The landscape of regional missense mutational intolerance quantified from 125,748 exomes

Abstract

Missense variants can have a range of functional impacts depending on factors such as the specific amino acid substitution and location within the gene. To interpret their deleteriousness, studies have sought to identify regions within genes that are specifically intolerant of missense variation^1-12. Here, we leverage the patterns of rare missense variation in 125,748 individuals in the Genome Aggregation Database (gnomAD)¹³ against a null mutational model to identify transcripts that display regional differences in missense constraint. Missense-depleted regions are enriched for ClinVar¹⁴ pathogenic variants, de novo missense variants from individuals with neurodevelopmental disorders (NDDs)^15,16, and complex trait heritability. Following ClinGen calibration recommendations for the ACMG/AMP guidelines, we establish that regions with less than 20% of their expected missense variation achieve moderate support for pathogenicity. We create a missense deleteriousness metric (MPC) that incorporates regional constraint and outperforms other deleteriousness scores at stratifying case and control de novo missense variation, with a strong enrichment in NDDs. These results provide additional tools to aid in missense variant interpretation.

Main text

Over the last decade, exome and genome sequencing have enabled variant discovery across hundreds of thousands of individuals^13,17-21. These large reference databases have provided the opportunity to study selective forces acting on the human genome and to identify genomic regions under selective constraint by, for example, identifying regions with fewer variants than expected based on mutational models^13,18,22-25. Gene-level metrics of predicted loss-of-function (pLoF) variant depletion have proven to be valuable in variant classification and identification of novel disease genes^15,16,26-28. The functional impact and selective pressures relevant to missense variation, by contrast, remain challenging to predict, as the effect of a missense variant is governed by the gene housing the variant, the position of the variant in the gene, and the specific amino acid substitution caused by the variant. To address this, prior work has sought to identify regions within coding genes that are specifically intolerant of missense variation as a way to improve interpretation^1-12. Here, we expand upon previous work¹ and show a sub-genic measure of missense intolerance leveraging population-level variation facilitates variant classification and risk stratification for association studies with de novo, rare, and common variants.

We explored the patterns of rare missense variant presence or absence in 125,748 exomes in the Genome Aggregation Database (gnomAD) v2.1.1 on GRCh37 to quantify missense depletion at the sub-genic level. We searched 18,629 canonical protein-coding transcripts for variability in missense constraint, quantified as the number of rare (allele frequency [AF] < 0.1%) missense variants observed in gnomAD divided by the number expected under neutral evolution as estimated from previously described mutational models¹³(observed/expected [OE]). For each transcript, we applied a recursive search based on likelihood ratio tests over all potential rare missense sites looking for breaks that divide the transcript coding sequence (CDS) into distinct missense constraint regions (MCRs; Fig. 1a, b). We discover 5,127 transcripts (28%) harbor regional variability in missense constraint (Fig. 1c), i.e., have two or more MCRs (minimum coding length 49bp, median 461bp; Supplementary Fig. 1). We thus refine the resolution of missense constraint for 42% of coding sites (coding space in the 5,127 transcripts vs. 18,629 total assessed). After recalibrating the missense OE distribution over all potential sites of missense variants using MCR-wide rather than transcript-wide missense OE measurements, we discover widespread signatures of negative and neutral selection that are obscured when quantifying over the unit of whole transcripts (Fig. 1d). We find a larger proportion of the exome lies within strongly constrained sequences (5.6% vs. 1.7% at OE < 0.4; see Supplementary Note for OE threshold selection), and the mode of the distribution shifts toward an OE indicative of evolutionary neutrality at approximately 1 (40.6% vs. 36.5% at 0.9 < OE ≤ 1.1).

