We found that population-focused methods do best for identifying highly impactful variants (de novo’s in individuals with developmental disorders for instance), while the deep learning methods are better at prioritizing inherited variation in biobanks

Interestingly, these scores also provide additional insight into genes regulated by these regions, even those underpowered by previous constraint metrics:

Gnocchi extends our constraint metrics to the non-coding genome, highlighting for instance, disease-associated non-coding CNVs

We built a new metric we called gnocchi (genomic non-coding constraint of haploinsufficient variation), building on methods that find depletions of variation (natural selection), which we show can prioritize functional variation

CHARR operates only on homozygous alternate sites and scales very well (“cost per 1M samples” might be my new favorite metric):