Evaluating drug targets through human loss-of-function genetic variation

Naturally occurring human genetic variants that are predicted to inactivate protein-coding genes provide an in vivo model of human gene inactivation that complements knockout studies in cells and model organisms. Here we report three key findings regarding the assessment of candidate drug targets using human loss-of-function variants. First, even essential genes, in which loss-of-function variants are not tolerated, can be highly successful as targets of inhibitory drugs. Second, in most genes, loss-of-function variants are sufficiently rare that genotype-based ascertainment of homozygous or compound heterozygous ‘knockout’ humans will await sample sizes that are approximately 1,000 times those presently available, unless recruitment focuses on consanguineous individuals. Third, automated variant annotation and filtering are powerful, but manual curation remains crucial for removing artefacts, and is a prerequisite for recall-by-genotype efforts. Our results provide a roadmap for human knockout studies and should guide the interpretation of loss-of-function variants in drug development.

Building on these insights, here we leverage pLoF variation in the Genome Aggregation Database (gnomAD) 7 v2 dataset of 141,456 individuals to answer open questions in the interpretation of human pLoF variation in disease biology and drug development.

Constraint in human drug targets
We compared constraint in the targets of approved drugs extracted from DrugBank 17 (n = 383) versus all protein-coding genes (n = 17,604). Drug targets were, on average, just slightly more constrained than all genes (mean 44% versus 52%, nominal P = 0.00028, D = 0.11, two-sided Kolmogorov-Smirnov test), but the two gene sets had a qualitatively similar distribution of scores, ranging from intensely constrained (0% obs/exp) to not at all constrained (≥100% obs/exp) (Fig. 1a). Constraint scores showed clear divergence between categories of genes (Extended Data Table 1) expected to be more or less tolerant of inactivation (Fig. 1b), as previously reported 7,10 , validating the usefulness of constraint as a measure of gene essentiality. Nonetheless, when drug targets were stratified by drug effect (Fig. 1b), modality, or indication (Extended Data Fig. 1), no statistically significant differences between subsets of drug targets were observed.
The slightly but significantly lower obs/exp value among drug targets may superficially appear to provide evidence that constrained genes make superior drug targets. Stratification of drug targets by protein family, human disease association, and tissue expression, however, argues against this interpretation. Drug targets are strongly enriched for a few canonically 'druggable' protein families, for genes known to be involved in human disease, and for genes with tissue-restricted expression; each of these properties is in turn correlated with either significantly stronger or weaker constraint (Extended Data Fig. 2). Although controlling for these correlations does not abolish the trend of stronger constraint among drug targets, the correlation of so many observed variables with the status of a gene as a drug target argues that many unobserved variables probably also confound interpretation of the lower mean obs/exp value among drug targets.
The overall constraint distribution of drug targets (Fig. 1a) also argues against the view that a gene in which LoF is associated with a deleterious phenotype cannot be successfully targeted. Indeed, 19% of drug targets (n = 73), including 52 targets of inhibitors, antagonists or other 'negative' drugs, have lower obs/exp values than the average (12.8%) for genes known to cause severe diseases of haploinsufficiency 18 (ClinGen level 3). To determine whether this finding could be explained by a particular class or subset of drugs, we examined constraint in several well-known example drug targets (Fig. 1c, Extended Data Table 2). Some heavily constrained genes are targets of cytotoxic chemotherapy agents such as topoisomerase inhibitors or cytoskeleton disruptors, a set of drugs intuitively expected to target essential genes. However, genes with near-complete selection against pLoF variants also include HMGCR and PTGS2, the targets of highly successful, chronically used inhibitors-statins and aspirin.
These human in vivo data further the evidence from other species and models that essential genes can be good drug targets. Homozygous knockout of Hmgcr and Ptgs2 are lethal in mice [19][20][21] . Drug targets exhibit higher inter-species conservation than other genes 22 . Targets of negative drugs include 14 genes with lethal heterozygous knockout mouse phenotypes reported 23 and 6 reported as essential in human cell culture 24 .

