Episignatures in practice: independent evaluation of published episignatures for the molecular diagnostics of ten neurodevelopmental disorders

Abstract

Variants of uncertain significance (VUS) are a significant issue for the molecular diagnosis of rare diseases. The publication of episignatures as effective biomarkers of certain Mendelian neurodevelopmental disorders has raised hopes to help classify VUS. However, prediction abilities of most published episignatures have not been independently investigated yet, which is a prerequisite for an informed and rigorous use in a diagnostic setting. We generated DNA methylation data from 101 carriers of (likely) pathogenic variants in ten different genes, 57 VUS carriers, and 25 healthy controls. Combining published episignature information and new validation data with a k-nearest-neighbour classifier within a leave-one-out scheme, we provide unbiased specificity and sensitivity estimates for each of the signatures. Our procedure reached 100% specificity, but the sensitivities unexpectedly spanned a very large spectrum. While ATRX, DNMT3A, KMT2D, and NSD1 signatures displayed a 100% sensitivity, CREBBP-RSTS and one of the CHD8 signatures reached <40% sensitivity on our dataset. Remaining Cornelia de Lange syndrome, KMT2A, KDM5C and CHD7 signatures reached 70-100% sensitivity at best with unstable performances, suffering from heterogeneous methylation profiles among cases and rare discordant samples. Our results call for cautiousness and demonstrate that episignatures do not perform equally well. Some signatures are ready for confident use in a diagnostic setting. Yet, it is imperative to characterise the actual validity perimeter and interpretation of each episignature with the help of larger validation sample sizes and in a broader set of episignatures.

Fig. 1. Episignature predictive performances.

Panel (a) and (b) display sensitivities and specificities, respectively, according to various clustering parameters. Error bars indicate 95% confidence intervals based on a binomial distribution.

Fig. 2. Visual inspection of three typical examples of robust, unstable and weak signatures.

Respectively for (A) Kabuki/KMT2D, (B) MXRSCJ/KDM5C and (C) RSTS/CREBBP, each panel displays the heatmap with simultaneous hierarchical clustering of validation and testing samples on the left. Blue indicates hypo-methylated positions while yellow indicates hyper methylated positions. First principal components are represented on the right. Percentage of explained variance is added in parenthesis on each axis.