henhao
Sign in
GeneticsNature
Peer-reviewed journal article

A genome-wide mutational constraint map quantified from variation in 76,156 human genomes

Siwei CheniD, Laurent C. FrancioliiD, Julia K. GoodrichiD, Ryan L. CollinsiD, Masahiro KanaiiD, Qingbo S. WangiD, Jessica AlföldiiD, Nicholas A. Watts, Christopher Vittal, Laura Doyle Gauthier, Timothy PoterbaiD, Michael W. WilsoniD, Yekaterina Tarasova, William PhuiD, Riley H. Grant, Mary T. Yohannes, Zan Koenig, Yossi FarjouniD, Eric BanksiD, Stacey DonnellyiD, Stacey GabrieliD, Namrata GuptaiD, Steven Ferriera, Charlotte Tolonen, Sam Novod, , David B. GoldsteiniD

Abstract

The depletion of disruptive variation caused by purifying natural selection (constraint) has been widely used to investigate protein-coding genes underlying human disorders, but attempts to assess constraint for non-protein-coding regions have proven more difficult. Here we aggregate, process, and release a dataset of 76,156 human genomes from the Genome Aggregation Database (gnomAD), the largest public open-access human genome reference dataset, and use this dataset to build a mutational constraint map for the whole genome. We present a refined mutational model that incorporates local sequence context and regional genomic features to detect depletions of variation across the genome. As expected, proteincoding sequences overall are under stronger constraint than non-coding regions. Within the non-coding genome, constrained regions are enriched for known regulatory elements and variants implicated in complex human diseases and traits, facilitating the triangulation of biological annotation, disease association, and natural selection to non-coding DNA analysis. More constrained regulatory elements tend to regulate more constrained protein-coding genes, while non-coding constraint captures additional functional information underrecognized by gene constraint metrics. We demonstrate that this genome-wide constraint map provides an effective approach for characterizing the non-coding genome and improving the identification and interpretation of functional human genetic variation.

PDFDOIData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
1,735
FWCI
287.35
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (10 Oct 2022) · CC BY 4.0

Rare-variant depletion in 1kb windows of 76,156 gnomAD genomes, versus feature-adjusted mutation expectations, marks regulatory elements and outperforms seven alternative scores at flagging functional non-coding variants.

Gene constraint metrics cover protein-coding sequence (under 2% of the genome), while existing non-coding approaches ignore regional mutation-rate determinants beyond flanking bases, rely on curated disease-variant training sets, or lack power for recent human-lineage selection. Pairing 76,156 diverse gnomAD genomes with a mutation model jointly fitting trinucleotide context, CpG methylation and regional features, this work quantifies constraint genome-wide in 1kb windows.

Key results

  1. ShownWindows with protein-coding bases are far more depleted of rare variation than purely non-coding windows, yet a small fraction of non-coding windows reach exon-like constraint; the score also correlates with a separate structural-variant (SV) constraint metric.

    Median Z 1.47 for windows with any coding bases vs 0.08 for non-coding windows (Wilcoxon P<10^−200); concatenated exons median 3.17; 3.12% and 0.05% of non-coding windows match the 50th and 90th exonic percentiles; adjusted proportion of singletons (APS) tracks Z across 116,184 SVs (beta=0.01, P=4.3×10^−65). · Quantifying mutational constraint across the…; Correlation between constraint Z score and APS

  2. ShownThe most constrained non-coding windows (Z≥4) are enriched for ENCODE candidate cis-regulatory elements (cCREs), most strongly promoter-like ones, and for genome-wide association study (GWAS) and fine-mapped trait variants, whereas CTCF-only elements (a negative control) are not.

    cCRE odds ratio (OR)=2.77 (Fisher P<10^−200): promoter-like 7.28, proximal/distal enhancer-like 4.35/2.14, CTCF-only 0.82; FANTOM enhancers 3.58, super enhancers 3.41; GWAS Catalog 837/19,471 constrained windows vs 51,430/1,843,559 genome-wide (OR=1.57, P=2.5×10^−32), replicated OR=2.08, fine-mapped OR=3.24 (P=3.0×10^−10). · Investigating genomic properties of non-coding…; Correlation between constraint Z score and…

  3. ShownAmong eight genome-wide scores, constraint Z gives the highest area under the curve (AUC) for separating putative functional non-coding variants from frequency-matched background; conservation scores do worse, and context-only rebuilds of Z lose performance.

