Extended Data Fig. 8: Power of constraint detection.
From: A genomic mutational constraint map using variation in 76,156 human genomes
a,b, The sample size required for well-powered non-coding constraint detection. The percentage of non-coding regions powered to detect constraint (Gnocchi ≥ 4) at a 1kb (a) and 100bp (b) scale under varying levels of selection (depletion of variation) is shown as a function of log-scaled sample size. Lighter color indicates milder deletion of variation (weaker selection), which requires a larger sample size to detect constraint; the grey dashed vertical line indicates the current sample size of 76,156 genomes. Dotted curves (left to right) benchmark the 95th, 90th, and 50th percentile of depletion of variation observed in coding exons of similar size. The number of samples required to obtain an 80% detection power is labeled at corresponding benchmarks. c, AUCs of Gnocchi scores computed on different window sizes in identifying putative functional non-coding variants. 1kb (used in this study) presents the optimal window size with high performance while maintaining reasonable resolution. d, AUCs of Gnocchi scores computed from different subsets of gnomAD in identifying putative functional non-coding variants. While with an equal sample size, the downsampled dataset with diverse ancestries presents higher performance than the Non-Finnish European (NFE)-only dataset.
