In-silico Analysis of NF1 Missense Variants in ClinVar: Translating Variant Predictions into Variant Interpretation and Classification

Abstract

Background: With the advent of next-generation sequencing in genetic testing, predicting the pathogenicity of missense variants represents a major challenge potentially leading to misdiagnoses in the clinical setting. In neurofibromatosis type 1 (NF1), where clinical criteria for diagnosis may not be fully present until late infancy, correct assessment of variant pathogenicity is fundamental for appropriate patients' management. Methods: Here, we analyzed three different computational methods, VEST3, REVEL and ClinPred, and after extracting predictions scores for 1585 NF1 missense variants listed in ClinVar, evaluated their performances and the score distribution throughout the neurofibromin protein. Results: For all the three methods, no significant differences were present between the scores of "likely benign", "benign", and "likely pathogenic", "pathogenic" variants that were consequently collapsed into a single category. The cutoff values for pathogenicity were significantly different for the three methods and among benign and pathogenic variants for all methods. After training five different models with a subset of benign and pathogenic variants, we could reclassify variants in three sharply separated categories. Conclusions: The recently developed metapredictors, which integrate information from multiple components, after gene-specific fine-tuning, could represent useful tools for variant interpretation, particularly in genetic diseases where a clinical diagnosis can be difficult.

Keywords: ClinPred; NF1; REVEL; VEST3; missense variants; variant interpretation.

Figure 1

Prediction scores of the three computational methods for the LB (LEANING BENIGN), VUS (variants of uncertain significance), CI (conflicting interpretations), and LP (LEANING PATHOGENIC) NF1 missense variants. Median and 95% confidence intervals are reported for each category. The P value of non-parametric ANOVA is reported for each predictor (Kruskal-Wallis test).

Figure 2

Overall performances of the three predictors on the NF1 missense variants. The area under the curve (AUC) is reported for each predictor.

Figure 3

Prediction score values in the principal NF1 functional domains annotated at InterPRO (RAS-GTPase aa 1210–1549; CRAL-TRIO LIPID BINDING DOMAIN aa 1581–1726; PH-LIKE Pleckstrin homology domain aa 1727–1837; Armadillo type fold aa 1849–2676; NO FUNCTION all amino acid not comprised in the previous domains). Medians with 95% confidence intervals are reported. The P value of non-parametric ANOVA is reported for each predictor (Kruskal-Wallis test).

Figure 4

Prediction score values after reclassification of NF1 missense variants in the three categories LEANING BENIGN, VUS, and LEANING PATHOGENIC. For each category, median and 95% confidence intervals are reported.

Figure 5

REVEL prediction score values distribution throughout the NF1 protein. Above, the location of the four NF1 InterPRO domains (RAS-GTPase aa 1210–1549; CRAL-TRIO LIPID BINDING DOMAIN aa 1581–1726; PH-LIKE Pleckstrin homology domain aa 1727–1837; Armadillo type fold aa 1849–2676). Green triangles are LEANING BENIGN scores, yellow diamonds VUS scores, red circles LEANING PATHOGENIC scores. Shaded in red are regions with clustering of LEANING PATHOGENIC variants with high prediction scores (aa 1200–1440, and 1725–1250), while green shading indicates regions where no or few LEANING PATHOGENIC variants are present (aa 2530–2849 and 600–700, respectively).