The mutational constraint spectrum quantified from variation in 141,456 humans

Key results

1. 01

   ShownIn a high-precision callset, LOFTEE preferentially removes common, artifact-enriched predicted loss-of-function (pLoF) calls yet spares rare and ClinVar-pathogenic ones, leaving 443,769 high-confidence pLoFs; selection holds methylated CpG stop-gained sites well below synonymous saturation.

   In two gold-standard samples, precision >99% (SNVs) and >98.5% (indels), recall >90% and >82%; de novo recall 97.3% (SNVs), 98% (indels) on 375 validated variants; 413,097 pLoFs on canonical transcripts of 16,694 genes; 1,815 genes (2,636 variants) curated as likely tolerant of biallelic loss; observed 52% of possible methylated CpG stop-gained vs 84.9% of synonymous variants. · A high-quality catalogue of variation; Identifying loss-of-function variants; Figure legends

2. 02

   ShownA refined mutation model enables a continuous, confidence-bounded constraint score (LOEUF, 90% upper confidence bound of the observed/expected pLoF ratio) that supersedes dichotomous pLI and reveals constraint as a continuum, with most genes under at least moderate selection against pLoF variants.

   Synonymous counts fit by the model with r = 0.979; median expected pLoFs per gene 17.9 (ExAC 13.2); 72.1% of genes have >10 expected pLoFs (ExAC 62.8%); median o/e 48%; 19,197 genes binned in deciles of ~1,920; first-decile aggregate o/e ~6%; zero observed pLoFs gives LOEUF 1.34 at 2 expected vs 0.03 at 100 expected. · The LoF intolerance of human genes; Figure legends

3. 03

   ShownLOEUF matches orthogonal data: haploinsufficient genes cluster at the constrained extreme, recessive genes sit mid-range and olfactory receptors are unconstrained, while constrained genes carry fewer rare exonic deletions and are more often essential in heterozygous mouse knockouts and human cells.

   6,735 rare exon-overlapping deletion structural variants from 6,749 genomes correlate with LOEUF (r = 0.13; p = 9.8 × 10^−68); 389 mouse heterozygous-lethal orthologs mean LOEUF 0.488 vs 0.962 for 18,808 others (p = 10^−78); 678 CRISPR-essential genes 0.63 vs 0.964 (p = 9 × 10^−71); 777 non-essential genes 1.34 vs 0.936 (p = 3 × 10^−92). · Validation of LoF intolerance score; Figure legends

4. 04

   ShownDe novo pLoF excess in neurodevelopmental cases concentrates in the most constrained LOEUF decile, schizophrenia rare-variant burden is higher in constrained genes, and common-variant heritability enrichment increases with constraint, most strongly for schizophrenia and educational attainment.

   Intellectual disability/developmental disorder (ID/DD; 5,305 cases vs 2,179 controls): de novo pLoF rate 15-fold higher in the most constrained decile (Fig. 5a legend: ~11-fold), 2.9-fold in the second; autism (6,430 patients): 4.4-fold; LD Score regression on 658 traits: enrichment scales with LOEUF decile, robust to gene size, brain expression and baseline annotations. · Constraint informs disease etiologies; Figure legends

5. 05

   ShownConstrained genes have more protein-interaction partners and broader, higher expression; within genes, constrained transcripts are more highly expressed; at matched sample size, African/African-American samples show lower LOEUF, in line with a larger effective population size.

   Protein-interaction degree vs LOEUF r = −0.14 (p = 1.7 × 10^−51, length-adjusted); breadth across 38 GTEx tissues r = −0.31 and expression level ρ = −0.28 (both p < 1 × 10^−100); 1,740 genes: constrained transcripts 6.34 TPM higher (p = 2.2 × 10^−14); 927 genes at n = 8,128: LOEUF 0.49 (African/African-American) vs 0.62 (European), p = 4.1 × 10^−14. · Biological properties of constraint; Figure legends