Gene-specific application of computational prediction tools aids the classification of rare missense variants in the diagnosis of hereditary endocrine tumour syndromes

Introduction: The successful implementation of clinical genetic testing relies on accurate variant interpretation, as misclassification can result in significant harm to the patient and wider family. Missense single nucleotide variants (SNVs) pose a particular challenge, with current interpretation methods often unable to differentiate pathogenic variants from rare neutral variants, resulting in high numbers of variants of uncertain significance (VUS), and diagnostic uncertainty. In silico tools are frequently used during interpretation, but established methods lack specificity and are inconsistently applied. Here, we assess the utility of state-of-the-art computational tools in the classification of missense SNVs in five hereditary endocrine tumour genes (MEN1, NF1, RET, SDHB, VHL).

Methods: Fourteen recently reported computational variant prediction tools based on DNA sequence (n=8) or protein structure (n=6) were used to assess four groups of missense SNVs (‘benign’, ‘pathogenic, ‘VUS’ and ‘GnomAD rare’) identified from publicly available repositories (ClinVar, LOVD, GnomAD), totalling >7,400 unique variants. Relevant protein structures were obtained from Protein Data Bank and AlphaFold2. The performance of tools was assessed using multiple statistical metrics.

Results: The majority of sequence-based tools (e.g. ClinPred, VARITY, MutPred2) demonstrated good performance at standardised ‘pathogenicity’ cut-offs for differentiating known benign and pathogenic variants (e.g. sensitivity ~70-100%, specificity ~60-90%) and generally outperformed structure-based tools (Rhapsody and SNPMuSiC performing well for specific genes). However, all tools lacked discriminatory ability when classifying ‘VUS’ and ‘GnomAD rare’ SNVs with high proportions of deleterious variants predicted. The development of gene-specific ‘pathogenicity’ cut-offs for each tool improved specificity and the stratification of variant groups, which was further enhanced when concordance between combinations of the highest-performing tools was assessed.

Conclusions: Here, we demonstrate the utility of recently described computational variant prediction tools when applied to several hereditary endocrine tumour genes and advocate a gene-specific approach that incorporates combinations of tools to optimise specificity and clinical utility.