Interpretation of mRNA splicing mutations in genetic disease: review of the literature and guidelines for information-theoretical analysis

Introduction

Pre-mRNA splicing is a necessary step in the production of a functional protein product. It consists of the recognition of intron/exon boundaries, and the subsequent excision of the introns. It is important to distinguish between alternate splicing isoforms and mutant splice forms. The former consists of using different combinations of splice sites for the same gene. It is estimated to occur in over 60% of human genes, some of which will have multiple alternate isoforms^1,2. For example, NF1 is reported to produce 46 splice variants³. The cell regulates this naturally occurring process through the availability of tissue-specific splice factors. Alternative splicing is not generated by changes in the unspliced RNA sequence, whereas mutations that produce non-constitutive splice forms are the result of dysregulation of natural splice site recognition. Mutations can have various consequences to RNA processing, such as exon skipping, cryptic splicing, intron inclusion, leaky splicing, or less frequently, introduction of pseudo-exons into the processed mRNA. A broad range of molecular phenotypes are possible depending on the type and severity of the mutation, making it imperative to understand the consequences of splicing mutations. For the purposes of this review, we consider sequence changes in genes that affect transcript structure or abundance to be mutations, regardless of their allele frequencies. Although spliceosomal recognition and RNA binding factors are operative in mutation-derived and normal alternative mRNA splicing events, this review is focused on aberrant sequence changes that alter constitutive splicing, and often result in clinically abnormal phenotypes.

The process of U1/U2-based mRNA splicing involves the recognition of a number of key sequence components^4,5, with exons defined by both intronic and exonic features^4,6. The exonic and intronic sequences flanking the 5´ end of an intron is termed the donor site and the 3´ end, the acceptor site. In typical mRNA splicing, the natural donor and acceptor splice sites span intervals of 10 and 28 bases in length, respectively. It is a common misconception that these sequences (especially the dinucleotides immediately intronic to the exon) are invariant. Although highly conserved, these sequences vary at different splice junctions within a gene as well as between genes. The particular combination of nucleotides at each position within the same splice site determines its overall strength, which dictates the likelihood of recognition by the U1 and U2 spliceosomes.

In addition, binding sites for splicing regulatory elements have been shown to reside over a range of distances from the corresponding natural splice sites⁷; the impact of these sites appears to be related to their binding affinities to the cognate RNA binding proteins and to their distance from the proximate intron/exon boundary⁸. Recognition sites for these regulatory proteins can reside either within introns or exons. Those within exons are commonly referred to as exonic splice enhancers or silencers (ESE or ESS, respectively), whereas the corresponding designations for intronic elements are ISE or ISS. Sequence variants affecting these protein-binding sites (or mutations in the binding proteins themselves) have been documented as contributing to aberrant splicing and pathogenic phenotypes. We focus on variants occurring in cis with target genes, as opposed to those in the splicing complex (in trans), leading to abnormal splicing. The efficiency and specificity of splicing depends on the combination of natural splice site strengths and the binding of splicing regulatory proteins that orchestrate exon recognition⁹.

Mutations that affect pre-mRNA splicing account for at least 15% of disease-causing mutations¹⁰ with up to 50% of all mutations described in some genes^11,12. Interpreting the effects that these variants have on splicing is not straightforward because natural and regulatory splice sites exhibit considerable sequence variation. Furthermore, performing in vitro experiments to verify the consequences of each variant is costly and time consuming, and may not be practical. In silico prediction methods have become essential resources for analyzing these variants. Software programs for splicing analysis use a wide variety of bioinformatic approaches. Several splice site prediction tools compare the predicted mutant sequence to a consensus sequence, based on a set of functional acceptor or donor splice sites¹³. A drawback of this approach is that low-frequency nucleotides present in functional splice sites are not represented, which can lead to misinterpretation and false-positive mutation predictions. One example of this was illustrated by Rogan and Schneider (1995), in which the variant, IVS12-6T>C in MSH2, described by Fishel et al. (1993) was predicted to be benign, despite being located 6 nt from the natural acceptor splice junction^14,15. The consensus sequence fails to indicate that C and T at this position are nearly equally probable, which reclassified this transition as a polymorphism rather than a pathogenic variant. This conclusion is supported by evidence that ~10% of normal individuals without predisposition to non-polyposis colon cancer harbour this alternate allele¹⁶.

Over the last 20 years, we and others have developed an information theory (IT)-based approach for prediction of splicing mutations, and their impact on mRNA structure and abundance. The effects of these mutations is founded on the formal relationship between IT and the second law of thermodynamics, in that the change in information ascribed to a sequence variant within a splice site is directly related to thermodynamic entropy and free energy of binding^17,18. A weight matrix consisting of the Shannon information (product of the probability of each nucleotide and –log₂ of its probability) at each position of the splice site is constructed. The individual information for a splice site (R_i, in bits) is defined as the dot product of this weight matrix and the unitary vector of a particular splice site sequence. The magnitude of the information content of a nucleotide within a given site is an indication of its level of conservation relative to a set of functional sites. This method retains all of the sequence variability inherent in each model of donor and acceptor splice sites. By contrast, each base in the consensus sequence has the maximum R_i value, which is actually rare in the human genome, and is generally not representative of the preponderance of natural splice sites. Prior to the introduction of IT-based approaches, consensus sequence-based methods were widely used¹³. Also, the use of neural networks, trained on sequences experimentally determined to be “bound” and “unbound”, was another early approach used to predict splice sites¹⁹. However, these unbound set of sequences are known to harbour some contaminating functional sites^20,21, which can limit the sensitivity and specificity of these networks²².

There are instances when IT does not accurately predict the consequences of a splice variant. This can often be attributed to instances involving multiple sites or multiple regulatory factors, which are not components of current splicing models. In addition, splicing regulatory proteins can share overlapping and degenerate binding sites, and may exert conflicting effects (for example, serine-arginine [SR] vs. hnRNP proteins), making in silico prediction less reliable and accurate in these cases²³. Finally, functional cryptic splicing motifs occurring deep within the introns can be challenging to identify, because they tend to be less well conserved than natural splice sites^24,25.

Nevertheless, a number of authors have recommended IT methods for analysis of splice site variants (N = 29; Supplementary Table 1). In fact, this approach has been described as equivalent to using a general reference textbook as a diagnostic tool, which complemented by functional assays, may provide a complete molecular diagnosis²⁶. Most of the applications of IT for splicing mutation analysis have involved predominantly rare diseases, as well as some low frequency variants associated with more common genetic conditions. This is because IT has been used to assess how well computed changes in binding affinity conform to levels of expression and/or patient phenotypes.