Fig. 1:

28% of protein-coding genes in the human genome are discovered to harbor regional variation in population-level missense depletion.

a, An example transcript that has two missense constraint regions (MCRs) with significantly different levels of population-wide missense depletion, defined as the number of missense variants observed in gnomAD at rare frequency (AF < 0.1%) divided by the number of rare missense variants expected under neutral evolution (observed/expected or OE). Lower OE values correspond to greater variant depletion in the population and suggest stronger constraint. b, Flow chart describing the process of searching for breakpoints that divide a transcript into multiple MCRs. Searching for breakpoints is recursive and leverages likelihood ratio tests at a significance threshold of p = 0.001. c, The number of MCRs within the 5,127 transcripts discovered to harbor regional differences in missense constraint. The other 13,502 transcripts are deemed to have a single MCR (that is, a constant level of constraint across their entirety) and are not shown. d, The distribution of local missense OE at all coding sites in canonical transcripts. Local missense OE is defined as the OE calculated over the whole transcript (for “transcript OE”) or over the MCR (for “MCR OE”) where the site is located. Transcript OE and MCR OE are equivalent for transcripts with one MCR.

Furthermore, we find that constrained MCRs overlap established disease-associated mutational hotspots, including critical protein domains. One example is in the well-characterized KCNQ1, a voltage-gated potassium channel gene, in which pathogenic variants cause cardiac disorders such as long QT syndrome. We discover one moderately constrained MCR (missense OE = 0.60) overlapping the highly conserved C-terminus ²⁹ and another (missense OE = 0.66) encompassing the voltage-sensing and pore domains (Fig. 2a). Both the C-terminus of KCNQ1 and its voltage-sensing domain are established “hotspot” regions (specific missense-constrained regions with ACMG/AMP hotspot/functional domain moderate support [PM1] for pathogenicity)^29-31. All but two ClinVar pathogenic/likely pathogenic (P/LP) missense variants in this gene fall within these two missense-constrained MCRs.

Fig. 2:

Missense constraint regions (MCRs) and the distribution of ClinVar pathogenic/likely pathogenic (P/LP) missense variants in two genes associated with early-onset developmental disorders. Exons are delineated with black outlines and MCRs are delineated by color. MCRs are colored based on their missense observed/expected (OE) ratio, and MCRs with missense OEs not significantly different from 1 (p > 0.001) are shaded gray. a, KCNQ1. Only two of the 210 P/LP missense variants in KCNQ1 do not fall within either constrained MCR. The first constrained MCR encompasses the voltage-sensing and pore domains of this gene, and the other constrained MCR overlaps the C-terminus. Both domains contain previously reported hotspot regions, with some regions reaching moderate level (PM1) support for pathogenicity³¹. Ion transport protein: domain that contains both the transmembrane voltage-sensing and pore domains. KCNQ channel: C-terminal cytoplasmic domain that overlaps four helices (A-D). b, BAP1. Variants in this gene can lead to cancer-predisposition syndromes, increased risk of certain cancers, or the neurodevelopmental disorder Kury-Isidor syndrome³². All of the ClinVar P/LP variants associated with Kury-Isidor fall within the first MCR with a highly depleted missense OE of 0.33. An additional five variants reported in Kury et al.³² but not ClinVar fall within either highly constrained MCR in this gene. P/LP variants associated with Kury-Isidor are colored in black, and all other cancer-associated P/LP variants are colored in gray. UCH: Ubiquitin carboxyl-terminal hydrolase isozyme L5 domain. ClinVar data are from the October 15, 2023 release.

We also find that missense constraint within MCRs is able to identify regions associated with severe, early-onset disease. One example is in BAP1, which plays a key role in chromatin modeling by mediating histone deubiquitination. Disease-causing variants in this gene are linked to cancer or, as recently discovered, Kury-Isidor syndrome³². The first highly missense-constrained MCR (missense OE = 0.33) in BAP1 encompasses the ubiquitin C-terminal hydrolase domain connected to Kury-Isidor syndrome³² (Fig. 2b), and all 11 variants reported to be causal for Kury-Isidor fall within this MCR. The only ClinVar P/LP variants that do not fall within any missense-constrained MCRs in BAP1 are associated with cancer phenotypes, which may be under weaker selection than neurodevelopmental disorders (NDDs).

