human genome and found that a new variable, the SD-Score, which represents a common logarithm of the frequency of a specific 5′ splice site, efficiently predicts the splicing consequences of these minigenes. We also employed the information contents (R_i) to improve the prediction accuracy. We validated our algorithm by analyzing 32 additional minigenes as well as 179 previously reported splicing mutations. The SD-Score algorithm predicted aberrant splicings in 198 of 204 sites (sensitivity = 97.1%) and normal splicings in 36 of 38 sites (specificity = 94.7%). Simulation of all possible exonic mutations at positions −3, −2 and −1 of the 189 249 sites predicts that 37.8, 88.8 and 96.8% of these mutations would affect pre-mRNA splicing, respectively. We propose that the SD-Score algorithm is a practical tool to predict splicing consequences of mutations affecting the 5′ splice site.

INTRODUCTION

In eukaryotes, splicing of the nuclear mRNA precursor (pre-mRNA) takes place mostly within the U2-dependent spliceosome, a complex of five uridine-rich small nuclear (sn) ribonucleoproteins (RNPs): U1, U2, U4, U5 and U6 snRNPs and numerous non-snRNP proteins. In the first step of spliceosome formation, U1 snRNP recognizes the 5′ splice site and regulates initiation of pre-mRNA splicing (1).

The 5′ splice site is composed of the last three nucleotides of an exon (positions −3, −2 and −1) and the first six nucleotides of an intron (positions +1 to +6). The consensus sequence of the U2-dependent 5′ splice sites is (C/A)AG|GT(A/G)AGT (2), where the vertical line (|) represents the exon–intron boundary, and the ‘GT’ dinucleotide at the 5′ end of an intron is invariable (Figure 1A) (3). In the latter stage of pre-mRNA splicing, U1 snRNA dissociates from the 5′ splice site, and U6 snRNA subsequently binds to nucleotides at positions +2, +5 and +6 (Figure 2A) (4–6).

Figure 1.

(A) The 5′ end of U1 snRNA recognizes three nucleotides at exonic positions −3, −2 and −1 and six nucleotides at intronic positions +1 to +6. Note that the consensus sequence, 5′-(C/A)AG|GT(A/G)AGT-3′, is complementary to U1 snRNA at most positions. (B and C) The ‘two-point’ analysis reveals that NCp-nucs at exonic positions −3, −2 and −1 are compensated for by Cp-nucs at positions +4, +5 and +6, and that NCp-nucs at intronic positions +4, +5 and +6 are compensated for by Cp-nucs at positions −3, −2, −1 and +3. For example, when a complementary C is used at position −3 (68 353 sites), the frequency of a complementary T at position +6 is 42.9% (29 307 of 68 353). In contrast, when a noncomplementary T is used at position −3 (22 667 sites), the frequency of a complementary T at position +6 is increased to 60.2% (13 636 of 22 667). The concordance ratio is calculated as (60.2–42.9)/42.9 = 0.403 (arrow). This means that when position −3 is a noncomplementary T, we observe a complementary T at position +6 40.3% more frequently than when position −3 is a complementary C. A positive concordance ratio at a specific position indicates that a Cp-nuc to U1 snRNA is preferentially used to compensate for an NCp-nuc at another position.

Figure 2.

(A) The U6 snRNA base pairs with nucleotides at positions +2, +5 and +6. (B) RT–PCR analysis of minigene constructs transfected into HEK293 cells with artificial U1 or U6 snRNA. A single Cp-nuc is introduced into U1 or U6 snRNA while retaining the mismatch at the mutation. Wt, wild-type construct; Mt, mutation observed in a patient. For example, +6U1 indicates that a nucleotide on U1 snRNA corresponding to position +6 is substituted to match the 5′ splice site. Circles and squares represent nucleotides that become complementary to the artificial U1 and U6 snRNAs, respectively. The open arrowhead indicates a normally spliced fragment, whereas the closed arrowhead indicates an exon-skipped fragment. The rightmost column shows the densitometric ratio of the exon-skipped fragment. The asterisk indicates a mixture of fragments due to activated 5′ splice sites four and 13 nucleotides downstream of the native 5′ splice site at the 5′ exon of pSPL3. The dagger indicates a heteroduplex formed by normally spliced and exon-skipped products. Uppercase nucleotides are complementary to U1 snRNA, whereas lowercase nucleotides are not.

