LAUSR

Identification of a pathogenic variant in NF1 is diagnostic for neurofibromatosis, but is often impossible at the moment of variant detection due to many factors including allelic heterogeneity, sequence homology, and the lack of functional assays. Computational tools may aid in interpretation but are not established for NF1. Here, we optimized our random forest-based predictor RENOVO for NF1 variant interpretation. RENOVO was developed using an approach of “database archaeology”: by comparing versions of ClinVar over the years, we defined “stable” variants that maintained the same pathogenic/likely pathogenic/benign/likely benign (P/LP/B/LB) classification over time (n = 3579, the training set), and “unstable” variants that were initially classified as Variants of Unknown Significance (VUS) but were subsequently reclassified as P/LP/B/LB (n = 57, the test set). This approach allows to retrospectively measure accuracy on prediction with insufficient information, reproducing the scenario of maximal clinical utility. We further validated performance on: (i) validation set 1: 100 NF1 variants classified as VUS at the time of RENOVO development and subsequently reclassified as P/LP/B/LB in ClinVar; (ii) validation set 2: 15 de novo variants discovered in a prospective clinical cohort and subsequently reclassified per ACMG criteria. RENOVO obtained consistently high accuracy on all datasets: 98.6% on the training test, 96.5% in the test set, 82% in validation set 1 (but 96.2% for missense variants) and 93.7% on validation set 2. In conclusion, RENOVO-NF1 accurately interprets NF1 variants for which information at the time of detection is insufficient for ACMG classification and may overcome diagnostic challenges in neurofibromatosis.