What gnomAD pLI Measures and When to Trust It
pLI is the probability that a gene belongs to the class of genes intolerant to a single inactivating copy. It comes from a three-class mixture model fitted across all genes in gnomAD. For each gene the model compares the number of observed predicted loss-of-function (pLoF) variants against the number expected under a sequence-context mutational model. Genes with far fewer pLoF variants than expected get pLI near 1. The conventional cutoff is pLI ≥ 0.9, and it is a statement about haploinsufficiency in the population. It does not describe any variant you happen to carry.
The score answers one narrow question: has natural selection removed heterozygous truncating variants in this gene from the sampled population? It does not tell you whether a given stop-gain is pathogenic, whether a gene causes recessive disease, or how severe a phenotype would be. Questions of that kind require separate evidence and, for anything clinical, a clinician and a diagnostic laboratory.
How the score is computed
The expected pLoF count comes from a mutational model keyed on trinucleotide context, corrected for sequencing depth and base-level coverage. Every possible single nucleotide change in the coding sequence is assigned a mutation rate. Those rates are summed over changes that would create a stop-gain, a canonical splice site disruption, or a frameshift-equivalent event as annotated by VEP with the LOFTEE plugin. Only high-confidence LOFTEE calls are counted. That excludes variants in the last 50 base pairs of the penultimate exon, variants in non-canonical splice contexts, and other annotation artifacts.
The observed count is the number of distinct high-confidence pLoF variants seen in the sample. It is not the allele count. A variant carried by 400 people counts once, the same as a singleton. That is deliberate. Constraint is about the absence of new variation, so the model cares about how many distinct sites survive rather than how common any one of them is.
The mixture model then assigns each gene a posterior probability of falling into one of three classes: tolerant of homozygous loss, tolerant of heterozygous loss only, and intolerant of heterozygous loss. pLI is the posterior weight on that third class. The companion values pRec and pNull are the other two weights, and the three sum to one.
Getting the files and joining them correctly
The gnomAD v4.1 constraint table is a single flat file in a public bucket:
gsutil cp gs://gcp-public-data--gnomad/release/4.1/constraint/gnomad.v4.1.constraint_metrics.tsv .
Read the header before writing any parser, because the column names changed between major versions. In v4.1 you want gene, transcript, and mane_select. You also want lof.pLI, lof.oe, and lof.oe_ci.upper (LOEUF). Finally take lof.obs and lof.exp along with mis.z_score and constraint_flags. In v2.1.1 the file is gnomad.v2.1.1.lof_metrics.by_gene.txt.bgz. The equivalent columns are pLI, oe_lof, and oe_lof_upper. The rest are obs_lof, exp_lof, and mis_z.
Constraint is computed per transcript, so the first thing to do is filter to one transcript per gene. For v4, take rows where mane_select is true. Joining on gene symbol alone will silently give you a non-MANE transcript for some genes and produce different pLI values than you expect. Symbols also drift between releases, so join on Ensembl gene ID where you can.
To annotate your own variants, run VEP with LOFTEE and keep only high-confidence calls on the MANE Select transcript:
vep -i sample.vcf.gz --cache --assembly GRCh38 --vcf \
--mane_select --canonical --symbol \
--plugin LoF,loftee_path:$LOFTEE,human_ancestor_fa:$HA_FA,gerp_bigwig:$GERP \
-o sample.loftee.vcf
Then join the SYMBOL and Feature fields against the constraint table. Discard anything where the LOFTEE LoF field is LC, and read the LoF_filter reason before you decide a variant is real.
The threshold and what falls outside it
Roughly three thousand genes clear pLI ≥ 0.9 in a given release, out of about twenty thousand protein-coding genes. The number moves between versions. That set is heavily enriched for genes underlying dominant developmental disorders, which is why gene discovery studies routinely use constraint as a prioritization filter when they have too many candidates and too few cases. Exome studies of clubfoot, congenital cardiac laterality defects, and autism have all used constraint metrics this way to narrow rare variant candidate lists 123.
Three things are worth internalizing about the cutoff. First, pLI saturates. A gene with pLI = 0.999 and a gene with pLI = 1.000 are not meaningfully rankable. Second, pLI near zero is not evidence of tolerance when the gene is small. Third, recessive disease genes routinely sit at pLI ≈ 0, because heterozygous carriers are not selected against. A pLoF variant in a recessive gene can matter enormously for reproductive planning while the gene’s pLI says nothing about it.
Why pLI changed between v2 and v4
gnomAD v2.1.1 constraint used about 125,000 exomes. v4 uses roughly 730,000. More samples means more observed pLoF variants per gene, and the observed count is the numerator of the constraint ratio. Genes that had zero observed pLoF in v2 often have several in v4, which pulls pLI down sharply even when the underlying biology is unchanged.
