Integrating 730,947 exome sequences with clinical literature improves gene discovery
Jérémy GueziD, Julia K. GoodrichiD, Mikhail A. MoldovaniD, Katherine R. ChaoiD, Prathitha KariD, Ruchit PanchaliD, Michael W. WilsoniD, Kristen M. LaricchiaiD, Greg Rohlicek, Dmitry BibaiD, Daniel Marten, Qin HeiD, Philip W. Darnowsky, Riley H. Grant, Ben WeisburdiD, Samantha M. BaxteriD, Joshua Nadeau, Wenhan LuiD, Steve Jahl, Sophie Parsa, Abdallah LamaneiD, Stephanie P. DiTroiaiD, Jack M. FuiD, Xuefang ZhaoiD, Elissa Alarmani, , Konrad J. KarczewskiiD
Abstract
Accurate estimates of allele frequencies aid in genetic discovery, including rare disease diagnosis, common disease investigations, and population genetics. Here, we present the Genome Aggregation Database version 4 (gnomAD v4), including 730,947 with exome sequences, a fivefold increase over previous releases. We demonstrate that statistical power to detect strong selective constraint continues to increase with sample size. We develop a new loss-of-function annotation pipeline, which learns genomic features predictive of nonsense-mediated decay and splicing effects from selection signals, achieving 90% precision for distinguishing likely true versus false positive loss-of-function variants. This improved pipeline, along with incorporation of highly deleterious missense variants into measures of loss-of-function intolerance, improves disease gene detection, particularly for short genes and those with gain-of-function mechanisms. To improve disease gene prediction, we systematically extract gene-disease associations from biomedical literature, map these to gene-level biological features, and integrate both with refined constraint metrics within a Bayesian framework, yielding state-of-the-art prediction of gene-disease relevance. We highlight genes under strong constraint but with limited clinical characterization, which are enriched in embryonic lethal and fertility phenotypes, thus prioritizing previously under-characterized disease genes. Together, these advances establish a unified framework for accelerating gene discovery and improving rare disease diagnosis.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 26
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (25 Mar 2026) · CC BY 4.0
gnomAD v4 (730,947 exomes) brings LOFTEE-2 loss-of-function annotation (0.90 precision, strict mode), missense-aware constraint, and LLM-mined literature evidence combined with constraint to rank disease genes.
Nearly half of rare disease patients remain undiagnosed after genetic testing, and in 2021, 65% of genes in the most loss-of-function-constrained decile had no OMIM disease entry, while the earlier LOFTEE annotator still passed false positive calls. With a fivefold larger exome set, the work recalibrates loss-of-function annotation from selection signals, adds deleterious missense variants to constraint, and fuses LLM-based literature curation with constraint in a Bayesian gene score.
Key results
Shown730,947 unrelated exomes yield 19,940,035 high-quality coding variants. Highly mutable synonymous sites are near-saturated, yet power to separate strongly constrained genes keeps rising with sample size.
Share of possible synonymous variants seen: 96.3% methyl-CpG transitions, 62.0% unmethylated CpG, 32.6% non-CpG transitions, 12.8% transversions. AUPRC (area under precision-recall curve) for severe haploinsufficient (HI) genes rises with sample size; weak vs neutral gene separation dips slightly. Analytical models project power gains through 10 million samples for genes under strong purifying selection. · Overview of the gnomAD v4 dataset; Mutational saturation and recurrence reshape…
ShownLOFTEE-2 calibrates nonsense-mediated decay (NMD) position rules and Pangolin splice thresholds against pneutral, a per-variant probability of neutrality fitted from allele frequencies in strongly constrained genes. This improves the precision of predicted loss-of-function (pLoF) calls.