Prospects for finding human 'knockouts'
Athough constraint alone is not adequate to nominate or exclude drug targets, the study of individuals with single hit (heterozygous) or two-hit ('knockout') LoF genotypes in a gene of interest can be highly informative about the biological effect of engaging that target 5 . To assess prospects for ascertaining knockout individuals, we computed the cumulative allele frequency (CAF) of pLoF variants in each gene (Methods), and then used this to estimate the expected frequency of two-hit individuals under different population structures (Fig. 2) in the absence of natural selection.
Whereas gnomAD is now large enough to include at least one pLoF heterozygote for most (15,317 out of 19,194; 79.8%) genes, ascertainment of total knockout individuals in outbred populations will require 1,000-fold larger sample sizes for most genes: the median expected two-hit frequency of a gene is just six per billion (Fig. 2a). Even if every human on Earth were sequenced, there are 4,728 genes (24.6%) for which identification of even one two-hit individual would not be expected in outbred populations. Intuitively, because the sample size of gnomAD today is larger than the square root of the world population, variants so far seen in zero or only a few heterozygous individuals are not likely to ever be seen in a homozygous state in outbred populations, except where variants prove common in populations not yet well-sampled by gnomAD.
Because population bottlenecks can result in very rare variants present in a founder rising to an unusually high frequency, we also considered knockout discovery in bottlenecked populations, using Finnish individuals in gnomAD as an example 8 . Although this population structure can enable well-powered association studies for the small fraction of genes in which pLoF variants drifted to high frequency due to the bottleneck, overall, identification of two-hit pLoF individuals    Table 1). For drug effect, 'positive' indicates agonist, activator or inducer, whereas negative indicates antagonist, inhibitor or suppressor, for example. c, Examples of drug targets and corresponding drug classes from across the constraint spectrum. Details in Extended Data Table 2. for a pre-specified gene of interest appears equally or more difficult in Finnish individuals than in outbred populations (Fig. 2b, Extended Data Fig. 3), because rare variants not present in a founder have been effectively removed from the population.
In consanguineous individuals, parental relatedness greatly increases the frequency of homozygous pLoF genotypes. The n = 2,912 individuals in the East London Genes & Health (ELGH) cohort 25 who report having parents who are second cousins or closer have on average 5.8% of their genomes autozygous. Here, the expected frequency of two-hit individuals is many times higher than in outbred populations, at five per million for the median gene (Fig. 2c).
These projections allow us to draft a roadmap for discovery of human knockouts across 19,194 genes (Fig. 2d, e). Online Mendelian Inheritance in Man (OMIM) already describes human disease association for 3,367 genes (18%), although the discovery of LoF individuals in population databases will still be valuable for assessing penetrance and identifying LoF syndromes of known gain-of-function genes. Another 3,421 genes (18%) without known human disease association have two-hit pLoF genotypes reported in gnomAD 7 , ELGH 26 , PROMIS 27 , deCODE 28 or UK Biobank 29 , which suggests that this genotype may be tolerated. An additional 2,190 genes (11%) appear intolerant of heterozygous inactivation (pLI score > 0.9) in gnomAD-a set expected to be enriched for genes with severe heterozygous and lethal homozygous LoF phenotypes. Another 2,781 genes (14%) have no pLoF variants observed in gnomAD, but our sample size is not yet large enough to robustly infer LoF intolerance. For these genes, observation of outbred two-hit individuals is not expected, and we cannot yet assess the feasibility of identifying consanguineous two-hit individuals because we lack an estimate of pLoF allele frequency. This leaves 7,435 genes (39%) for which one or more pLoFs are observed in gnomAD, but strong LoF intolerance cannot be determined, two-hit genotypes have not been observed, and a human disease phenotype is not known. We projected the sample sizes required to identify knockout individuals for these genes (Fig. 2e). In outbred populations, current sample sizes would need to increase by approximately 1,000-fold before ascertainment of a single two-hit LoF individual would be expected for the typical gene. By contrast, around a 10-to 100-fold increase from current consanguineous sample size, meaning hundreds of thousands of individuals in absolute terms, would identify at least one two-hit LoF individual for the typical gene. Among other simplifying assumptions (Methods), these projections presume that complete knockout is tolerated. When only one or a few two-hit individuals are expected in a dataset, the absence of any such individuals can be due to either early lethality, a severe clinical phenotype incompatible with inclusion in gnomAD, or simply chance. Thus, the ability to infer lethality of the two-hit genotype based on statistical evidence will lag behind the identification of two-hit individuals where they do exist (Fig. 2e). For some genes, inference of lethality will always remain impossible in outbred populations, though it may be feasible in consanguineous individuals.

