The Performance of In Silico Prediction Tools for Variant Curation in a Panel of Cancer Genes

Rare single base pair changes in genes are an important cause of disease, as they can reside in key regions of the gene influencing biological function by impacting the protein conformation and protein interactions. Generation of the necessary experimental evidence to define the outcome of the presence of these gene variants is time-consuming and costly. These challenges have led to the development of a plethora of in silico prediction tools. These tools frequently use similar sources of information and are trained on overlapping multigene 'truth' datasets. However, frequently there has been no quantitative validation of the performance of these in silico tools for individual genes. Here, we have applied the ClinGen Sequence Variant Interpretation Working Group's recommended in silico score thresholds and AlphaMissense predictions to a set of predisposition gene variants with established pathogenicity/benignity. Of the genes assessed (BRCA1, BRCA2, TP53, TERT and ATM), when recommended thresholds were used, in silico tool predictions showed inferior sensitivity (< 65%) for pathogenic TERT variants and inferior sensitivity (≤ 81%) for benign TP53 variants. AlphaMissense outperformed the other tools for TP53 but did not improve predictive accuracy for TERT variants. This validation study highlights that in silico tool performance can be gene-specific and is dependent on the 'training set' on which the algorithm is built. Where there are sufficient numbers of established benign and pathogenic missense variants based on clinical and functional evidence, the use of in silico tool scores should be validated for individual genes. For genes where this is not possible and gene-agnostic in silico score cut-offs are used, consideration of missense variant-protein structural impact relationships is suggested.