    Outperformed Orion, CDTS, gwRVIS, DR, phyloP, phastCons and GERP on 9,229 replicated GWAS Catalog, 2,191 fine-mapped, 140 high-confidence and 288 ClinVar pathogenic variants; conservation scores correlate with Z at Spearman 0.017-0.19; 204 variants unique to Z's 99th percentile show 3.45-fold recombination and 2.74-fold methylation vs other positives. · Comparing constraint Z score to other…; Comparison of constraint Z score and other…

  4. ShownCopy number variants (CNVs) in developmental delay (DD) cases overlap constrained non-coding windows far more often than control CNVs, an excess that survives (attenuated) adjustment for size and gene content and restriction to non-coding CNVs; locus examples highlight constrained critical regions.

    CNVs touching Z≥4 non-coding windows: 7,239/17,004 (42.6%) in cases vs 10,403/83,526 (12.5%) in controls (OR=5.21, Fisher P<10^−200); 18/19 known pathogenic DD loci and 85.5% of ClinVar pathogenic CNVs; log[OR]=1.06 conditioning on gene constraint (P<10^−100); gene-free IHH duplication region median Z=2.52 (P=1.3×10^−3), top window Z=4.12 at its major enhancer. · Exploring non-coding dosage sensitivity in the…; Correlation between constraint Z score and…; Figure legends

  5. ShownEnhancer constraint is higher for constrained than less-constrained genes, and for genes with ≤5 expected loss-of-function (LoF) variants than for powered unconstrained genes; adding enhancer Z to LOEUF (LoF observed/expected upper bound fraction) improves classification of underpowered genes.

    Median enhancer Z 2.71 vs 1.99 (Wilcoxon P=1.3×10^−26) and 2.64 vs 2.27 (P=9.8×10^−4) for the two contrasts; LOEUF+enhancer Z AUC 0.80 vs 0.73 for LOEUF alone on 77 held-out underpowered genes (bootstrap P=0.03); enhancer constraint predicts constrained genes conditional on LOEUF (P=7.4×10^−11); ASCL2 enhancer Z=5.58. · Leveraging non-coding constraint to improve gene…; Incorporation of non-coding constraint of…

Methods, data and reproducibility

Non-overlapping 1kb windows are scored by a signed Z contrasting observed rare variants with expectations from trinucleotide mutation rates (estimated in 1,000 downsampled genomes; CpG sites stratified by germline methylation) scaled by per-context effects of 13 regional features fitted on de novo mutations by logistic regression.

