The Performance of In Silico Prediction Tools for Variant Curation in a Panel of Cancer Genes.
Summary
This study assesses the performance of in silico prediction tools for genetic variant curation within a panel of cancer predisposition genes. Researchers applied ClinGen SVI Working Group recommended thresholds and AlphaMissense predictions to variants in genes such as BRCA1, BRCA2, TP53, TERT, and ATM, which have established pathogenicity or benignity. The findings indicated insufficient sensitivity for pathogenic TERT variants and benign TP53 variants. The study emphasizes that the performance of these tools can be gene-specific and is influenced by their training datasets. Consequently, it is crucial to validate these tools for individual genes, especially for missense variants.
Analysis
This bioinformatic study critically evaluates the performance of in silico prediction tools for genetic variant classification, a crucial aspect of variant curation. It highlights that, despite often using similar information sources and training datasets, these tools can exhibit variable and gene-specific performance. Researchers applied ClinGen SVI Working Group recommended thresholds and AlphaMissense predictions, a deep learning model, to assess missense variants in cancer predisposition genes. The results demonstrated sensitivities below 65% for pathogenic TERT variants and below or equal to 81% for benign TP53 variants, indicating significant limitations. AlphaMissense outperformed other tools for TP53 but did not improve accuracy for TERT. The study emphasizes the need for quantitative, gene-specific validation of in silico scores, particularly for missense variants, to improve the reliability of classifications and potentially guide clinical variant review workflows.