Computational Tools for Splicing Defect Prediction in Breast/Ovarian Cancer Genes: How Efficient Are They at Predicting RNA Alterations?

In silico tools for splicing defect prediction have a key role to assess the impact of variants of uncertain significance. Our aim was to evaluate the performance of a set of commonly used splicing in silico tools comparing the predictions against RNA in vitro results. This was done for natural splice sites of clinically relevant genes in hereditary breast/ovarian cancer (HBOC) and Lynch syndrome. A study divided into two stages was used to evaluate SSF-like, MaxEntScan, NNSplice, HSF, SPANR, and dbscSNV tools. A discovery dataset of 99 variants with unequivocal results of RNA in vitro studies, located in the 10 exonic and 20 intronic nucleotides adjacent to exon-intron boundaries of BRCA1, BRCA2, MLH1, MSH2, MSH6, PMS2, ATM, BRIP1, CDH1 , PALB2, PTEN, RAD51D, STK11, and TP53, was collected from four Spanish cancer genetic laboratories. The best stand-alone predictors or combinations were validated with a set of 346 variants in the same genes with clear splicing outcomes reported in the literature. Sensitivity, specificity, accuracy, negative predictive value (NPV) and Mathews Coefficient Correlation (MCC) scores were used to measure the performance. The discovery stage showed that HSF and SSF-like were the most accurate for variants at the donor and acceptor region, respectively. The further combination analysis revealed that HSF, HSF+SSF-like or HSF+SSF-like+MES achieved a high performance for predicting the disruption of donor sites, and SSFlike or a sequential combination of MES and SSF-like for predicting disruption of acceptor sites. The performance confirmation of these last results with the validation dataset, indicated that the highest sensitivity, accuracy, and NPV (99.44%, 99.44%, and 96.88, respectively) were attained with HSF+SSF-like or HSF+SSF-like+MES for donor sites and SSF-like (92.63%, 92.65%, and 84.44, respectively) for acceptor sites. We provide recommendations for combining algorithms to conduct in silico splicing analysis that achieved a high performance. The high NPV obtained allows to select the variants in which the study by in vitro RNA analysis is mandatory against those with a negligible probability of being spliceogenic. Our study also shows that the performance of each specific predictor varies depending on whether the natural splicing sites are donors or acceptors.