Next, we sought to determine if the signatures of selection revealed by MCRs recapitulated biological and disease relevance of coding sequences. Overall, most transcripts that are intolerant to pLoF variation (as measured by the loss-of-function observed/expected upper bound fraction [LOEUF] score¹³) also tend to be intolerant to missense variation. This trend is markedly more prominent when measuring missense constraint at the sub-genic level vs. the transcript-level (Supplementary Note; Supplementary Fig. 2). We also discovered that 64% (1697/2659) of genes that are both LOEUF- and MCR missense-constrained do not have disease associations in OMIM³³, suggesting the existence of many undocumented genes containing variants of significant consequence for disease (Supplementary Fig. 3). In a set of 730 strongly mutationally intolerant genes, defined here as exhibiting both population depletion of pLoF variants (first three LOEUF deciles) and association with a developmental phenotype (high-confidence membership in any non-cancer Gene2Phenotype [G2P]³⁴ gene list with dominant inheritance), we observed strong transcript-wide missense depletion that was even stronger for genes with multiple MCRs (Fig. 3a and Supplementary Fig. 4; Wilcoxon p < 10⁻⁵⁰). Given that we have greater power to detect missense constraint variability over longer sequences (Supplementary Fig. 5), we controlled for transcript length but still found that intolerant transcripts are eight times more likely to harbor multiple MCRs (p < 10⁻⁵⁰). These strongly intolerant transcripts are highly enriched for severely depleted regions (three times more likely to have minimum MCR OE < 0.4 after regressing out transcript length, p < 10⁻¹⁸), whereas the most constrained MCRs in not strongly intolerant transcripts are less depleted and more evenly distributed across the OE spectrum. Finally, we observe a group of genes with strong overall missense depletion in which we did not detect multiple MCRs (n = 459 with missense OE < 0.4; Supplementary Table 1), suggesting these genes are robustly intolerant to missense variants across their length. When comparing missense constraint to selection over longer timescales (measured by evolutionary conservation in placental mammals, phyloP³⁵), we found that genes with more conserved coding sequences also tended to be more overall depleted of human missense variation (Spearman ρ = 0.56, p < 10⁻⁵⁰). However, a substantial number of strongly constrained MCRs appear widely unconserved across mammals, potentially pointing to human-specific negative selection pressures that are obscured at the whole-transcript level (Supplementary Fig. 6).

Fig. 3:

Regional missense depletion reveals constraint obscured by gene-level measures. a, Left: The distribution of transcript-wide missense observed/expected (OE) across 18,629 transcripts stratified by the combination of two factors: whether the transcript is strongly mutationally intolerant (within first three LOEUF deciles and association with a developmental phenotype in Gene2Phenotype [G2P]³⁴) and whether we detect multiple missense constraint regions (MCRs). Number of transcripts in each category are: strongly intolerant with multiple MCRs (n=581; red), strongly intolerant with one MCR (n=149; yellow), not strongly intolerant with multiple MCRs (n=4,546; dark gray), not strongly intolerant with one MCR (n=13,353; light gray). X-axis is cut off at 1.6 for visibility. Right: Minimum MCR missense OE using the same groupings. Minimum MCR missense OE is the same as transcript missense OE for transcripts with a single MCR. b, MCR missense OE at all sites of possible exome-wide missense variants vs. sites of de novo missense variants in controls, autistic individuals (AUT), or individuals with DD. De novo variants from individuals with developmental phenotypes are enriched in more constrained sequences, with a more pronounced enrichment in DD than autism. c, Enrichment in per-variant heritability explained by common (AF > 5%) protein-coding SNPs stratified by MCR missense OE quintile, relative to the average SNP genome-wide. Enrichment is estimated by linkage disequilibrium score regression, accounting for number of SNPs in each quintile, and is averaged across 268 independent traits 198 in UKBB and other large genome-wide association studies. Black dashed line at 1 indicates no enrichment. Blue dashed line at 3 indicates average coding enrichment. Error bars represent 95% confidence intervals.