In the course of identification of exonic splicing mutations in genetic forms of Parkinson's disease, we found two splicing mutations at the 5′ splice site that compromise binding to U1 snRNA. To clarify how exonic mutations at the 5′ splice site cause aberrant splicing, we analyzed 31 minigenes in vitro and examined 189 249 putative U2-dependent 5′ GT splice sites of the entire human genome in silico. We found that a new variable, the SD-Score, in combination with the information contents (R_i), which represents the amount of information in bits (7,8), can efficiently predict the splicing consequences of exonic mutations at the 5′ splice site. We validated our algorithm with 32 additional minigenes and with 179 previously reported splicing mutations, and found that the SD-Score algorithm has a sensitivity of 97.1% and a specificity of 94.7%. We believe that the SD-Score algorithm is a practical tool for predicting the splicing consequences at the 5′ splice site of mutations causing human disease.

MATERIALS AND METHODS

Exon trapping vector

To examine splicing consequences of exonic mutations, we constructed minigenes in an exon-trapping vector, pSPL3 (a discontinued product of Invitrogen, Carlsbad, CA, USA), which was kindly provided by Dr Kazunori Imaizumi, Department of Anatomy, University of Miyazaki, Miyazaki, Japan. We introduced a cytomegalovirus promoter in place of the simian virus 40 promoter by means of a megaprimer-based, site-directed mutagenesis method (9) and also introduced a PacI recognition sequence at the multiple cloning site. Because the nonsense-mediated mRNA decay can degrade a specific splicing product with a premature termination codon and cause misinterpretation of splicing assays (10), we eliminated a premature stop codon and constructed pSPL3 vectors with three reading frames to insert any chimeric exons in-frame (Supplementary Figure 1).

Minigene constructs and mutagenesis

We used the polymerase chain reaction (PCR) to amplify an exon and the flanking introns of 200 bp of the genes of our interest using normal human genomic DNA extracted from HEK293 cells. NotI and PacI recognition sites were introduced to the 5′ and 3′ ends, respectively, of the PCR product. Each amplicon was inserted into one of the three pSPL3 vectors so that the reading frame of the chimeric exon was retained.

For the PARK7 and DYSF genes, wild-type pSPL3 constructs yielded a large proportion of exon-skipped products. We thus amplified a genomic segment spanning the mutation-harboring exon, the flanking introns and the neighboring exons, and then inserted the amplicon into the pcDNA3.1(+) mammalian expression vector (Invitrogen). Different splicing consequences between pSPL3 and pcDNA3.1(+) constructs likely represent the complexity of splicing analysis.

The U1 snRNA gene with its own promoter was kindly provided by Dr Alan M. Weiner, Department of Biochemistry, University of Washington, Seattle, WA, USA. The U6 snRNA gene with the 5′ promoter region of 367 bp and a 3′ end region of 149 bp was amplified using normal human genomic DNA extracted from HEK293 cells and was inserted into the pGEM-T Easy Vector (Promega, Madison, WI, USA).

Naturally occurring and artificial mutations were introduced into the inserts with the QuikChange Site-Directed Mutagenesis Kit (Stratagene, La Jolla, CA, USA). We confirmed by sequencing that there were no artifacts in any insert.

Transfection and RNA analysis

HEK293 cells were maintained in the Dulbecco's minimum essential medium (DMEM; Sigma–Aldrich, St Louis, MO, USA) with 10% fetal bovine serum (FBS; Sigma–Aldrich). At ∼50% confluency (∼5 × 10⁵ cells) in a 6-well plate, 1 ml of fresh Opti-MEM I (Invitrogen) was substituted for DMEM, and 1 µg of a minigene with 3 µl of the FuGENE6 Transfection Reagent (Roche Diagnostics, Indianapolis, IN, USA) were then added. After 4 h, 2 ml of DMEM with 10% FBS was overlaid, and the cells were incubated overnight. The transfection medium was replaced with 2 ml of fresh DMEM with 10% FBS, and the transfected cells were incubated for 48 h before RNA extraction. When artificial U1 or U6 snRNA vector was used, 50 ng of a minigene and 950 ng of each snRNA vector were introduced. Total RNA was extracted using the GenElute Mammalian Total RNA Kit (Sigma–Aldrich). DNA was degraded on-column with the DNase I (Qiagen, Valencia, CA, USA). Twenty percent of the isolated RNA was used as a template for cDNA synthesis with the Oligo(dT)^12–18 primer (Invitrogen) and the SuperScript II Reverse Transcriptase (Invitrogen). Ten percent of the synthesized cDNA was used as a template for reverse transcriptase (RT)–PCR amplification with primers SD6 (5′-TCTGAGTCACCTGGACAACC-3′) and SA2 (5′-GTGAACTGCACTGTGACAAGCTGC-3′), both of which were on the pSPL3 vector. For minigenes in pcDNA3.1(+), gene-specific primers were employed. Amplification was performed for 30 to 35 cycles of denaturation at 94°C for 20 s, annealing at 52°C for 20 s and extension at 72°C for 45 s. We measured the signal intensities of the normal and aberrant fragments with the NIH Image 1.63 program. When the ratio of the aberrant product of the mutant construct was increased by 2.5-fold compared with that of the wild-type construct, we considered the mutant construct to have resulted in aberrant splicing. We tried several different thresholds and found that the threshold of 2.5-fold best represents the results of our visual inspections (data not shown).