  • gnomAD v3.1.2 whole-genome callset on GRCh38, joint-called from 153,030 genomes and filtered to unrelated, consented samples; 36,811 (48.3%) of non-European ancestry · human · n = 76,156 genomes; 644,267,978 short nuclear variants
  • De novo mutations from two family-based whole-genome sequencing studies, contrasted with random background sites to fit regional mutation-rate effects · human · n = 413,304 de novo mutations; 4,104,879 background sites
  • Regulatory annotations: ENCODE cCREs (promoter-like, enhancer-like, CTCF-only), FANTOM5 enhancers, SEdb super enhancers, plus GENCODE miRNA and FANTOM6 lncRNA elements · human · n = 926,535 cCREs; 63,285 enhancers; 331,601 super enhancers
  • Functional non-coding variant sets: GWAS Catalog hits (P≤5.0×10^−8), biobank fine-mapped variants (posterior inclusion probability ≥0.9), ClinVar pathogenic variants · human · n = 111,308 GWAS (9,229 replicated); 2,191 fine-mapped; 288 ClinVar
  • Copy number variant morbidity map contrasting developmental delay and birth-defect cases with healthy controls · human · n = 100,530 CNVs (17,004 in cases, 83,526 in controls)
  • UK Biobank fine-mapping results (SuSiE, uniform prior) reweighted using normalized constraint Z as a functional prior · human (UK Biobank) · n = 119 traits; 60,121 variants in 6,592 credible set-trait pairs
  • Roadmap Epigenomics enhancer-gene links merged across reference epigenomes, plus GTEx median expression for 11 matched tissue types · human (Roadmap Epigenomics, GTEx) · n = 127 reference epigenomes; 11 matched tissues
  • Gene sets: constrained (ClinGen haploinsufficient, MGI essential, OMIM autosomal dominant, LOEUF first decile) vs unconstrained (olfactory receptors, LOEUF last decile) · human (MGI essential genes mapped to human orthologs) · n = 3,220 ClinGen/MGI/OMIM; 1,920 LOEUF first decile; 356 olfactory; 189 last decile
Methods & tools
GATK Best Practices pipeline, Hail (joint variant calling), CrossMap (liftover to GRCh38), bedtools, bigWigAverageOverBed, SuSiE (prior UK Biobank fine-mapping results reweighted with a constraint-based functional prior), Logistic regression with PCA-decorrelated genomic features, Dominance analysis (relative predictor contribution), Orion, CDTS, gwRVIS and DR (population-genetic constraint comparators), phyloP, phastCons and GERP (conservation comparators), LOEUF gene constraint metric, APS structural-variant constraint score
Data & accessions
https://gnomad.broadinstitute.org/
Reagents & environment
GRCh38; gnomAD v3.1.2 harmonized via GATK Best Practices and joint-called in Hail; CrossMap liftover; bedtools and bigWigAverageOverBed for feature values; GTEx RNASeQCv1.1.9 expression; regression coefficients shared as pickle files.
Reproducibility gaps
Code availability is deferred to Supplementary Information with no repository URL in the main text; sample inclusion required consent for aggregate variant data sharing, so released data are summary-level; software versions are largely unstated.
Limitations (6)
  • Non-coding sequence lacks a consequence framework analogous to synonymous versus missense or LoF classes, and per-variant expression effects cannot yet be predicted, so constraint cannot be resolved by consequence and selection strength may differ by mechanism. · stated by the authors
  • The score encodes no tissue, cell-type, developmental-stage or environmental context, so interpreting a constrained element requires layering in outside annotations such as tissue-specific enhancers. · stated by the authors
  • Because detection rests on depletion of variation, selection acting after reproductive age is invisible, so regions underlying late-onset phenotypes are likely under-detected. · stated by the authors
  • Power is limited: 76,156 genomes reliably detect only constraint as extreme as the 90th percentile of similarly sized coding exons; an estimated ~340K genomes are needed for the 50th percentile and ~5.3M samples for 100bp rather than 1kb resolution. · stated by the authors
  • The LOEUF-augmentation gain (AUC 0.80 vs 0.73, bootstrap P=0.03) appears to rest on one 50% train/test split scored on 77 underpowered genes, with olfactory receptors as the sole negative class, so its generality is uncertain. · noted in this summary
  • Allele frequency (AF) cutoffs for the 390,393,900 variants behind the map are reported inconsistently (≤1% in Results vs ≤0.1% in Methods), complicating exact reproduction of observed counts. · noted in this summary

Who should read:Read this if you interpret non-coding variants, CNVs or GWAS hits and want a population-based, 1kb-resolution constraint score, its benchmarks against Orion, CDTS and conservation scores, and its power limits.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHavrilla '18Barroso '26Lake '24Hasenahuer '22Di '25Sullivan '23
76% semantic matchGenomics
James M. Havrilla, Brent S. Pedersen, Ryan M. Layer+1 more·Nature Genetics·Nov 28, 2018· 246 citations