Many IT studies have focused on sequence variants in individual disorders or genes. Our synopsis of the broader implications of this work sets the stage for this compilation of peer-reviewed variants with accompanying IT analyses. We cover all publications retrieved through PubMed and Google Scholar that cite the use of IT (N = 367; Supplementary Bibliography) before September 2014. These items include primary research articles, review articles, presentations, and theses. Of all references, 216 publications reported variants or other results or analyses pertinent to this review (Supplementary Table 2). In the remaining studies, analyses were either not performed, insufficient information was provided to reproduce the reported result, or authors described unrelated applications of IT-based analysis. We summarize the spectrum of variants analyzed to obtain a global perspective of splicing mutations resulting in genetic disease. We also highlight common errors that can occur in variant analysis and interpretation, and offer guidelines for optimal use of our software programs for interpretation of splicing mutations.

Information theory and splice site analysis

IT was first introduced by Claude Shannon in 1948 and is now used in a variety of disciplines to express the average number of bits (i.e. the information content) needed to communicate symbols in a message²⁷. Bits are the basic unit used in computing and can have one of two values (typically the answer to a yes/no, true/false, or +/- problem). In nucleic acid molecular biology, the symbols in the message comprise a group of related, aligned sequences, with the average number of bits in the set corresponding to the amount of information in the message. This is determined from the information content at each position in the sequence, summed over all positions²⁸. The average information is depicted graphically by a sequence logo, which stacks the individual nucleotides at each position ranked by frequency, and where the height of the stack is the position-specific contribution to the average information²⁹. If the set of sequences are functional binding sites recognized by the same factor, the individual information in each site (i.e. R_i value) is related to thermodynamic entropy, and thus, to the free energy of binding¹⁸.

The information content of a nucleic acid binding site is related to the affinity of its interaction with proteins and other macromolecular complexes, such as the case during mRNA splicing¹⁸. Information theory-based position weight matrices (PWM; R_i [b,l] - also referred to as a ribl - where b and l correspond to the nucleotide and position in the splice site) can be determined for set of known binding sites, in this case, for the purpose of calculating individual and average sequence information²⁸. Figure 1 shows an example of sequence logos for the canonical acceptor (or 3´, recognized by the U2 spliceosome) and donor (or 5´, recognized by the U1 spliceosome) splice sites, computed from the majority of constitutive sites at annotated splice junctions in the human genome³⁰. The information contained within the natural splice donor site is distributed between the last codon of each exon and the adjacent 6 nucleotides of intronic sequence, whereas the acceptor sites are almost entirely intronic, extending 26 nucleotides upstream from the exon boundary.

Figure 1. Distribution of deleterious natural site variants relative to information content.

A) The sequence logo for human acceptor and donor splice sites based on the positive (+) strand of the October 2000 (hg5) genome draft. The logo shows the distribution of information contents (R_i in bits) at each position over the region of 28 nucleotides for acceptor [-25, +2] and 10 nucleotides for donor [-3, +6] from the first nucleotide of the splice junction (position 0). Nucleotide height represents its frequency at that position. The horizontal bar atop each stack indicates the standard deviation at that position. This figure was modified from Rogan et al. (2003) to include splice sites in genes on both strands of the annotated human reference genome³⁰. B) The distribution of deleterious single-nucleotide variants reported at the natural acceptor (left) and donor (right) splice sites. The variants used to populate this graph (Supplementary Table 8) were included only if they were reported to negatively affect splicing (N = 419 for acceptors, 599 for donors). The image was aligned to the sequence logo (A) to illustrate potential correlation of number of splicing variants at a position to the information content at that position.

The distributions of R_i values for these sets are approximately Gaussian, with a couple of important exceptions, namely the distribution has defined upper and lower bounds¹⁸. The upper limit corresponds to the consensus sequence, as it is not possible to have stronger binding than an exact match to this sequence. The theoretical lower limit corresponds to R_i = 0 bits. An R_i value less than zero implies that energy would be required (ΔG > 0 kcal/mol) for a stable binding complex to form, i.e. that the event would not occur spontaneously without an exogenous source of energy. The minimum strength site is zero bits, the equilibrium state (ΔG = 0). Assuming the contacts at each position in the same binding site form independently, this approach is accurate and quantitative. Altering a nucleotide with high information (implying high prevalence and conservation at that position) will have a greater impact on binding, than if a less-well conserved base were altered. The change in information due to a mutation in a site (ΔR_i) is the difference between R_i,final and R_i,initial values, where R_i,final is the information of the sequence containing the variant, and R_i,initial the information of the reference (wild-type) sequence. The minimum fold change in binding affinity resulting from the mutation is an exponential function based on ΔR_i, or ≥ 2^ΔRi (Ref.¹⁸).

Software resources

Natural sites

The early splice site recognition literature often oversimplified the composition of the U1/U2-type 5´ donor and 3´ acceptor sites by presenting only consensus sequences and truncating the positions in each site^13,45,46. However, the conserved tracts extend well beyond the canonical GT and AG dinucleotides adjacent to intron/exon junctions. Furthermore, a small, albeit significant, proportion of natural donor sites (~800, or 0.7%) contain cytosine at position +2 in the genome. This is reflected by a corresponding small decrease in average information at this position (Figure 1). Sequences adjacent to these positions are more variable, but are nevertheless essential for the accurate recognition by the spliceosome. Specifically, the donor site is defined by the three terminal nucleotides of each exon and the first seven bases of the downstream intron. Conversely, acceptor sites are represented by the first two bases of the exon and the last 26 bases of the upstream intron. Because ASSA and ASSEDA use an integer-based coordinate system, there is a zero coordinate at the first intronic base of each splice site (Figure 1), which is not used in the conventional numbering system. The coordinate ranges for the donor and acceptor site positions are therefore [-3, +6] and [-25, +2], respectively. Individual information analysis computes the R_i values over these intervals for normal and variant-containing splice sites. As discussed below, information content present in intronic intervals justifies sequencing and analyses of sequences well beyond the locations of the splice junctions themselves.

Certain variants within donor and acceptor sites are tolerated and may even have benign effects, while others have a deleterious impact on spliceosomal recognition. IT accounts for all of these possible outcomes. Unusual donor sites (i.e. with cytosine at position +2) are detected by information analysis, but could be falsely called deleterious by consensus sequence-based methods. Although the terminal position of exons contributes significantly to donor splice sites with a preference for G, a significant proportion of sites naturally possess A or U at this position, or less frequently, C.