Against 329 manually curated variants: precision 0.66 (LOFTEE v1) vs 0.77 (LOFTEE-2 relaxed) and 0.90 (strict); recall 1 vs 0.93 and 0.72. GTEx allelic-expression NMD ratio (pLoF vs non-pLoF): 4.24 vs 5.78 and 6.05. LOEUF (loss-of-function observed/expected upper bound fraction) AUPRC for severe HI genes: 0.19 vs 0.224 and 0.253; filtering by pneutral gives 0.277. · Improving loss-of-function variant annotation…
ShownThe 1% most deleterious missense variants are more depleted than pLoF variants. Genes showing this are enriched for gain-of-function (GoF) or dominant-negative (DN) mechanisms, and adding these variants to LOEUF (LOEUF-MIS) improves neurodevelopmental disorder (NDD) gene detection in short genes.
Observed/expected ratio at the top 1%: AlphaMissense 0.28, ESM1v 0.39, PopEVE 0.16 vs 0.55 for pLoF; the excess disappears within HI genes. Enrichment vs LOEUF-matched controls: curated GoF/DN-prone genes 1.7 (p = 8.5 × 10^-3), ion channels 2.47, oncogenes 1.82, tumor suppressors 1.48 (p = 0.14). NDD AUPRC among genes with < 50 expected pLoF: 0.126 (LOEUF) vs 0.176 (LOEUF-MIS). · Combining deleterious missense and…; Figure legends
ShownAn agentic LLM pipeline over PubMed abstracts scores each gene's literature-supported clinical impact (PEPPERLLM). An XGBoost proxy trained on gene features (PEPPERXGB), combined with LOEUF-MIS in a Bayesian model (OMELETXGB), predicts NDD genes better than either input.
47,839 gene-disease associations extracted, recovering 95.7% of definitive GenCC associations and adding 1,961 highly penetrant ones absent from GenCC. On 585 NDD vs 16,582 other genes: OMELETXGB AUPRC 0.504 vs 0.344 (PEPPERXGB) and 0.291 (LOEUF-MIS), bootstrap p < 5 × 10^-4. The literature-based OMELETLLM reaches 0.686, but the NDD list is itself literature-curated. · Integrating literature and constraint improves…; Figure legends
SuggestedContrasting constraint with literature-derived impact (Discovery Potential, DisPo) flags constrained, clinically uncharacterized genes enriched for mouse embryonic-lethal and infertility phenotypes and testis and fetal expression; a feature- vs literature-based contrast (ΔPEPPER) yields 220 constrained candidates.
Median DisPo percentile 61.9 (mouse embryonic-lethal) and 56.8 (mouse infertility) vs 41.7 (GenCC disease genes); p = 7.07 × 10^-58, 4.19 × 10^-16. High-DisPo odds ratios: 1.83 (testis-specific GTEx expression), 1.48 (broad expression in 15 fetal tissues). 220 candidates (PEPPERXGB above the 95th percentile, no GenCC association, ΔPEPPER > 25): median LOEUF 0.31 vs 0.92 (p = 2.4 × 10^-87). · Integrating literature and constraint improves…; Discordance between constraint and literature…; Data Availability
Methods, data and reproducibility
Samples were jointly called (GRCh38, Hail VariantDataset) and quality-controlled to 730,947 unrelated exomes. Saturation of constraint was modeled analytically and by downsampling; a Bayesian mixture model on pLoF frequencies estimated pneutral and heterozygous selection (shet) to calibrate LOFTEE-2; LLM-curated PubMed scores (or an XGBoost proxy) served as Bayesian priors with gnomAD counts as likelihood.