For the PINK1, CHM, BRCA1, DYSF, F8 and DMD genes, we cloned and sequenced all RT–PCR fragments to confirm that the expected normal and aberrant splicings indeed had taken place in these minigenes.

In silico analysis

We extracted all the nonredundant 5′ GT splice sites in the entire human genome using the CDS tags in the NCBI RefSeq Database Build 36.2. Each 5′ splice site on the genome is counted once, even if it is used multiple times in alternatively spliced transcripts. The analysis was performed with the PrimePower HPC2500/Solaris 9 supercomputer (Fujitsu Ltd., Tokyo, Japan). Using the JMP-IN Ver. 5.1.2 software (SAS Institute, Cary, NC, USA), we statistically determined a threshold for each variable using the default settings.

In humans, ∼0.1–0.3% of introns are spliced by the minor U12-dependent spliceosome (2,11,12), and ∼70% of the U12-dependent introns have GT-AG terminal dinucleotides (13). Previous in silico analyses of the human genome identified 275 (12), 469 (14) and 487 (13) GT-AG U12-dependent introns. We thus eliminated 487 U12-dependent 5′ GT splice sites from our analysis, according to the U12 Intron Database (http://genome.imim.es/cgi-bin/u12db/u12db.cgi). Our training and validation data sets (see ‘Results’ section) did not include any of the known U12-dependent splice sites.

RESULTS

Screening of exonic splicing mutations in genes causing Parkinson's disease

To identify exonic splicing mutations in genetic forms of Parkinson's disease, we analyzed 57 missense, nonsense and synonymous mutations deposited in the Human Gene Mutation Database (http://www.hgmd.cf.ac.uk/) (15) in the SNCA, PARK2, PINK1 and PARK7 genes using minigenes (Supplementary Table 1). We found that no mutation affected an exonic splicing enhancer (ESE) or silencer (ESS). Two mutations at the 5′ splice site, however, resulted in skipping of the mutation-harboring exon likely by compromising binding to U1 snRNA. One mutation was E417G in PINK1 (16) and the other mutation was E64D in PARK7 (17). To understand how complementary nucleotides (Cp-nucs) to U1 snRNA at the 5′ splice site compensate for noncomplementary nucleotides (NCp-nucs) at other positions, we introduced a single Cp-nuc to the mutant PINK1 and PARK7 minigenes while retaining the mutation (Supplementary Figure 2A and Supplementary Table 2). We employed these results to develop an algorithm to predict splicing consequences of mutations at the 5′ splice site.

Recapitulation of aberrant splicings due to previously reported exonic splicing mutations

We next recapitulated aberrant splicings of six previously reported exonic splicing mutations at position −2 in the CHM, FAH, HMBS, UROS, BRCA1 and CYP27A1 genes (Supplementary Figure 2B and Supplementary Table 2). All mutations except for CYP27A1 caused the mutation-harboring exon to be skipped. The CYP27A1 mutation is exceptional because it introduces a Cp-nuc rather than disrupting complementarity (18). We similarly introduced a Cp-nuc to the mutant constructs while retaining the mutation. These results were also used for developing a prediction algorithm of splicing mutations.