Of the published IT-based variant analyses, single nucleotide variants (SNVs) that were reported to affect a natural splice site (multi-nucleotide and insertion/deletion variants are listed separately in Supplementary Table 3) were compiled and reanalyzed. After reducing this set to only those variants occurring within the intervals covered by the splice site information weight matrices described above, 1152 SNVs were reported to affect the strengths of either natural donor or acceptor sites. A variant was considered deleterious if it was predicted to affect splicing (either leaky expression or exon skipping), or if it was experimentally shown to reduce or abolish splicing of the corresponding exon. In instances where prediction and validation did not concur, the latter were used to determine the effect of the variant. Variants predicted to have a neutral effect but demonstrated to be deleterious in the validation study were classified as damaging. In total, 1010 deleterious natural splice site variants were analyzed (Supplementary Table 4).

Sequence conservation has long been considered a surrogate measure of evolutionary constraint and, by inference, functional significance. The average information quantitates the relative conservation at each of the positions within a binding site. We compiled the mutation spectra for all mutations that significantly affected the strengths of donor and acceptor splice sites and compared these with the average information contents at each position. The panels in figure 1b respectively indicate, at each position of the natural acceptor and donor sites, the frequencies of variants deemed deleterious by information analysis. Interestingly, when the sequence logo is overlaid with the histogram of the corresponding mutation spectra, the relative frequencies of deleterious mutations and the average information are comparable. Indeed, these frequencies and the information contents across each type of site are strongly correlated (r=0.95 for acceptors and 0.89 for donors). Our interpretation is that the susceptibility to deleterious mutation at a position is related to its overall conservation within the splice site, which reflects the contribution of that ribonucleotide to the stability of the interaction with the corresponding spliceosome. Nevertheless, there is an unstated bias in ascertainment in these mutation spectra. Variants occurring at sites with low information and/or that are benign are underrepresented, as they are less likely to be associated with genetic disease, and were less likely to be reported. Also, the distribution is dependent on the region sequenced by the authors of the reviewed publications; in early work, the full sequence interval containing the entire splice site was sometimes not included or unavailable for analysis.

An interactive website was created to summarize this set of SNVs. This software application renders interpretations of variant effects in a more practical, useful way than the corresponding table of supplemental data (Supplemental Table 10). The “Splicing Mutation Calculator” (SMC; http://splicemc.cytognomix.com) is a web service that amalgamates all published results for the same type of substitution in a natural splice site, regardless of genic context. Variants that create cryptic splice sites were not included, because we consider these cases to be sequence-specific as opposed to positional. With this program, users have the option of exploring mutation data (at present, only SNVs can be analyzed) linked to the original literature citations. SMC processes and provides literature support for the variants that occur within the defined regions spanned by natural splice sites. The user first selects the type of site (donor or acceptor), position (based on ASSEDA’s integer-based system), wild-type or reference nucleotide, and the alternate substitution at that position (Figure 2a). The software tool outputs the ΔR_i and the number of variants that have been reported and analyzed to date using IT (Figure 2b). SMC provisionally classifies the reported variants based on the degree to which these predicted effects are expected to decrease spliceosomal affinity, and consequently splicing. The following criteria are empirically based on affinity changes and a summary of published phenotypes associated with these changes: “Deleterious” (if the site is weakened by more than 7.0 bits), “Probably Deleterious” (if the site is weakened such that -4.0 bits ≥ ΔR_i ≥ -7.0 bits), “Leaky” (the site is weakened such that -1.0 bits ≥ ΔR_i ≥ -4.0 bits), or “Benign, probable polymorphism” (if the site is weakened by less than 1.0 bits). In this first release of SMC, we have omitted “benign” variants, which are likely polymorphisms; these will be catalogued and included in a later version. It is important to appreciate that the ΔR_i is a constant for a specific nucleotide change at a specific position, though the absolute strength of the splice site depends on the sequence context of the mutation. This context varies between mutations, and R_i,initial is not the same for each case, which can result in different R_i,final values for different mutations.

Figure 2. Sample retrieval of average change in information content (ΔR_i) with splicing mutation calculator (SMC) for published mutations.

A) Example mutation input for SMC (T>A at the 3^rd intronic position of natural acceptor). The type of splice site is selected by clicking on the corresponding sequence logo (acceptor [left] or donor [right]). The purple slider bar appearing below the logo is used to select the position of the mutation. The reference and mutant nucleotides are then designated, and the variant is submitted to the software (‘Submit your selection’). SMC outputs a table indicating the user input, the number of instances in the literature where this substitution has been analyzed using IT, and the computed ΔR_i values (in bits) using both the old (1992; top) and new (2003; bottom) ribls. The cell color for ΔR_i values indicates the predicted severity of the inputted variant according to defined thresholds^22,168. B) Tabular output detailing each instance of the selected mutation from the source table. The user may view, in a separate window, extensive details of all variants referred to in SMC output (Supplementary Table 10).

Besides published sources, the software also can predict effects of mutations by computing ΔR_i values directly. Particular substitutions that have not been reported in Supplemental Table 10 can nonetheless be provisionally interpreted. The ΔR_i value is computed and reported from the ribl. While SMC enables rapid exploration of results for validated and novel mutations, it is, however, not a replacement for ASSEDA or the Shannon Pipeline, since it does not consider the sequence context, which can also influence the interpretation of deleterious, leaky, or benign variants.

Branch-point mutations

Although branch-point site (BPS) recognition occurs independently and post-exon definition, mutations in this sequence have also been described, due to its proximity to the natural acceptor site. Following the recognition of and binding to the 5´ss (upstream donor site) by the U1 snRNP, the U2 is recruited to the 3´ss (downstream acceptor) and recognizes the BPS, resulting in the formation of the pre-spliceosome⁵⁵. Association of U2 with the BPS is essential, as it is the first energy-requiring step, allowing for the tri-snRNP complex of U4/U6 × U5 to be recruited to the BPS, which produces a catalytically active spliceosome⁵⁶. The BPS typically contains a conserved adenosine and a downstream polypyrimidine tract. It is located within 40 nt of the natural 3´ss, however there are reported cases where it can be up to 400 nt away.