Curation of pLoF variants
Where pLoF variants can be identified, they are a valuable resource for assessing the effect of lifelong reduction in gene dosage. To highlight the challenges and opportunities of identifying such variants, we manually curated gnomAD data and the scientific literature for six genes associated with gain-of-function (GoF) neurodegenerative diseases, for which inhibitors or suppressors are under development [30][31][32][33][34][35] : HTT (Huntington's disease), MAPT (tauopathies), PRNP (prion disease), SOD1 (amyotrophic lateral sclerosis), and LRRK2 and SNCA (Parkinson's disease). The results (Fig. 3, Extended Data Table 3) illustrate four points about pLoF variant curation.
First, other things being equal, genes with longer coding sequences offer more opportunities for LoF variants to arise, and so tend to have a higher cumulative frequencies of LoF variants, unless they are heavily constrained. Ascertainment of LoF individuals is thus harder for shorter and/or more constrained genes, even though these may be good targets.
Second, many variants annotated as pLoF are false positives 6 , and these are enriched for higher allele frequencies, so that both filtering and curation have an outsized effect on the cumulative allele frequency of LoF. Studies of human pLoF variants lacking stringent curation can therefore easily dilute results with false pLoF carriers.
Third, after careful curation, cumulative LoF allele frequency is sometimes sufficiently high to place certain bounds on what heterozygote phenotype might exist. For example, GoF mutations causing genetic prion disease have a genetic prevalence of approximately 1 in 50,000 36 and have been known for three decades, with thousands of cases identified, making it unlikely that a comparably severe and penetrant haploinsufficiency syndrome associated with PRNP would have gone unnoticed to the present day despite being more than twice as common (roughly 1 in 18,000). Similar arguments can be made for HTT, LRRK2 and SOD1 genes (Extended Data Tables 3, 4). Of course, this does not rule out a less severe or less penetrant heterozygous LoF phenotype.
Finally, careful inspection of the distributions of pLoF variants can reveal important error modes or disease biology. HTT, MAPT   and PRNP genes each have different non-random positional distributions of pLoF variants (Fig. 3). High-frequency HTT pLoF variants cluster in the polyglutamine/polyproline repeat region of exon 1 and appear to be alignment artefacts (Fig. 3a). True HTT LoF variants are rare and the gene is highly constrained, which might suggest some fitness effect in a heterozygous state in addition to the known severe homozygous phenotype 37,38 , although the frequency of LoF carriers still argues against a penetrant syndromic illness, consistent with the lack of phenotype reported in heterozygotes identified so far 38,39 . High-frequency MAPT pLoF variants cluster in exons not expressed in the brain in GTEx data 14,40 , and all remaining pLoFs appear to be alignment or annotation errors (Fig. 3b). No true LoFs are observed in MAPT, although our sample size is insufficient to prove that MAPT LoF is not tolerated-among constitutive brain-expressed exons, we expect 12.6 LoFs and observe 0, giving a 95% confidence interval upper bound of 23.7% for obs/exp values. PRNP-truncating variants in gnomAD cluster in the N terminus; the sole C-terminal truncating variant in gnomAD is a dementia case (Extended Data Table 5), consistent with variants at codon ≥145 causing a pathogenic gain-of-function through change in localization (Fig. 3c). Within codons 1-144, PRNP is unconstrained (Extended Data Table 3), and no neurological phenotype has been identified in individuals with truncating variants so far, consistent with the hypothesis that N-terminal truncating variants are true LoF and are tolerated in a heterozygous state 41 .