- gnomAD v4 exomes after QC, drawn from 1,108,389 aggregated exomes and genomes (gnomAD v2/v3, UK Biobank, other cohorts) · human · n = 730,947
- Manually curated pLoF variants used to benchmark LOFTEE versions · n = 329
- GTEx allelic expression (stop-gained variants) and tissue expression · human (GTEx) · n = 54 tissues
- PubMed abstracts processed by the agentic LLM curation pipeline · n = 47,839 gene-disease associations extracted
- GenCC gene-disease associations (validation and disease gene labels)
- NDD gene benchmark for disease gene prediction · n = 585 NDD-positive and 16,582 negative genes
- Mouse embryonic-lethal and mouse infertility gene sets · mouse
- Fetal tissue expression of high-DisPo genes vs LOEUF-matched controls · n = 15 fetal tissues; 901 gene pairs
- Methods & tools
- Hail, LOFTEE, LOFTEE-2, Pangolin, SpliceAI, ESM1v, AlphaMissense, PopEVE, XGBoost, GenCC, OMIM, GeneBayes
- Code
- https://github.com/broadinstitute/gnomad_qc/releases/tag/v4.1https://github.com/atgu/gnomAD_v4_flagship_paper
- Data & accessions
- https://gnomad.broadinstitute.org/data#v4
- Software & environment
- QC code is pinned to the gnomad_qc v4.1 release tag; no software versions, containers or requirement files are stated.
- Reproducibility gaps
- The LLM(s) used for agentic curation are not named in the main text; released data are aggregate (sites-level VCF, Hail Tables, TSV); most method details are deferred to Supplementary Material not included here.
Limitations (6)
- pneutral can be inferred directly only in genes under strong purifying selection, and saturation at gnomAD v4 scale lowers LOEUF's power to separate weakly selected from neutral genes. · stated by the authors
- Because the NDD gene lists were curated from literature, scoring LLM-derived metrics against them mainly measures curation quality, not predictive power. · stated by the authors
- Stronger missense than pLoF constraint need not mean GoF/DN: clonal expansion (e.g., LZTR1, pLoF observed/expected 2.08) or transcriptional adaptation via paralogs can produce it, and separating these is left to future studies. · stated by the authors
- LLM curation still needs testing on edge cases, manually curated benchmarks, and comparisons across LLMs, including biomedical fine-tuned versus general-purpose models. · stated by the authors
- In the text, OMELETXGB's AUPRC is reported only against its own inputs despite claims of outperforming existing methods; the one external method (GeneBayes, AUPRC 0.148) is in Extended Data Fig. 7b, whose LOEUF-MIS AUPRC (0.22) differs from the main text's 0.176, likely because Fig. 4d uses short genes only. · noted in this summary
- Prioritized genes (220 ΔPEPPER candidates, 182 top-1% DisPo genes) rest on enrichment statistics, a GenCC submission-year trend and one case later confirmed by an independent study (DENND2B); no systematic prospective or functional validation is reported. · noted in this summary
Who should read:Read this if you use gnomAD allele frequencies or LOEUF for variant interpretation or gene prioritization and want to know how LOFTEE-2, LOEUF-MIS and the literature-informed OMELET and DisPo scores change disease gene ranking.
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
- Improving the diagnostic yield of exome-sequencing, by predicting gene-phenotype associations using large-scale gene expression analysis
Patrick Deelen, Sipko van Dam, … Lude Franke·Nature CommunicationsPeer-reviewed·Jun 28, 2019
- Inference of elevated mutation rates and variant effects using 700k exomes
Prathitha Kar, Mikhail A. Moldovan, … Shamil R. Sunyaev·bioRxiv·Jun 10, 2026
- Systematic single-variant and gene-based association testing of thousands of phenotypes in 426,370 UK Biobank exomes
Konrad J. Karczewski, Matthew Solomonson, … Benjamin M. Neale·Cell GenomicsPeer-reviewed·Aug 15, 2022
- Divine: Prioritizing Genes for Rare Mendelian Disease in Whole Exome Sequencing Data
Changjin Hong, Jean Rene V. Clemenceau, … TaeHyun Hwang·bioRxiv·Aug 21, 2018
- Transcript expression-aware annotation improves rare variant discovery and interpretation
Beryl B. Cummings, Konrad J. Karczewski, … David B. Goldstein·NaturePeer-reviewed·May 27, 2020
- Exome-by-phenome-wide rare variant gene burden association with electronic health record phenotypes
Joseph Park, Anastasia Lucas, … Daniel James Rader·Nature MedicinePeer-reviewed·Jan 1, 2021