Recognition of the BPS is thus a crucial step in proper splicing, and sequence variants can disrupt this event, impede lariat formation, and intron excision. The complete list of BPS variants analyzed using the ASSA and ASSEDA server can be found in Supplementary Table 5. The variants range in distance from 0–76 nt from the natural acceptor junction, and either weaken, abolish or strengthen the BPS. When validation assays were performed, the prediction by the server was correct in 9/11 cases. We deemed the two other cases to be partially discordant (NM_004628:c.413-24A>G and NM_005902:IVS8-55A>G). ASSEDA predicted these variants to abolish the BPS, but leaky and normal splicing was observed, respectively. The predictions are partially concordant with experimental findings because ASSEDA also predicted the existence of nearby alternative BPS, which if used, could account for the observed phenotype.

Although IT-based prediction of a variant effects on BPS has been accurate, the number of validated sites used to compute the ribl is substantially smaller (N = 20), and it is not as reliable as those used to determine R_i values of natural acceptor and donor sites. Furthermore, these motifs are short and relatively frequent in unspliced mRNA. One possible explanation for the rarity of BPS mutations is that compensatory, alternative BPS sequences can be recognized and used. Furthermore, the weak constraint on the precision of the distance between the BPS and the 3´ (acceptor) splice site (Figure 3) further enables activation of these alternative sites. These factors increase the chance that a variant will be falsely predicted to affect a BPS. For example, variants within donor splice site sequences are routinely predicted to alter strength of false BPS. This error is easily avoidable if the potential recognition sequence is filtered for the genomic context of the variant, as well as its proximity to acceptor splice sites.

Figure 3. Ribl used for the prediction of a variant’s effect on branch-point sites.

Sequence logo for information model for the branch-point site, created using 20 annotated branch-point sequences.

Activation of cryptic splicing

It has been estimated that 1.6% of disease causing missense mutations can affect splicing and recent predictions suggest that approximately 7% of exonic variants in the general population may disrupt splicing, which includes cryptic splicing^57,58. The genome is replete with pseudo (or decoy) splice sites with varying degrees of similarity to natural sites that are not recognized in constitutive splicing⁵⁹. However, mutations that alter the strengths of either these decoys or the natural splice site of the same polarity may shift the balance of isoforms towards non-constitutive splice isoforms that predominate over or eliminate normal mRNAs (Figure 4). Mutations can create a cryptic splicing event by creating or strengthening a site in either intronic or exonic regions (Figure 4, Type 1), weaken the natural site while simultaneously altering an overlapping decoy site (Figure 4, Type 2), or exclusively weaken the natural site, leading to the activation of a pre-existing decoy site (Figure 4, Type 3). Although the contributions of cryptic splicing to genetic disease have long been recognized, IT analysis correctly predicts most, but not all, cases (Figure 4). The challenges in identifying potential cryptic sites or determining activation are attributable to our incomplete understanding of the requirements for activation^60–62, which include exon length, processivity of donor site recognition, and involvement of splicing regulatory factors. A database of aberrant 3´ and 5´ splice sites has been compiled⁶².

Figure 4. Outcomes of cryptic splicing mutations.

A prototypical internal exon (in purple) with flanking exons (in blue); introns are represented by black solid, and dashed lines (top). The three types of cryptic splice site activation are then illustrated. Type 1 cryptic splice site activation (left) is caused by the activation (green arrow) of a cryptic site by strengthening a pre-existing site, or by creating a novel splice site (blue). Type 2 (middle) results from the simultaneous weakening or abolition (red arrow) of the natural splice site while strengthening or creating (green arrow) a cryptic site. Type 3 (right) involves the activation of a pre-existing cryptic site due to the weakening or abolition of the natural splice site (indicated by orange triangle). The number of cases that have been reported in the literature that have been analyzed by IT for each type is indicated, with the percent accuracy in parentheses. The bottom row represents the resulting mRNA structure due to the activated cryptic splice site.

Figure 5. Distribution of activated cryptic sites.

The frequency of validated cryptic splice acceptors (A) and donors (B) occurring at positions relative to the natural splice site. Positions are given using ASSEDA coordinates. Lower panel expands the cryptic site distribution of the region circumscribing the natural splice site.

Another bioinformatic method for cryptic site recognition relies on a training set composed of cryptic sites that are known to be used⁶³. There are a number of drawbacks to this approach: the training set is itself not representative of all cryptic sites; and sites that are altered but unused cannot be discriminated from those that are activated (since the latter group also depends on the strength of the corresponding natural splice site). IT-based methods rank cryptic and cognate natural site strength in a way that predicts whether the site will be activated, as well as the abundance of each pair of splice isoforms. Furthermore, the structures of the prospective isoforms are presented by ASSEDA with relative quantitation of each, both prior to and post-mutation.

During our review, we noted 203 variants with experimental support for cryptic splicing (Supplementary Tables 6–8). Of these, 38 variants resulted in Type 1 cryptic splicing. From those, site activation (existence of the site and strength ≥ 2.4 bits²²) was correctly predicted by ASSEDA in 34 cases (89.5%). We identified 56 variants resulting in Type 2 splicing, 38 of which (67.7%) were accurately predicted, while the remaining 119 variants resulted in Type 3 cryptic splicing and 99 (90.8%) of the alternate splice sites matched predictions.

Prediction of Type 3 cryptic splicing was more accurate than Types 1 or 2. The criteria for concordance with experimental data were that ASSEDA predicted both the cryptic site and that the variant weakened the natural site. However, the strength of a site is not the sole determinant of whether or not a site is activated. Unlike natural sites, novel cryptic sites are not under selection to maintain binding to the spliceosome, and their genomic context is less constrained than natural splice sites. The presence of cooperative splicing enhancer or repressor elements adjacent to cryptic sites, which could influence cryptic splice site activation, is not yet predictable. Additionally, many of the reported activated cryptic sites have been confirmed using non-quantitative approaches, and these may not constitute the predominant splice forms relative to constitutive exons with stronger natural sites. Finally, certain isoforms may not be detected; as aberrant transcripts are often subject to degradation and the tools used to evaluate functional splicing consequences do not always have sufficient resolution to distinguish small differences in isoform structure. All of these factors can affect the concordance of predicted cryptic site activation with experimental validation.