Site-directed mutagenesis of a single nucleotide of U1 snRNA and U6 snRNA

Because the 5′ splice site is recognized by both U1 and U6 snRNAs at different stages of pre-mRNA splicing (Figures 1A and 2A), we wondered which nucleotide of the 5′ splice site is most important for binding each snRNA. To this end, we introduced a single Cp-nuc to U1 or U6 snRNA while retaining the mismatch between U1 or U6 snRNA and the mutation.

Among 11 experiments with artificial U1 snRNAs, corrections of U1 snRNA corresponding to position +6 in FAH and to +5 in CHM ameliorated aberrant splicings, while the others were inefficient (Figure 2B). Among four experiments with artificial U6 snRNAs, only a correction corresponding to position +5 in CHM partially normalized aberrant splicing (Figure 2B). In contrast to manipulation of the splicing cis-elements, introduced artificial snRNAs are competed by endogenous snRNAs, and hence their effects tend to be compromised. In addition, substitution of these nucleotides might have modified the core secondary structure of U1 or U6 snRNA and made it nonfunctional. Nevertheless, it is interesting to note that corrections of U1 and U6 snRNAs ameliorate aberrant splicings in some mutants even in the presence of mismatch at the mutation. To our knowledge, no similar study has been performed in this scale (19), but the study size was still too small to draw a definite conclusion.

In silico analysis of human 5′ splice sites: the consensus sequence

To examine how the human 5′ splice sites are organized and why the identified exonic mutations resulted in aberrant splicings, we analyzed the 5′ splice sites of the entire human genome. According to the NCBI RefSeq Database, the human genome comprises 28 714 annotated genes with 192 643 5′ splice sites. Of these sites, 189 718 (98.5%) sites carry an invariant GT dinucleotide at positions +1 and +2, whereas 1859 (1.0%) and 311 (0.2%) sites have GC and AT dinucleotides, respectively. The remaining 755 (0.4%) sites carry other dinucleotides and likely include erroneous annotations. We excluded 487 U12-dependent 5′ GT splice sites (see ‘Materials and Methods’ section) and extracted nine nucleotide segments spanning positions −3 and +6 from the remaining 189 249 5′ GT splice sites.

Analysis of nucleotide frequencies at each position showed that the ‘winner sequence’ comprising the most frequently used nucleotides was CAG|GTAAGT (Table 1), which is entirely complementary to U1 snRNA (Figure 1A). The frequency of Cp-nuc was highest at position −1, followed in descending order by positions +5, +4, −2, +3, +6 and −3 (excluding positions +1 and +2, which are invariant in our analysis).

Table 1.

Nucleotide frequencies (%) at U2-dependent 189 249 human 5′ GT splice sites

Position	−3	−2	−1	+1	+2	+3	+4	+5	+6
A	33.4	63.5	10.0			59.5	69.4	8.9	17.9
C	36.1	10.9	2.8			2.9	7.7	5.7	15.0
G	18.5	11.6	80.3	100.0		34.6	11.8	77.6	19.4
T	12.0	14.0	6.8		100.0	3.0	11.1	7.8	47.7
Consensus sequence	C/A	A	G	G	T	A/G	A	G	T

Nucleotides that are complementary to U1 snRNA are underlined. In this study, we calculated the CV (21) with the equation , where F(n, l) is a ratio of a nucleotide ‘n’ at position ‘l’. Similarly, we calculated the R_i (7,8) with the equation . We ignored the error function of R_i, because F(n, l) values are calculated using a large number of observations (189 249 sites), and hence the contribution of the error function should be negligible.

In silico analysis of human 5′ splice sites: the ‘two-point’ analysis

We next analyzed how an NCp-nuc to U1 snRNA at a specific position is compensated for by a Cp-nuc at the other positions. The ‘two-point’ analysis revealed that NCp-nucs at positions +4, +5 and +6 are compensated for by Cp-nucs at positions −3, −2, −1 and +3 (Figure 1C). Conversely, NCp-nucs at positions −3, −2 and −1 are associated with high concordance ratios at positions +4, +5 and +6 (Figure 1B). These results suggest that a stretch of Cp-nucs either in an exonic or an intronic region is essential for proper splicing, which also conforms to the notion that consecutive base pairings with U1 snRNA contribute to recognition of the 5′ splice site (20). It is also interesting to note that a Cp-nuc at position +6 most frequently compensates for an NCp-nucs at position −3, −2 and −1, although the frequency of a Cp-nuc at position +6 is only 47.7% in the human genome.