The effect is not uniform. Short genes gain the most proportionally, and the mixture model’s class boundaries shift when the whole distribution moves. Do not compare pLI values across versions. Pin a version in your pipeline, record it, and recompute rather than reuse when you upgrade.
Population composition matters too. gnomAD aggregates many ancestry groups, and rare variant frequencies differ substantially between them. Analyses of admixed cohorts have shown that the rare variant contribution to disease depends on local ancestry in ways that a single aggregated frequency obscures 4. Constraint is an aggregate summary and inherits that limitation.
pLI or LOEUF
We use LOEUF and keep pLI as a secondary sanity check. LOEUF is the upper bound of the 90% confidence interval on the observed-to-expected pLoF ratio. It is continuous, it degrades gracefully in short genes by returning a wide interval rather than a confident zero, and it can be used to rank genes rather than sort them into two bins. The commonly cited cutoff is LOEUF < 0.6, though with v4 sample sizes the deciles in the file are a better ranking tool than any single threshold.
pLI is still useful for one thing: it is what most published filters and most clinical literature were built on, so reproducing someone else’s gene list usually requires it. Report both.
Failure modes worth knowing
Low expected counts break the score. A gene with lof.exp under about 10 cannot reach high pLI, because the model never has enough statistical power to distinguish depletion from chance. Check lof.exp before you believe a low pLI.
The constraint_flags column flags genes with no expected pLoF, outlier observed-to-expected ratios, or poor coverage. Filter on it. Genes in segmental duplications and highly paralogous families produce mapping artifacts that inflate observed counts and deflate constraint.
Constraint says nothing about missense variants. Use mis.z_score for missense depletion, and be aware that gene-level missense constraint averages over domains that differ greatly in tolerance. Regional constraint methods are better where they exist.
Finally, constraint is a human population statistic. It does not transfer to other species, and cross-species LoF catalogues built with similar annotation pipelines need their own reference panels and their own mutational models 5. The same caution applies to applying constraint to gene sets defined by function rather than by inheritance, such as host proteins that interact with a virus 6.
Questions people also ask
What does a pLI of 1 mean for a variant I carry?
It means the gene is depleted of heterozygous truncating variants across the population. It does not classify your variant. A high-confidence pLoF in a high-pLI gene is a reason for a clinician and a diagnostic lab to look, and nothing more on its own.
Why did my gene’s pLI drop from 0.99 to 0.3 in v4?
Sample size. With 730,000 exomes the gene now has observed pLoF variants where it had none at 125,000. Compare LOEUF instead, and compare within a single release.
Should I use pLI ≥ 0.9 or LOEUF < 0.6?
Use LOEUF as your primary metric and deciles rather than a hard cutoff. Keep pLI when you need to reproduce a published filter.
Does pLI apply to recessive conditions?
No. pLI models selection against heterozygous loss. Most recessive disease genes have pLI near zero, and carrier status in those genes is a separate question requiring separate analysis.
Footnotes
-
Wu-Lin Charng, Momchil Nikolov, Isabel Shrestha, et al. Exome sequencing of 1190 non-syndromic clubfoot cases reveals HOXD12 as a novel disease gene. Journal of Medical Genetics, 2024. https://doi.org/10.1136/jmg-2024-109846 ↩
-
Archana Rai, Jonathan Klonowski, Bo Yuan, et al. Genomic rare variant mechanisms for congenital cardiac laterality defect: A digenic model approach. The American Journal of Human Genetics, 2025. https://doi.org/10.1016/j.ajhg.2025.05.014 ↩
-
Vijay Gupta, Afif Ben-Mahmoud, Bonsu Ku, et al. Identification of two novel autism genes, TRPC4 and SCFD2, in Qatar simplex families through exome sequencing. Frontiers in Psychiatry, 2023. https://doi.org/10.3389/fpsyt.2023.1251884 ↩
-
Courtney Astore, Shivam Sharma, Sini Nagpal, et al. The role of admixture in the rare variant contribution to inflammatory bowel disease. Genome Medicine, 2023. https://doi.org/10.1186/s13073-023-01244-w ↩
-
R. Alan Harris, Muthuswamy Raveendran, Wes Warren, et al. Whole Genome Analysis of SNV and Indel Polymorphism in Common Marmosets (Callithrix jacchus). Genes, 2023. https://doi.org/10.3390/genes14122185 ↩
-
Suvobrata Chakravarty. COVID-19: The Effect of Host Genetic Variations on Host–Virus Interactions. Journal of Proteome Research, 2020. https://doi.org/10.1021/acs.jproteome.0c00637 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.