We also separated each sub-group of cryptic splice variants by location (intronic vs. exonic) and computed the average difference in strength between pairs of natural (post-mutation) and the activated cryptic sites. For intronic Type 1 variants, activated cryptic sites were 0.86 ± 5.28 bits stronger than the corresponding natural site (N = 12). There were eight Type 1 variants (4 at acceptors and 4 at donors) that were missed, because the R_i,final value of the natural site exceeded the strength of the corresponding cryptic site by ≥ 1.0 bits (variants with ΔR_i < 1.0 bits are not reliably detected experimentally). We hypothesize that these cases could be explained by concomitant changes in surrounding regulatory binding site sequences. Exonic Type 1 variants were often slightly weaker than their cognate natural sites (-1.1 ± 3.8 bits; N = 26). Nearly all of these involved ectopic donor site activation (12 of 13), consistent with a processive mechanism for donor site recognition, which searches downstream from the acceptor splice site to the first donor site of sufficient strength to form an exon³⁵. The opposite pattern was observed with intronic Type 2 cases, in which 20 of 21 exceptions occurred at acceptor sites. On average, the activated cryptic site exceeded the strength of the cognate natural site (1.3 ± 4.6 bits; N = 57). Activated, exonic Type 2 acceptor cryptic sites tended to be weaker than their natural site counterparts (-2.2 ± 3.3 bits; N = 4). This result may be attributable a low sample size, with 2 of these mutations exhibiting natural sites that were stronger (≥ 1.0 bits) than the corresponding cryptic site (1 donor and 1 acceptor). Finally, Type 3 activated intronic cryptic sites exhibited the greatest difference between the strengths of cryptic sites and cognate natural splice sites (6.3 ± 4.9 bits; N = 104). This category contained the fewest number of exceptional cryptic sites, with R_i values less than those of natural sites (5 acceptors and 3 donors). This is consistent with the idea that the intronic cryptic sites are generally not under selection for adjacent functional regulatory binding sites, and, in order to be activated, are required to be substantially stronger than the natural site. Although R_i,final values were stronger (2.1 ± 1.9 bits; N = 20) than the natural site, exonic Type 3 cryptic splice sites did not show as great a difference in strength with a single exceptional case (of an acceptor). Despite these exceptions, activated cryptic splice sites are generally stronger than the corresponding natural splice sites²².

Combinatorial effects

While functional natural splice sites and an intact BPS are integral for accurate and efficient splicing, other genetic elements have been shown to make essential contributions to exon definition⁶⁴. Introns will often contain more than one potential splice site recognition sequence, but nevertheless, the correct natural site is consistently selected⁵⁹. Differences among the strengths of potential sites, as determined by IT analysis, are a major, but not the sole, determinant of splice site utilization. The implication is that additional sequences within the gene are necessary to ensure specificity and precision of exon recognition. Studies of facultatively expressed alternative exon structures have revealed cis-acting sequence elements that function to enhance or repress exon recognition. These sequences cooperate with factors that recognize natural splice sites, whose sequences and relative strengths can vary considerably. Depending on their context, these elements have been referred to as exonic splicing enhancers (ESEs), exonic splicing silencers (ESSs), intronic splicing enhancers (ISEs) or intronic splicing silencers (ISSs). In general, these elements serve as binding sites for trans-acting elements, which will either promote or impede the spliceosomal recognition of a splice site. The majority of enhancer elements will act through the recruitment of SR proteins and associate components of the U1 and U2 spliceosomes^65,66. Silencers are often of the hnRNP class, which act through a diversity of mechanisms including steric hindrance, the formation of dysfunctional complexes, or blocking processiveness^67–69. To add to the complexity of splicing regulation, it has recently been shown that SR protein function is dependent on context, i.e. whether the corresponding binding site is intronic or exonic^70,71.

To improve accuracy of exon definition, the strengths of regulatory elements (i.e. their R_i values) have been incorporated into splicing mutation prediction. The significance of regulatory elements in disease has been demonstrated in many cases. For example, in the NF1 gene, ESE disruption is the primary cause of exon skipping⁷². Many other genes, including APC, SMN, BEST1, PDHA1^73–76 have been proven to harbour variants that disrupt ESEs and have a confirmed impact on mRNA splicing.

Adding to the complexity, the recognition sequences for these RNA binding factors, while well defined, tend to be short, and can vary to the degree that the same sequence may contain overlapping elements of binding sites for multiple factors. However, this does not necessarily imply that such a sequence is bound with similar affinity by each factor or that it contributes to exon definition. At the same time, these sequences tend to be evolutionarily conserved and may be required for proper splicing^77,78.

ASSEDA optionally incorporates PWMs for regulatory binding sites for mutation analysis (Table 1) in addition to the default donor and acceptor sites. The program selects the most proximate predicted ESE/ISS to the natural splice site when calculating R_i,total. The molecular phenotype, which dictates the splice isoforms that are predicted and their relative abundance, accounts for both the potential effect on the natural site and the most relevant splicing regulatory site. For these regulatory binding sites, a second gap surprisal term specific to the ESE/ISS of interest is applied to the R_i,total calculation³⁵. The gap surprisal functions for SF2/ASF and SC35 have been previously described³⁵, where the most common distance of the ESE/ISS is within 10nt of the natural site. The gap surprisal penalty gradually increases with distance from the natural site. Gap surprisal distributions for ELAVL1, TIA1 and SRp55 show a similar pattern, while hnRNPA1 and PTB binding sites are strongly clustered around splice junctions. It should be feasible to include the contributions of multiple splicing regulatory binding sites of the same or different RNA binding proteins in determining R_i,total; however this capability had not yet been implemented. Currently, if multiple sites of the same type are altered, the strongest (before or after mutation) is chosen by ASSEDA software.

Table 1. Splicing regulatory protein binding sites ASSEDA scans for and their associated effect on
splicing.

Splicing Factor	Rsequence (bits)	Sequence Logo	Location-dependent effect on splicing
			Intronic	Exonic
hnRNPH1	8.9 ± 1.8		E79,80 / S81,82	S / E83
hnRNPA1i	4.6 ± 1.5		S / E84	S
TIA1	7.6 ± 3.1		E	N/A
SRSF6 (SRp55)	5.2 ± 1.4		E / S82	E / S
SRSF5 (SRp40)	4.5 ± 1.5		E / S82	E / S85
SRSF2 (SC35)	4.5 ± 1.6		E / S86	E / S87
SRSF1 (SF2/ASF)88	5.8 ± 1.5		E / S86,89,90	E / S
PTBii	4.9 ± 1.9		S / E91	S
ELAV1	9.6 ± 3.4		S / E / N92,93	S

Reported dominant effect is bolded. E – Enhancer; S – Silencer; N - Neutral.

ⁱEnhancer activity by hnRNP A1 occurs at the junction⁸⁴. ⁱⁱPTB does not directly enhance splicing, but can do so indirectly by preventing the binding of splicing repressors⁹¹.