Discussion
Studying human gene inactivation can illuminate human biology and guide the selection of drug targets, complementing mouse knockout studies 42 , but analysis of any one gene requires genome-wide context to set expectations and guide inferences. Here we have used gnomAD data to provide context to aid in the interpretation of human LoF variants. Targets of approved drugs range from highly constrained to completely unconstrained. There may be several reasons why some genes apparently tolerate pharmacological inhibition but not genetic inactivation. LoF variants in constitutive exons should affect all tissues for life, whereas drugs differ in tissue distribution and timing and duration of use. Many drugs known or suspected to cause fetal harm are tolerated in adults 43 , and might target developmentally important genes. Constraint is thought to primarily reflect selection against heterozygotes 13 , the effective gene dosage of which may differ from that achieved by a drug. Constraint measures natural selection over centuries or millennia; the environment of our ancestors presented different selective pressures from what we face today. Actions of small-molecule drugs may not map one-to-one onto genes [44][45][46][47] . Regardless, these human in vivo data show that even a highly deleterious knockout phenotype is compatible with a gene being a viable drug target.
For most genes, the lack of total knockout individuals identified so far does not yet provide statistical evidence that this genotype is not tolerated. Indeed, for many genes, such evidence may never be attainable in outbred populations. Bottlenecked populations, individually, are unlikely to yield two-hit individuals for a pre-specified gene of interest, although the sequencing of many different, diverse bottlenecked populations will certainly expand the set of genes accessible by this approach. Identification of two-hit individuals will be most greatly aided by increased investment in consanguineous cohorts, in which the sample size required for any given gene is often orders of magnitude lower than in outbred populations. Our analysis is limited by sample size, insufficient diversity of sampled populations, and simplifying assumptions about population structure and distribution of LoF variants, so our calculations should be taken as rough, order-of-magnitude estimates. Nonetheless, this strategic roadmap for the identification of human knockouts should inform future research investments and rationalize the interpretation of existing data.
Recall-by-genotype efforts are only valuable if the variants in question are correctly annotated. Automated filtering 7 and transcript expression-aware annotation 14 are powerful tools, but we demonstrate the continued value of manual curation for excluding further false positives, assessing and interpreting the cumulative allele frequency of true LoF variants, and identifying error modes or biological phenomena that give rise to non-random distributions of pLoF variants across a gene. Such curation is essential before any recontact efforts, and establishing methods for high-throughput functional validation 48 of LoF variants is a priority. Our curation of pLoF variants in neurodegenerative disease genes is limited by a lack of functional validation and detailed phenotyping; a companion paper demonstrates a deeper investigation of the effects of LoF variants in the LRRK2 gene 49 .
Drug development projects may increasingly be accompanied by efforts to phenotype human carriers of LoF variants. With the cost of drug discovery driven overwhelmingly by failure 50 , successful interpretation of LoF data to select the right targets and right clinical pathways will yield outsize benefits for research productivity and, ultimately, human health.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2267-z.

Calculation of pLoF constraint
The calculation of constraint values for genes has been described in general elsewhere 10,12 and for this dataset specifically by Karczewski et al. 7 . Constraint calculations used LOFTEE-filtered ('high confidence') single-nucleotide variants (which for pLoF means nonsense and essential splice site mutations) found in gnomAD exomes with minor allele frequency <0.1%. Only unique canonical transcripts for protein-coding genes were considered, yielding 17,604 genes with available constraint values. For curated genes (Extended Data Table 2), the number of observed variants passing curation was divided by the expected number of variants to yield a curated constraint value. For PRNP, the expected number of variants was adjusted by multiplying by the ratio of the sum of mutation frequencies for all possible pLoF variants in codons 1-144 to the sum of mutation frequencies for all possible pLoF variants in the entire transcript, yielding 6 observed out of 6.06 expected. For MAPT, the expected number of variants was taken from Ensembl transcript ENST00000334239, which includes only the exons identified as constitutively brain-expressed in Fig. 3b (exon numbering previously described 51 ).

