The performance of genetic-constraint metrics varies significantly across the human noncoding genome

A longstanding goal in human genetics is to prioritize noncoding loci that, when disrupted, lead to developmental disorders and other Mendelian traits. In pursuit of this goal, multiple metrics have been developed to distinguish neutrally evolving sequences from those subjected to purifying selection. These metrics are commonly evaluated genome-wide, e.g., by computing a precision-recall curve on windows tiling the entire noncoding genome. Here, we identify parts of the noncoding genome where these metrics are distorted due to "bias" in the underlying models of neutral genetic variation and underperform due to a low "signal-to-noise ratio" in the genetic data. The most extreme effects are found for Gnocchi (Chen et al. 2024), the performance of which declines as GC content increases. We trace the source of the bias in Gnocchi to a mismatch between the training set of de novo mutations (DNMs) used to adjust expected SNP counts and the genomic windows on which those expected counts are computed. Resolving that mismatch produces a more calibrated version of Gnocchi, significantly improving its interpretability in applications. Looking forward, we suggest annotating constraint scores of noncoding genomic intervals with robust measures of the bias of the corresponding model, allowing users to gauge confidence in those scores.