Although the disruption of splicing regulatory sequences can cause aberrant splicing, the interpretation of variants affecting these sites is not as straightforward. Due to their degenerate nature, short sequence, and a lack of understanding of the context of their use, altered regulatory sites should be functionally validated before being deemed pathogenic⁷. Using variants from a number of different studies, ASSEDA accurately predicted experimentally determined changes in binding at a splicing regulatory site 75% of the time (N = 12)³⁵. However, there were instances where regulatory sequences had been analyzed by IT, and considered to contribute to disease, but the results were not reproducible. For example, Kölsch et al. (2009) described SNPs associated with Alzheimer’s Disease, one of which strengthened and created SRp40 and SRp55 sites, respectively, but were reported by authors to be abolished⁹⁴. This study did not report any evidence to support the significance of these predictions.

Functional validation of the effects of these mutations could contribute to understanding the roles of these factors in regulating constitutive splicing. Similarly, there is still little understanding on how multiple regulatory binding sites within the same region function as a unit. Using a pull-down assay, Olsen et al. (2014) demonstrated how different variants affect the binding of multiple regulatory proteins. One mutation was predicted to create and strengthen multiple hnRNPA1 sites and slightly strengthen an SF2/ASF (SRSF1) site. The pull-down studies showed up-regulation of hnRNPA1 binding and a decrease in SF2/ASF binding. However, SF2/ASF binding increased when a mutation disrupting hnRNPA1 affinity was introduced, suggesting that the strong hnRNPA1 sites outcompete the weaker SF2/ASF site.

In some instances, alterations in regulatory splice site recognition sequence and natural splice strength occurred concomitantly, with both predicted to have similar effects on splicing. Alteration of a regulatory sequence can sometimes provide a plausible explanation for discordant in silico prediction and experimental validation. As an example, Smaoui et al. (2004) analyzed a donor site mutation (NM_001040667:c.1327+4A>G) in HSF4 in a family with congenital cataracts⁵⁰. This variant was predicted to cause leaky splicing (R_i,final = 5.4 bits; ΔR_i = -2.6 bits; 67.5% residual binding), however RT-PCR showed complete exon skipping. Our further analysis showed that it is predicted to also create an overlapping hnRNPA1 site (R_i,final = 4.2 bits; ΔR_i = 17.1 bits). Another case involved a mutation in the XPC gene (NM_004628:c.2033+2T>G) that created a novel intronic cryptic site 4 nt downstream of a natural donor site⁹⁵. However, a weaker site 68 nt downstream from the natural site was activated. A possible explanation could be that activation of the cryptic site is influenced by a neighbouring hnRNPA1 site that is itself strengthened (R_i,final = 5.2 bits; ΔR_i = 2.2 bits) and an SRp55 site that is significantly weakened (R_i,final = 1.9 bits; ΔR_i = -4.0 bits).

The effects of changes in regulatory binding site strengths may ascribe potential functions to previous VUS. For example, Maruszak et al. (2009) present a PIN1 variant associated with late-onset Alzheimer’s Disease (NM_006221:c.58+64C>T)⁹⁶. Based on IT, it is expected to abolish an intronic SC35 site, which could have either an enhancing or silencing effect (Table 1). A 2.82-fold decrease in transcript levels was demonstrated, which is concordant with previous findings reporting decreased PIN1 levels in the brains of Alzheimer’s Disease patients. Another study described an exonic missense variant within the ETFDH gene in a patient with multiple acyl-CoA dehydrogenase deficiency (NM_004453:c.158A>G) that showed evidence of exon skipping. The variant was predicted to be “benign” or “tolerated” when evaluated with PolyPhen and SIFT²³. ASSEDA, on the other hand, predicted the creation of an hnRNPA1 site (R_i,final = 5.9 bits; ΔR_i = 17.1 bits), a slightly strengthened hnRNPH site (R_i,final = 4.0 bits; ΔR_i = 0.2 bits), the abolition of an SRp40 site (R_i,final = -3.3 bits; ΔR_i = -6.3 bits) and two novel, weak SF2/ASF sites (R_i,final = -4.6 bits; ΔR_i = 0.8 bits and R_i,final = -2.4 bits; ΔR_i = 0.4 bits)²³. The natural donor site was unaltered by the mutation. As indicated earlier, the mutation was confirmed experimentally to increase hnRNPH and hnRNPA1 and decrease SRp40 and SF2/ASF binding.

Validation of results

A number of early mutation studies did not perform expression analysis and relied solely on the ASSEDA or ASSA server to interpret potential mutations. This is not recommended, as there are limitations to any in silico predictive method, which impacts accuracy and precision of the prediction. Assuming that the impact of the mutation on expression can be detected, experimental validation of IT-based mutation analysis can reveal its limitations. We describe the various validation methods that were employed in the articles where expression data were available. Below, advantages and disadvantages of these approaches are explored, as well as how lower sensitivity validation can result in misinterpretation. Finally, we determine the accuracy of IT-based prediction, and point out some instructive, discordant cases.

Accuracy of IT-based prediction

We previously evaluated the accuracy of IT-based prediction using a set of validated splicing mutations (85.2%; N = 61)²⁵. Other studies have also evaluated the accuracy of ASSA/ASSEDA while evaluating differences between multiple predictive programs and have shown varying levels of concordance (68.8%, N = 16; 90.1%, N = 22; 100%, N = 24)^51,104,116. With a comprehensive list of all published variants analyzed using IT-based methods (Supplementary Bibliography), we perform a meta-analysis of all of these variants to minimize bias in interpretation and impact of ascertainment of specific phenotypes from individual studies. The list of variants is more extensive than any previous study examining accuracy of IT-based methods. The variants are not restricted to a single or even group of diseases, but rather cover over 150 different conditions (see Supplementary Table 2).

In total, 905 variants were reported in 122 different publications to have been validated for their effects on splicing (1,727 total variants analyzed from 216 papers – Supplementary Table 9). In all cases, the authors performed information analysis; however, the validation experiments were sometimes contained in the original reports and in other cases, later studies. In a minority of mutations, the validation results were either uninformative (N = 36) or the methods did not directly imply an effect on splicing (N = 2); these mutations were therefore excluded in determining the accuracy of predictions (shaded in grey in Supplementary Table 9).