In our analysis, we assumed that only A at position +3 is a Cp-nuc. When we regarded both A and G as Cp-nucs at position +3, the concordance ratio at position +3 became always low (Supplementary Figure 3), likely because A or G is observed at position +3 in 94.1% of the human 5′ splice sites. This implies that the concordance ratio is less informative, when the frequency of a Cp-nuc is high.

In silico analysis of human 5′ splice sites: the SD-Score

The ‘two-point’ analysis disclosed interdependence between two nucleotides at the 5′ splice sites. However, we could not develop a scoring system using the ‘two-point’ analysis. We thus sought another quantitative measure to predict splicing consequences. We expected that the frequency of a specific 5′ splice site sequence in the human genome would represent the splicing signal intensity that we hoped to score. We analyzed the entire human genome and determined the frequency of each U2-dependent GT splice site sequence. The common logarithm of the frequency was calculated to give a new variable, the SD-Score (Supplementary Table 5). For example, the SD-Score for CAG|GTGAGG, which was observed at 2562 sites, is log (2562/189 249) = −1.868. The SD-Score of a splice site sequence that never appears in the human genome should be log (0/189 249) = −∞ but is defined as log (0.25/189 249) = −5.879 to simplify calculations. The correlation coefficients of the SD-Score with the R_i (7,8) and the consensus value (CV), which represents the similarity of a splice site sequence to the consensus splice site sequence (21), are 0.678 and 0.694, respectively, indicating that the SD-Score is similar to, but distinct from, the other variables.

Prediction of splicing consequences using the SD-Score algorithm

We next examined whether the SD-Score is indeed an effective scoring variable. To this end, we plotted the SD-Scores and the ΔSD-Scores of the 31 constructs of the PARK7, PINK1, CHM, FAH, HMBS, UROS, BRCA1 and CYP27A1 genes (Supplementary Table 2). The ΔSD-Score is calculated by subtracting the wild-type SD-Score from the mutant SD-Score. We found that a 5′ splice site with ΔSD-Score > −0.34 does not affect pre-mRNA splicing in 13 out of 13 sites, whereas a mutant site with ΔSD-Score < −0.34 and SD-Score < −2.9 causes aberrant splicing in 9 out of 10 sites. The 5′ splice sites with ΔSD-Score < −0.34 and SD-Score > −2.9 include a mixed population of three normal and five aberrant splicings. We thus employed the ΔR_i value, which is calculated by subtracting the wild-type R_i from the mutant R_i, and found that three out of three sites with ΔR_i > −1.45 are normally spliced, whereas five out of five sites with ΔR_i < −1.45 are aberrantly spliced. Therefore, these thresholds efficiently predict splicing consequences of our 31 minigene constructs (Figure 3).

Figure 3.

The diagram demonstrates the SD-Score algorithm to predict aberrant splicings due to mutations at the 5′ splice site. The algorithm is based on a training dataset of 31 minigenes and was validated with testing data sets of 32 additional minigenes and 179 naturally occurring splicing mutations (Supplementary Tables 2–4).

Previously unrecognized exonic splicing mutations at positions −2 and −1

To validate the SD-Score algorithm, we employed previously unrecognized splicing mutations at positions −2 and −1. To this end, we scrutinized 2477 exonic mutations from the Human Gene Mutation Database, and searched for mutations at positions −2 and −1. We then randomly selected three mutations in the DYSF, F8 and ABCD1 genes (Supplementary Table 2) whose splicing consequences have not been characterized. The SD-Score algorithm predicted that all the mutations would affect pre-mRNA splicings, and minigene analyses confirmed it (Supplementary Figure 2C and Supplementary Table 2).

We further introduced a single Cp-nuc to each mutant minigene while retaining the mutation. We thus constructed seven artificial minigenes, and six of these were spliced as predicted (Supplementary Figure 2C and Supplementary Table 2).