Calculation of pLoF heterozygote and homozygote/compound heterozygote frequencies
LOFTEE-filtered high-confidence pLoF variants with minor allele frequency <5% in 125,748 gnomAD exomes were used to compute the proportion of individuals without a loss-of-function variant (q); the CAF was computed as p = 1 − sqrt(q). This approach conservatively assumes that, if an individual has two different pLoF variants, they are in cis to each other and count as only one pLoF allele. For outbred populations (Fig. 2a), we used the value of p from all 125,748 gnomAD exomes, as this allows the largest possible sample size. This includes some individuals from bottlenecked populations, for which the distribution of p does differ from outbred populations, but these individuals are a small proportion of gnomAD exomes (12.6%). This also includes some consanguineous individuals, but these are an even smaller proportion of gnomAD exomes (2.3%), and any difference in the value of p between consanguineous and outbred populations is expected to be very small. Heterozygote frequency was calculated as 2p(1 −p) and homozygote and compound heterozygote frequency was calculated as p 2 . Lines indicate the size of gnomAD (141,456 individuals) and the world population (6.69 billion).
For bottlenecked populations (Fig. 2b), we used the value of p from the 10,824 Finnish exomes only. Lines indicate the number of Finnish individuals in gnomAD (12,526) and the population of Finland (5.5 million).
For consanguineous individuals (Fig. 2c), we again used the value of p from all gnomAD exomes, because p is not expected to differ greatly in consanguineous versus outbred populations. We used the mean proportion of the genome in runs of autozygosity (a) from individuals self-reporting second cousin or closer parents in East London Genes & Health, a = 0.05766 (rounded to 5.8%). Heterozygote frequency was calculated as 2p(1 − p) and homozygote and compound heterozygote frequency was calculated as (1 − a)p 2 + ap. Lines indicate the number of consanguineous South Asian individuals in gnomAD (n = 2,912, by coincidence the same number as report second cousin or closer parents in ELGH) based on F > 0.05 (a conservative estimate, because second cousin parents are expected to yield F = 0.015625), and the estimated number of individuals in the world with second cousin or closer parents (10.4% of the world population) 9 .
Several caveats apply to our CAF analysis. First, our approach naively treats genes with no pLoFs observed as having P = 0, even though pLoFs might be discovered at a larger sample size. Second, we naively group all populations together, even though the distribution of populations sampled in gnomAD does not reflect the world population 7 ; we believe that this is reasonable because CAF for many genes is driven by singletons and other ultra-rare variants for which frequency is not expected to differ appreciably by continental population 10 . (It is important to note that the histograms shown in Fig. 2 reflect the expected frequency of heterozygotes and homozygotes/compound heterozygotes, based on gnomAD allele frequency, rather than the actual observed frequency of individuals with these genotypes in gnomAD.) Third, we use only protein-truncating variants annotated as pLoF in gnomAD. Structural and non-coding variation resulting in a loss of function may be missed in exomes, and missense variants resulting in a loss of function cannot be rigorously annotated. Fourth, we naively treat genes with one pLoF allele observed as having P = 1/(2 × 125,748), even though on average singleton variants have a true allele frequency lower than their nominal allele frequency 10 . Fifth, the variants included in this analysis are filtered but have not been manually curated or functionally validated, so some will ultimately prove not to be true LoF. These false positives tend to be more common and will have disproportionately contributed to the cumulative LoF allele frequency. Sixth, as described in the main text, our calculations assume that complete knockout is tolerated, which will not be true for some genes. We therefore also include a projection of the sample size needed to infer lethality from the absence of two-hit knockout individuals (Fig. 2e). Points one to three will tend to lead to underestimation of the true complete knockout frequency, whereas points four to six will tend to lead to overestimation. On balance, our calculations may reflect an upper bound of complete knockout frequency for most genes owing to the strong influence of factors five and six. Finally, as a matter of comparison between population structures, the sample size for all gnomAD exomes (Fig. 2a, c) is larger than for only Finnish exomes (Fig. 2b). For a version of Fig. 2 with the global gnomAD population downsampled to the same sample size as the gnomAD Finnish population, see Extended Data Fig. 2.

Knockout roadmap
For the knockout 'roadmap' (Fig. 2d, e), we classified genes according to the current status of human disease association and LoF ascertainment. Genes were classified as having a Mendelian disease association if they were present in OMIM with the filters described in Extended Data Table 1.
Remaining genes were classified as '2-hit LoF reported' based on presence in one or more of the following gene lists: homozygous LoF genotypes in gnomAD curated as previously described 7 ; filtered homozygous LoF genotypes in runs of autozygosity with minor allele frequency <1% in canonical transcripts in the Bradford, Birmingham and ELGH 25 cohorts (total n = 8,925); observed number of imputed homozygotes >1 or number of compound heterozygous carriers where minor allele frequency <2% (for both variants) in deCODE 28 ; homozygous LoF reported in PROMIS 27 ; homozygous LoF with minor allele frequency <1% in UK Biobank 29 .

Genetic prevalence estimation
Here, we define 'genetic prevalence' for a given gene as the proportion of individuals in the general population at birth who have a pathogenic variant in that gene that will cause them to later develop disease. Genetic prevalence has not been well-studied or estimated for most disease genes.
In principle, it should be possible to estimate genetic prevalence simply by examining the allele frequency of reported pathogenic variants in gnomAD. In practice, three considerations usually preclude this approach. First, the present gnomAD sample size of 141,456 exomes and genomes is still too small to permit accurate estimates for very rare diseases. Second, the mean age of gnomAD individuals is approximately 55, which is above the age of onset for many rare genetic diseases, and individuals with known Mendelian disease are deliberately excluded, so pathogenic variants will be depleted in this sample relative to the whole birth population. Third and most importantly, a large fraction of reported pathogenic variants lack strong evidence for pathogenicity and are either benign or low penetrance 10,41 , so without careful curation of pathogenicity assertions, summing the frequency of reported pathogenic variants in gnomAD will in most cases vastly overestimate the true genetic prevalence of a disease.
Instead, we searched the literature and very roughly estimated genetic prevalence based on available data. In most cases, we took disease incidence (new cases per year per population), multiplied by proportion of cases due to variants in a gene of interest, and multiplied by average age at death in cases. In some cases, estimates of at-risk population or direct measures of genetic prevalence were available. Details of the calculations undertaken for each gene are provided in Extended Data Table 4.

Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this paper.

Code availability
Additional data and the R 3.5.1 and Python 2.7.10 source code for this study are available via GitHub (https://github.com/ericminikel/ drug_target_lof).

Extended Data Fig. 1 | Drug target constraint by modality and indication.
Mean (dots) and 95% confidence interval (line segments) for constraint in subsets of drug-targets sets (data sources and number of genes for each list are provided in Extended Data Table 1). Modality information was extracted from DrugBank and indication information from ATC codes; see Extended Data Table 1 for details. Fig. 2 | Drug-target gene set confounding. a, Forest plot of means (dots) and 95% confidence intervals of the mean (line segments) for gene sets evaluated for confounding with drug-target status. Data sources and number of genes for each list are provided in Extended Data Table 1. LoF obs/exp ratios differ significantly from the set of all genes for four canonically druggable protein families (top), human disease-associated genes (middle), and genes by broadness of tissue expression (bottom). Within each class, the genes that are drug targets have a lower mean obs/exp ratio (hollow grey circles) than the class overall. b, The druggable protein families, disease-associated genes, and genes expressed in some tissues but not others are enriched several-fold among the set of drug targets. Bars indicate fold enrichment and error bars indicate 95% confidence intervals. c-e, Composition of drug targets when broken down by protein family (c), disease association (d), or broadness of tissue expression (e). The enriched classes account for most drug targets. In a linear model, after controlling for protein family, disease association status, and number of tissues with expression >1 transcript per million (TPM), drug targets are still more constrained than other genes (−8.0% obs/exp, nominal P = 0.00011, t = −3.9, df = 17,325 for the contribution of drug_target in the linear regression obs/exp ~ drug_target + family + dz_assoc + n_tissues), but the probable existence of additional unobserved confounders cautions against over-interpretation of this observation (see main text). Fig. 3 | Expected frequency of individuals with one or two null alleles for every protein-coding gene across different population models, with sample size held constant. This is identical to Fig. 2 except as follows. As noted in the Methods, one caveat about Fig. 2 is that the sample size is larger for the plots using all gnomAD exomes (Fig. 2a, c) than for Finnish exomes (Fig. 2b). This figure shows the same analysis, but with the global gnomAD population downsampled to 10,824 randomly chosen exomes so that the same size is identical to that of Finnish exomes. Computation of P = 1 − sqrt(q) as described in the Methods is computationally expensive for downsampled datasets because it requires individual-level genotypes. Instead, this analysis uses 'classic' CAF, which is simply the sum of allele frequencies of all high-confidence pLoF variants each at allele frequency <5%, capped at a total of 100%, for both global and Finnish exomes. The results show that even when the sample size is held constant, the number of genes with zero pLoF variants observed is higher in a bottlenecked population than in a mostly outbred population. A constant y axis with no axis breaks is used in this figure to make this difference more clearly visible.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.

n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted

Software and code
Policy information about availability of computer code Data collection Analyses utilized Python 2.7.10 and R 3.5.1. Data and code sufficient to produce the plots and analyses in this paper are available at https://github.com/ericminikel/drug_target_lof Data analysis Analyses utilized Python 2.7.10 and R 3.5.1. Data and code sufficient to produce the plots and analyses in this paper are available at https://github.com/ericminikel/drug_target_lof For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design
All studies must disclose on these points even when the disclosure is negative.

Sample size
This study was opportunistic, and involved secondary use of all available genome and exome data. No sample size was predetermined. Our flagship analysis of gnomAD loss-of-function variants (

Replication
We did not attempt to reproduce any findings in a separate dataset, as no other exome or genome sequencing dataset of comparable size exists.
Randomization As this was a population-based study, and not a case-control study, no randomization was performed.

Blinding
As this was a population-based study, and not a case-control study, blinding was not relevant.

Reporting for specific materials, systems and methods
We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. As an opportunistic collection of data, the participants in gnomAD were not selected based on age, gender, or genotypic information. As described above, individuals with severe pediatric disease, and known first disease relatives of those with severe pediatric disease were excluded. The population and dataset inclusion criteria are described in more detail by Karczewski  Note that full information on the approval of the study protocol must also be provided in the manuscript.