More specifically, in order for experimental results and predictions to be considered concordant, one or more of the following criteria had to be met:

Cumulatively, 87.9% of variants documented by expression studies (762 of 867) that satisfied these criteria were accurately predicted by ASSEDA. A minority of papers reported variants to be “partially concordant” (3.1%; 27/867), meaning that while the cryptic site observed was predicted, it was not the most likely splice isoform relative to other expressed cryptic exons. Because this method of scoring met our criteria (see point d above), we included these in our determination.

Misinterpretation of variant effects

While preparing this review, several variants misinterpreted with IT-based tools were noted. These variants have been re-analyzed to disseminate the correct findings and to avoid making similar errors in the analysis of newly discovered variants. Supplementary Table 2 contains these results. The most common problems result from unfounded emphases on altered or pre-existing cryptic sites that are determined to be significantly weaker relative to the cognate natural site^{109,128–132}, and from selectively reporting a single change in the R_i value when, in fact, multiple significant changes can be detected^{48,128,133–136}. An example of the first type of error is exemplified by a variant in CGI-58 (ABDH5), where the natural splice site is 9.1 bits (or ≥ 549-fold) stronger than the reported cryptic site¹²⁹. Henneman et al. (2008) selectively reported the effect of a mutation that weakens a natural donor splice site in APOA5, however only a change in the information content of an SC35 binding site was indicated.

Other common problems include incorrect declaration of small ΔR_i values as significant changes^109,137,138, use of incorrect R_i,min values^139,140, and the computation of predicted binding strength changes on a linear scale¹⁴¹ rather than the correct exponential function (i.e. ≤ 2^ΔRi)¹⁸. Smaoui et al. (2004) described an 8.0 bit donor site as weak, which is actually equivalent in strength to R_sequence, the average strength³³. Allikmets et al. (1998) and Ozaltin et al. (2011) both described an inactivating mutation as leaky, because the weakened site remained above the R_i,min. However, the variant mutation produces a site with < 0.7% of its original binding affinity, which would substantially reduce exon recognition and lead to exon skipping¹¹⁶. Also, cryptic sites created in the promoter regions of genes should not be considered to be splicing mutations¹⁴². Variants that are predicted to create a cryptic site upstream or overlapping a natural site of the opposite polarity (i.e. cryptic donor upstream of a natural acceptor) have been reported^131,132,143, which would be inconsistent with established splicing mechanisms³⁵. A rare exception that could render such a site active is to the creation of a cryptic exon that occurs in conjunction with a proximate, correctly oriented, pre-existing cryptic splice site of opposite polarity^22,30. Insufficient numbers of examples of mutations creating cryptic exons have been reported to date for ASSEDA to accurately predict these exons by default.

Several results were generated by incorrect entry of mutations in to ASSA/ASSEDA. For example, altered cryptic splice sites have been confused with natural sites^{48,137,144,145}. Additionally, ‘residual binding strength’ displayed has been misinterpreted as a percent decrease^144,146. Strong, pre-existing cryptic sites outside of the default sequence analysis window (54 nt circumscribing the mutation) have also been missed because the window was not expanded to include these sites¹⁴⁷. Although the predicted isoform structure generated by ASSEDA will, by default, display skipping for mutated natural sites with ΔR_i ≥ -7.0 bits (or ≥ 128-fold)¹¹⁶, smaller decreases in natural site strength of an internal exon can sometimes partially induce exon skipping. This value is adjustable, and it may be advisable to explore different thresholds depending on the particular susceptibility of a splice junction to exon skipping. Sharma et al. (2014) used the default threshold from ASSEDA to interpret CFTR mutations c.2988G>A (9.6 to 6.6 bits, natural donor site of exon 18) and c.2657+5G>A (9.1 to 5.6 bits, natural donor site of exon 16), but exon skipping was documented. IT analysis was not discordant for these variants, which significantly weaken the corresponding splice sites by ≥8- and 11-fold, respectively, and has been shown in other genes to lead to exon skipping, leaky splicing, or both of these outcomes. Aissat et al. (2013) tabulated variants that were predicted to affect strengths of ESE binding sites, but did not comprehensively report all findings even though predictions by ASSA and ESEfinder were concordant (eg. CFTR: c.1694A>G). Alternate mutation entry methods, which do not use contextual gene name annotations, such as entry by rsID, report predicted binding changes on both strands. A report indicating abolition of SRp40 binding sites on the antisense strand was confused with binding sites for CYP46A1, which is transcribed from the sense strand¹⁴⁸.

Other problems include inadvertent mislabelling of splice site type or location^149–151, interchange of the terms information content and change in information (R_i and ΔR_i)¹²², and unclear variant interpretation (i.e. “run on into the intron”)¹⁵². Moriwaki et al. (2009) claim ASSA did not predict a mutated natural donor site, but in fact, the site was present in our reanalysis¹⁵³. Published R_i values from Rogan et al. (1998) and von Kodolitsch et al. (1999) are in some instances different from current values due to updates of the reference genome sequence. Nevertheless, the overall predicted effect did not change, but initial and final R_i values were inconsistent. Interpretations of certain mutations could not be reproduced in some instances^{103,145,154–156}. Finally, we noted that ASSEDA can sometimes improperly parse indels entered using c. or IVS notation. Such errors have led to published false results^{67,116,157,158}.

Interpretation of published variants in studies that use information analysis

Polymorphisms and splicing

Early studies suggested that common polymorphic sequence variations at splice sites corresponded to small ΔR_i values, consistent with these changes having little impact on mRNA abundance²². More recently, it has been appreciated that certain rare SNPs have significant genetic loads, can actively alter mRNA splicing profiles, and lead to non-obvious splicing phenotypes^58,189. Nevertheless, it is not uncommon for reports to solely analyze novel variants and ignore known SNPs^{136,156,158,192}, or limit results only to those that occur in the vicinity of natural splice sites¹⁸⁴. We find that 56.4% of common SNPs (with population frequencies ≥ 1% in Supplementary Table 2) within natural sites significantly alter their strength (12.8% abolish and 28.2% cause leaky splicing, 15.4% modestly strengthen sites [ΔR_i < 2.6 bits]), and 43.6% have insignificant ΔR_i values, as expected (N = 39). The mean R_i,final and ΔR_i values, for these natural sites are 7.9 ± 4.0 bits and -1.4 ± 3.0 bits, respectively, which suggests the effects of these polymorphisms on splicing are nil to limited. However, polymorphisms can significantly modulate splicing, as some common SNPs are predicted to abolish natural splicing (Supplementary Table 2: #1291, 1296, 1431, 1435, and 1436). These include rs10190751 in CFLAR, which modulates the production of two short isoforms, and is associated with an increased risk of lymphoma^189,193, rs3892097, which alters exon inclusion in CYP2D6³⁰ and leads to a non-functional protein and altered drug metabolism¹⁹⁴, and rs1805377 in XRCC4³⁴, which has been associated with oral cancer susceptibility¹⁹⁵ and increased risk of gliomas¹⁹⁶. There is also experimental support for common SNPs that have been predicted to affect splicing^{22,98,107,110,114,118,149,164,189,197}. For example, experimental evidence for increased exon inclusion has been described for three of six SNPs that increase strength of natural splice sites^189,190. Numerous common SNPs, which were either deemed neutral or predicted to affect splicing, have not been confirmed experimentally^{14,22,25,34,52,94,99,131,133,144,145,148,151,165,166,168,179,183,185,188,198–208}. Polymorphisms with significant information changes should be investigated, as they may not be completely benign and can have a significant impact on mRNA splicing.