Previously unrecognized splicing mutations at position −3

We next sought exonic splicing mutations at position −3, which has been reported only in two mutations (22,23) according to the Human 5′ Splice Site Database (http://www.uni-duesseldorf.de/rna/). In the 2477 exonic mutations described above, we identified six mutations that disrupt a complementary C nucleotide at position −3 (Supplementary Table 2). The SD-Score algorithm predicted that four mutations in the GLA, DMD and PARK2 genes would affect pre-mRNA splicing, whereas two mutations in the ABCD1 and NPC1 genes would not. We analyzed six mutant minigenes and found that all, except the PARK2 mutant, were spliced as predicted (Supplementary Figures 2D and 5, and Supplementary Table 2). We confirmed in patient's lymphocytes that a C-to-T mutation at position −3 in DMD exon 5 indeed caused the same aberrant splicing as we observed with the minigene (data not shown). The aberrant splicing due to a mutation in DMD exon 5, however, was likely successfully predicted by the SD-Score algorithm, because additional mutagenesis at position −4, which did not create a novel cryptic site, failed to show exon skipping (Supplementary Figure 5).

We also constructed seven artificial minigenes, and the SD-Score algorithm successfully predicted the splicing consequences of all the minigenes (Supplementary Figure 2D and Supplementary Table 2).

Splicing mutations in the literature database

To further validate the SD-Score algorithm, we employed other exonic and intronic splicing mutations in the literature database (Supplementary Tables 3 and 4). We randomly examined 2, 9, 26, 45, 3, 83 and 11 splicing mutations at positions −3, −2, −1, +3, +4, +5 and +6, respectively. Our algorithm correctly predicted aberrant splicings in 174 of the 179 reported mutations and falsely predicted normal splicings in five mutations.

DISCUSSION

Clinical implications of exonic splicing mutations

Although our analysis failed to detect ESE- and ESS-affecting mutations in genetic forms of Parkinson's disease, we identified two exonic splicing mutations at the 5′ splice site: E417G in PINK1 and E64D in PARK7. These mutations, as well as six other previously unrecognized exonic splicing mutations in the DYSF, F8, ABCD1, GLA and DMD genes (Supplementary Table 2), have been reported as synonymous, missense or nonsense mutations. Discrimination of splicing mutations from other types of mutations is essential for understanding human diseases, because different phenotypes and different therapeutic options should be considered for different disease mechanisms. For example, splicing abnormalities in the IKBKAP and SMN2 genes can be normalized with kinetin (24) and sodium valproate (25), respectively.

Prediction of aberrant splicings using the SD-Score algorithm

To predict aberrant splicings due to mutations at the 5′ splice site, we developed the SD-Score algorithm using a training dataset and tested it using a validation dataset. Except for the PINK1 and PARK7 genes, we selected mutations without any bias in both minigenes and previously reported splicing mutations in the literature database. Of the 63 minigenes examined in the present study, six normally spliced and seven aberrantly spliced minigenes required the use of ΔR_i values for analysis. In contrast, of the 179 splicing mutations in the literature database, only four mutations required the ΔR_i values for analysis. Artificial minigenes that we constructed to understand interdependence between Cp-nucs and NCp-nucs carry two nonnative nucleotides, whereas naturally occurring mutations carry a single nonnative nucleotide. The SD-Score alone may not be powerful enough to predict the splicing consequences of mutants carrying two or more nonnative nucleotides at the 5′ splice site.

Recognition of an exon, however, is dependent not only on the 5′ splice site sequence, but also on other splicing cis-elements, including the branch point, the polypyrimidine tract, the 3′ splice site and ESEs/ESSs and intronic enhancers/silencers. Lack of information about the other splicing cis-elements and possible errors in the NCBI RefSeq annotations make the SD-Score algorithm less accurate. In addition, our training dataset comprises exclusively minigenes, and minigenes are not always spliced in the same way as their endogenous counterparts (Supplementary Figure 4). Moreover, the SD-Score algorithm is not trained to predict if any of exon-skipping, activation of a cryptic site, and intron retention occurs due to a mutation. The SD-Score algorithm, however, can efficiently predict splicing consequences of our datasets with a sensitivity of 97.1% and a specificity of 94.7% (Table 2).

Table 2.

Sensitivity and specificity of the SD-Score algorithm

Prediction	Aberrantly spliceda	Normally spliceda
Aberrant splicingb	198 (24c/174d)	2 (2c/0d)
Normal splicingb	6 (1c/5d)	36 (36c/0d)
Total	204 (25c/179d)	38 (38c/0d)

The table shows the ^aactual and ^bpredicted splicing consequences of ^c63 minigenes and ^d179 splicing mutations at the 5′ splice site in the literature database. The overall sensitivity of the SD-Score algorithm is 97.1% (198 of 204) and the specificity is 94.7% (36 of 38). The specificity is dependent on only our minigene results, because no report has been made, in which a mutation at the 5′ splice site has no effect on pre-mRNA splicing.

