Fig. 3
a–d Evaluating the sensitivity and specificity of the 16p11.2del and CHD8+/− DNAm signatures. The classification models (details in Ref. [15]) are represented by the following classification plots: a The classification model based on the 16p11.2del DNAm signature sites (115 sites; q < 0.05, |Δβ| ≥ 5%) accurately classified CHD8+/− signature cases, and independent test cases consisting of CHD8 sequence variants (orange box with cross) and heterogeneous ASD cases (open gray triangle) as more similar to controls (green C). The model also accurately classified 7 of 9 independent 16p11.2del test cases (solid purple circle) as more similar to 16p11.2del signature cases (open purple circle), demonstrating 100% sensitivity. Two variants (arrows) correctly received negative classification scores: hatched arrow indicates a mosaic 16p11.2del case (2-0088-003), solid arrow indicates a 16p11.2del distal to and not overlapping the 600 kb typical deletion region (1-0616-003). b The 16p11.2del DNAm signature was tested on an independent set of compiled blood control DNAm data (n = 162) extracted from the Gene Expression Omnibus (GEO) Database. All GEO controls (turquoise C) were properly classified with experimental controls (green C), as distinct from 16p11.2del signature training cases (open purple circle), demonstrating 100% specificity. c The classification model based on the CHD8+/− DNAm signature sites for classification (103 sites; q < 0.01, |Δβ| ≥ 5%) accurately classified 16p11.2del signature cases and independent test cases consisting of 16p11.2del variants (purple circle with cross) and heterogeneous ASD cases (gray triangle) as more similar to controls (green C). The model also accurately classified 1 of 9 independent CHD8 test cases consisting of sequence variants (solid orange square) as more similar to CHD8+/− signature training cases (open orange square) (unable to report sensitivity with one positive case, indicated by arrow). Eight variants correctly received negative classification scores (7 missense, 1 with an in-frame deletion in the last exon of CHD8, not predicted pathogenic). Classifications are in agreement with in silico predictions. d The CHD8+/− DNAm signature was tested on the same GEO controls (turquoise C) as in b; all but one GEO control were properly classified with experimental controls (green C), as distinct from CHD8+/− signature cases (open orange square), demonstrating >99.3% specificity