Inference of variant pathogenicity by IT analysis

Recently, American College of Medical Genetics and Genomics recommendations for reporting incidental findings in sequencing have suggested that bioinformatic predictions are not sufficient to declare clinical significance²⁰⁹. Preceding the publication of these guidelines, numerous peer-reviewed articles suggested variants analyzed by IT to be causative/pathogenic/disease-causing, without confirmation of the predicted splicing effect^{101,135,137,150,160,167,178,205,210–218}. Other authors have qualified the interpretation of bioinformatic results with less certain terms (i.e. ‘suggest’ and ‘likely’ pathogenic)^{110,112,176,219–223}. Leclerc et al. (2002)¹⁶⁵ state that a predicted variant confirmed to affect splicing is likely deleterious, but could not be unequivocally shown to cause the observed phenotype. Although IT predictions can relate a sequence change to the resultant phenotype, caution should be exercised when deeming a predicted splicing variant as pathogenic in the absence of other functional evidence. The high level of concordance between IT mutation analysis and experimental findings indicates that this approach, in conjunction with other evidence, can be used to detect splicing effects, which may be used to explain disease phenotypes.

Comparisons to other software programs

There are now over a dozen other publically available splicing prediction tools, some using strategies similar (MaxEntScan [MES]) and others, which are quite different (NNsplice) that are compared with IT^224,225. Vreeswijk et al. (2009) assessed the applicability of different splice prediction programs to diagnose BRCA1/2 variants. These authors recommended that the outcome of 3 programs was sufficient for analysis, unless all three predictions were discordant from one another (2 for false positive predictions). Despite the obvious appeal of consensus between different analytical methods, a major caveat in using polling strategies for mutation assessment is that these approaches are prone to both systematic and sampling errors⁴⁰.

We summarize results of 36 publications that used both IT-based prediction tools and one or more alternate prediction tool (14 for 5´ and 3´ splicing, six for splicing regulatory proteins) to assess mutations^{23,39,97,99,103,111,114–117,123,130,132,141,147,156,158,165,166,171,179,185,189,197,210,218,226–235}. The analysis performed by the authors allowed us to compare the similarity of predictions to those programs and IT in Table 2a and Table 2b. Those most commonly used for 5´ and 3´ splice sites (NNsplice, MES, NG2, HSF and SSF) were highly concordant for natural sites (85.4% for donor and 77.6% for acceptor sites; Table 2a). Discordance of acceptor predictions may be due to methodologies that do not analyze the complete acceptor site (HSF analyzes only 14 intronic nucleotides upstream of acceptor splice sites)²³⁶. Some programs (SSF, HSF) exhibit greater concordance with IT for cryptic splice site prediction (96% for donor and 76.9% for acceptor sites). The level of discordance between IT and other commonly used software programs (59.5% for donor and 60% for acceptor sites) may be attributable to the empirically-derived scoring thresholds and the validation sets used to predict mutated splice sites. Models that are typically built (or trained) using known natural splice sites may be less sensitive for differentiating true cryptic splice forms from decoys in the genome, which tend to be weaker than natural splice sites. Tools are highly consistent when analyzing variants expected to be neutral with respect to splicing (100%; N = 71). Colombo et al. (2013) compared nine programs to evaluate accuracy in predicting mRNA splicing effects and reported that ASSA, along with HSF, demonstrated 100% informativeness and specificity.

ASSEDA has also been used to analyze RNA binding proteins that enhance or silence exon recognition (Table 2b). ESEfinder was used for 42.2% of these mutations in one or more regulatory binding sites^237,238. However, variants predicted by ESEfinder to have deleterious effects are discordant with some IT predictions (6 of 15; Table 2b). The discordance with ESEfinder may be associated with differences in the respective analytic methods, as several of the models (SF2/ASF, SC35, SRp40) used by ASSA and ESEfinder were created from the same source of experimental data^87,239. While the majority of the discordant results were cited in a single study¹¹⁴ (5/6 variants), the small size of the dataset (ranging from 28–34 sites) may artificially exacerbate differences between these results. In multiple instances, ASSA has been used to analyze SR proteins, but other programs were used to analyze 5´ and 3´ splice site mutations^23,99,115. This was surprising, since the donor and acceptor R_i values are generated by default by ASSA and ASSEDA. The advantage of performing both constitutive and regulatory splice site analysis with IT is that all results are reported on the same scale, and the strengths of all interactions, and effects of mutations are directly comparable to one another.

Other applications of information theory-based splice site analysis

The use of IT to analyze splicing is not limited to sequence variant analysis. The natural and alternative splicing of several genes have been characterized using this method^107,200,240. Khan et al. (2002) scanned all natural sites in the XPC gene and found a weak acceptor (-0.1 bits), and with RT-PCR found that this exon (exon 4) was skipped to a greater extent than another (exon 7), which possessed a strong acceptor, illustrating a relationship between the information content of a natural splice site and its level of alternative splicing. IT has also been used in genetic engineering in the design and alteration of binding sites, and has been used in the design of constructs for transgenic animal models^241–243. Thus, IT-based splice site analysis can be adapted for other important molecular genetic applications.

Guidelines for information theory-based splicing mutation analyses

Our comprehensive review of the use of IT in splicing mutation analysis has led us to propose general recommendations, which we formulate as guidelines. Adoption of these guidelines should ensure the accurate and comprehensive results from IT analyses of VUS and other pathogenic variants that alter mRNA splicing.

Data availability

F1000Research: Dataset 1. Dataset for mRNA splicing mutations in genetic disease, 10.5256/f1000research.5654.d38248²⁶²

Software availability