Comparison with the free energy, CV and R_i

Roca and colleagues (26) reported that the free energy between the 5′ splice site and U1 snRNA can be used to predict the 5′ splice site strength. The SD-Score can correctly predict 24 out of the 26 active and inactive cryptic sites in their series (data not shown). The Pearson's correlation coefficient between the SD-Score and the free energy in their series was 0.792, implying that these two parameters are likely mutually dependent.

We used our training dataset of 31 minigenes and validation dataset of 32 additional minigenes in an attempt to develop similar algorithms for the CV (21) and the R_i (7,8). We found that either the CV or R_i alone is not as efficient as the SD-Score algorithm for predicting splicing mutations (Figure 4). Both the CVs and R_is are based on the sum of information of each position in a sequence. On the other hand, the SD-Score represents information of the entire sequence of the 5′ splice site, which should include mutual interdependence between multiple positions. The SD-Score, CV and R_i, however, are mutually complementary, and our algorithm indeed achieved a high sensitivity and a high specificity with the help of ΔR_i values.

Figure 4.

Scatter graphs of the SD-Scores (A), the CVs (B) and the R_is (C) of 63 minigenes and 179 splicing mutations in the literature database. Thirty-nine normally spliced and 24 aberrantly spliced minigenes and 37 exonic and 142 intronic splicing mutations are plotted on each graph. Thresholds for the CVs and R_is were determined with the JMP-IN statistical software to give the best discrimination between normal and aberrant splicings. For the SD-Score, we used 31 minigenes as a training data set and other 32 minigenes as a validation data set. We obtained the similar thresholds of the SD-Score, even when we included 63 minigenes in our training data set (data not shown).

We also attempted to create a similar algorithm for the 3′ splice site but were unsuccessful, likely because the 3′ splice site includes at least three splicing cis-elements, and because a limited number of splicing mutations have been identified at the 3′ splice site.

Underestimated exonic splicing mutations

Most exonic splicing mutations affecting the 5′ splice site have been reported at position −1. On the other hand, to our knowledge, only 2 and 14 exonic splicing mutations have been reported at position −3 and −2, respectively (Supplementary Tables 2 and 3). When we introduced in silico all possible mutations that substitute an NCp-nuc for a Cp-nuc at positions −3, −2 and −1 into the 189 249 5′ splice sites in the human genome, the SD-Score algorithm predicted that 37.8%, 88.8% and 96.8% of these mutations would affect pre-mRNA splicings, respectively (Table 3). These percentages, as well as those of intronic mutations at the 5′ splice site, are much higher than we expected. We hope that the SD-Score algorithm serves as a practical tool to predict splicing mutations at the 5′ splice site and sheds light on underestimated aberrant splicings in human diseases.

Table 3.

Predicted ratios of exonic and intronic splicing mutations

Position	−3	−2	−1	+1	+2	+3	+4	+5	+6
Complementary nucleotide	C	A	G	G	T	A	A	G	T
A	1.8	–	93.7	–	–	–	–	93.9	56.9
C	–	89.6	99.7	–	–	99.9	94.4	98.6	75.4
G	35.0	90.5	–	–	–	48.7	96.2	–	56.7
T	76.7	86.2	97.1	–	–	99.9	94.3	97.0	–
All mutations	37.8	88.8	96.8	–	–	82.8	95.0	96.5	63.0

Numbers indicate the percentages of generating splicing mutations according to the SD-Score algorithm. The mutations are weighed by the number of occurrences of the native 5′ splice site. For example, the CAG|GTGAGG sequence, which is observed at 2562 splice sites in the human genome, has a SD-Score of −1.868. A C-to-T mutation at position −3 should generate TAG|GTGAGG, which is observed at 145 splice sites and has a SD-Score of −3.116. The ΔSD-Score of the mutation is thus −1.247. This mutation is predicted to cause aberrant splicing and is counted as 2562 mutations instead of one, because the chance that this mutation occurs should be higher than those of rare 5′ splice sites. Only mutations that substitute an NCp-nuc for a Cp-nuc are considered in this analysis, and 2 466 918 mutations have been simulated.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR online. Supplementary Table 5 is an Excel program to calculate the SD-Score algorithm.