In this study we molecularly characterized 24 UVs in the BRCA1
genes with potential effect at mRNA level. A total of 19 spliceogenic mutations were identified. These included all 11 variants located at invariant dinucleotides at the 5′ and 3′ intron ends, as expected, and 8 out of 13 UVs in less conserved positions of splicing regions. Sixteen mutations led to the synthesis of aberrant transcripts containing PTCs, 2 (BRCA1
c.134+3_134+6delAAGT and c.212G>A) to the up-regulation of naturally occurring PTC-containing isoforms, and one (BRCA2
c.8954−1_8955delGTTinsAA) to the in-frame deletion of 51 nucleotides at the 5′-end of exon 23, within the region coding for the DBD, a critical functional domain of the BRCA2 protein. Functional analyses revealed that the latter alteration caused the loss in the mutant protein of the ability to bind the partner protein DSS1 and ssDNA. Based on these observations, all spliceogenic mutations were classified as pathogenic or likely pathogenic, according to current guidelines for the interpretation of the results of in vitro
splicing analyses 
. These guidelines adopt the 5-class classification criteria proposed by Plon et al. 
, and classify spliceogenic mutations as of class 5 (probability of being pathogenic >99%) or of class 4 (probability of being pathogenic
95%–99%), depending on the relative amount of aberrant transcripts. Following this scheme, 15 mutations for which only expression of aberrant transcripts was observed, were considered of class 5, whereas the 2 mutations that maintained the ability to express normal in addition to aberrant transcripts were provisionally categorized as of class 4. To assess the relative amount of normal and aberrant transcripts expressed by these alleles, additional quantitative analyses are required. For the remaining 2 spliceogenic mutations the distinction in either class 4 or 5 could not be made due to the inability to assess allelic specific expression of the normal mRNA ( and ).
It must be remarked that a recent study, based on the analysis of LCL mRNA, reported 4 spliceogenic BRCA gene mutations introducing PTCs that were classified as uncertain or likely neutral by multifactorial likelihood analyses 
. Although it is likely, as the authors of the study reported, that this discrepancy depended on a reduced performance of the multifactorial analyses, due either to a paucity of information and/or the use of non specific prior probability of pathogenicity for the variants analyzed, these data suggest that the mutation effect detected in blood cells may not necessarily reflect that occurring in at-risk tissues, such as breast and ovarian epithelium. Another possible explanation for the inconsistency between the outcome of in vitro
splicing analyses and that of multifactorial models is the occurrence of spliceogenic mutations that maintain the ability to synthesize a normal in addition to an aberrant mRNA 
. These mutations may have an impact on cancer risk different from that of fully inactivating alterations. As mentioned above, we detected 2 such mutations (BRCA2 c.476−2A>G and c.8755−1G>A). However, quantitative analyses indicated that in both cases the contribution of the mutated allele to the total amount of normal mRNA was small. Assuming that most normal mRNA transcripts derive from the wild-type allele, we found that only approximately 10% originated from the mutated allele. Both the above mutations were detected in a single family each, and no sufficient data were available for a reliable classification using multifactorial models. It is interesting to note that, although splice site mutations producing both normal and aberrant transcript would be expected to be prevalently, if not exclusively, located in less conserved regions, both identified ‘leaky’ mutations were localized at the nearly invariant dinucleotides at the 5′ and 3′ intron ends. However, we could not formally rule out that expression of normal transcripts occurred also for other examined spliceogenic mutations, due to the relatively limited sensitivity of sequencing analyses in assessing allelic specific expression.
Of the 5 non spliceogenic variants, 2 were intronic and 3 introduced missense changes (p.Gly1366Asp and p.Asp1778Gly in BRCA1
and p.Pro3039Leu in BRCA2
). For all the latter substitutions, the Align-GVGD algorithm 
predicted a prior probability of pathogenicity of 1%. Therefore, following current guidelines 
, all non spliceogenic variants were classified as likely non pathogenic (class 2, probability of pathogenicity
0,1%–4,9%). This classification was in agreement with additional evidence from previous studies. In particular, BRCA1
p.Asp1778Gly located in the C-terminus transcriptional activation BRCT domain of the gene was predicted as neutral by 3 computational supervised learning algorithms based on features describing evolutionary conservation, impact of mutation on protein structure, and amino acid residue 
. This prediction has been recently confirmed by a comprehensive analysis using biochemical and cell-based transcriptional assays 
. In addition, the presence of the variant was not detected in the proband’s affected mother. Finally, the BRCA2
p.Pro3039Leu has been classified as neutral using a bioinformatics approach integrating information about protein sequence, conservation and structure in a likelihood ratio 
For 8 of the 13 variants that had been already investigated at the cDNA level, our findings were consistent with those of earlier reports, while for the remaining 5 variants (all spliceogenic) the observed transcript patterns differed from those described by previous studies (Table S6
). This was possibly due to the different experimental protocols that were used, suggesting that differences may occur in the ability of in vitro
analyses to detect mRNA transcripts, particularly those expressed at low level. Another potential source of inconsistency might be the use of different types of biological samples. Although no discrepancies emerged in the classification of the examined variants as spliceogenic or non-spliceogenic when comparing our data with those of previous studies, our observations emphasize the need of developing standardized methods for in vitro
characterization of UVs through gene transcript analyses, particularly when the outcomes of these analyses are used to counsel carriers of variants at splice sites.
In previous studies, bioinformatics analyses have been proposed as a first step to select variants predicted to affect mRNA splicing and, in particular, those located outside the nearly invariant dinucleotides at the 5′ and 3′ intron ends 
. To further verify the reliability and the usefulness of these programs for a priori
selection of spliceogenic UVs, we compared the computational splice-site predictions obtained from 9 commonly used programs with the experimental results derived from cDNA analyses. Consistent with previous reports 
, we found that most tested programs showed an incomplete informativeness, i.e. were not able to recognize all natural splice sites affected by the variants under analyses. Thus, the effect of nucleotide substitutions at these sites could not be subsequently computed, limiting the usefulness of these programs. In our analysis only 3 programs (MES, HSF and ASSA) exhibited 100% informativeness.
While the performance of a selective process is usually measured in terms of accuracy, i.e., the optimal compromise between sensitivity and specificity, it must be considered that UV classification in cancer predisposing genes is manly carried out for clinical purposes, i.e., to define risk estimates in carriers of such variants 
. Along this line, we reasoned that a mandatory pre-requirement of the procedures for BRCA1
variant selection for transcript characterization is 100% sensitivity. Therefore, in our study, we considered that a spliceogenic effect was predicted when an in silico
analysis measured a relative decrease of the SSPS/Ri values (of the natural splice site in the mutated compared to the wild-type sequence) higher than the lowest detected in the presence of an in vitro
verified spliceogenic mutation. Based on this assumption, we eventually verified the specificity, measured as the rate of false positive predictions, of each program. In our hands, this was found to be equal to 100%, i.e. no false positive prediction, for 5 programs: SSF, GS, HSF, SV and ASSA. In a general evaluation, the programs that performed better were HSF and ASSA, the only exhibiting 100% informativeness and 100% specificity.
The knowledge of the precise nature of aberrant transcripts is crucial for the assessment of the pathogenicity of spliceogenic mutations. For example, variant alleles producing transcripts carrying in-frame deletions not disrupting known functional domains are currently classified as of unknown clinical significance 
and some of them might actually be clinically neutral. This is supported by the observation that the BRCA2
c.6853A>G variant, resulting in increased exclusion of exon 12, is phenotypically indistinguishable from an allele with exon 12 deleted and wild-type BRCA2
in functional analyses using allelic complementation in Brca2-null mouse embryonic stem cells 
. Therefore, it is important to ascertain whether a spliceogenic mutation, in addition to abolishing the recognition of a natural splice site, leads to the creation of novel splice sites or the activation of cryptic ones. As already discussed, in this study the usage of alternative splice sites were observed in a relevant fraction of ascertained spliceogenic variants (10/19
42%). We sought to verify to which extent computational programs are able to predict such occurrences. We found that only 3 programs (MES, HSF and ASSA) recognized all experimentally ascertained alternative splice sites. However, these programs also detected other putative cryptic splice sites in the vicinity of the abolished naturally-occurring splice sites and, consistent with a previous report 
, we were unable to derive simple criteria, based on the outcomes of the in silico
analyses, for the prediction of the specific alternatively used splice sites. On the other hand, it is also possible that some of the cryptic sites predicted in silico
could be activated in mutant samples, but the corresponding aberrant transcripts were not observed in vitro
due to a limited sensitivity of the detection method we used.
Our study provides further evidences that in silico
tools may be used for the ascertainment of splice site variants to be submitted to in vitro
analyses. We performed a comparative analysis of 9 freely-available computational programs, and found that those that performed better in identifying variants affecting RNA splicing, under our analytical scheme, were HSF and ASSA. However, in vitro
analyses remain mandatory for the characterization of the exact nature of aberrant transcripts. Wider surveys within the frame of large collaborative consortia, such as the recently established ‘Evidence-based Network for the Interpretation of Germline Mutant Alleles’ (ENIGMA) 
, are looked-for, in order to define the more effective protocols for the use of bioinformatics analyses in the ascertainment of spliceogenic mutations.