We next aggregated de novo missense variants from 31,058 individuals with a severe developmental disorder¹⁵ (DD), 15,036 autistic individuals (AUT), and 5,492 siblings not diagnosed with a DD¹⁶ (Fig. 3b). The distribution of de novo missense variants across the missense OE spectrum in unaffected siblings largely mirrored the exome-wide missense OE distribution. In contrast, de novo missense variants in autistic individuals are enriched in missense-constrained sequences, and this pattern is more striking in individuals with DDs. For example, relative to unaffected siblings, the rate of de novo missense variants in MCRs with OE < 0.2 is 2-fold higher in autistic individuals (p < 10⁻²³) and 6.6-fold higher in individuals with DDs (p < 10⁻⁵⁰) (Supplementary Fig. 7; see Supplementary Note for OE threshold selection). This is consistent with the expectations that a small subset of de novo missense variants in individuals with developmental phenotypes are causal for those traits and that variants causal for DD are generally more selectively deleterious than those for autism.

Beyond large-effect rare and de novo variation in traits under strong negative selection, we additionally investigated whether our MCR metric, which was calculated using rare variants, correlates with functional effects of common variants. Prior work found that pLoF-constrained genes and their flanking 100kb sequences are enriched for SNP heritability across hundreds of independent traits in the UK Biobank (UKBB) and other large genome-wide association studies (GWAS) ¹³. We partitioned common (AF > 5%) variant heritability of the same 268 independent traits across MCRs to investigate relative enrichment. To establish a baseline, we computed the heritability enrichment over all coding sequences comprising MCRs (3-fold). The most constrained MCRs have the strongest heritability enrichment; the first quintile of MCR missense OE harbors a 41-fold enrichment (Fig. 3c). Coding SNPs in missense-unconstrained MCRs (e.g., in the two least constrained quintiles of MCR missense OE) harbor no detectable heritability enrichment relative to the average genome-wide SNP. These findings suggest that: 1) regions depleted of rare missense variation can help prioritize common coding variants important for complex traits (i.e., improve GWAS fine-mapping variant prioritization), and 2) there exists a subset of coding sequence with no appreciable heritability enrichment, which rare variant depletion can help identify.

We examined the localization of high-quality ClinVar³⁶ missense variants classified as P/LP within genes with both unconstrained (missense OE > 0.9) and constrained (missense OE < 0.2) MCRs and found that P/LP variants occur much more frequently in missense constrained MCRs (odds ratio [OR] = 15.2; p < 10⁻⁵⁰). We also examined the localization of P/LP and benign/likely benign (B/LB) variants within MCRs in autosomal dominant disease-associated genes and found that P/LP variants tend to localize to regions that are more strongly missense-constrained than the overall transcript (Wilcoxon p = 3.5x10⁻¹⁰), while B/LB variants show the opposite effect and tend to occur in regions with OEs closer to 1 (Wilcoxon p < 10⁻¹⁸; Fig. 4a). While more subtle, these same patterns are also significant in autosomal recessive disease-associated genes (Supplementary Fig. 8).

Fig. 4:

ACMG/AMP calibration of missense constraint. a, The distribution within genes with autosomal dominant disease associations of transcript-wide missense observed/expected (OE; gray) and missense constraint region (MCR) OE for ClinVar pathogenic/likely pathogenic missense variants (P/LP; red) and benign/likely benign missense variants (B/LB; blue). We filtered to 1,007 transcripts with at least one P/LP and one B/LB missense variant. For the P/LP and B/LB distributions, we annotated each variant with the missense OE across the MCR they fell in and collapsed these values within each transcript by taking the respective medians. b, Local posterior probabilities of pathogenicity given MCR missense OE in all transcripts. Gray shading indicates the one-sided 95% confidence interval on the more stringent side. Horizontal lines indicate thresholds required to meet ACMG/AMP evidence levels. From bottom to top: supporting, moderate, strong, very strong. MCR missense OE reaches supporting (OE ≤ 0.37) and moderate (OE ≤ 0.21) level evidence for PM1 (hotspot/functional domain).

To enable use of our missense constraint metric in ACMG/AMP clinical variant classification, we applied previously established probabilistic frameworks³⁷ to determine the MCR missense OE thresholds that met different levels of clinical evidence strengths evaluated under the hotspot/functional domain (PM1) and benign in silico prediction (BP4) criteria codes³⁰. MCR missense OE ≤ 0.37 met supporting (PM1_Supporting) and OE ≤ 0.21 met moderate (PM1) levels of evidence for pathogenicity (Fig. 4b), but no MCR missense OE threshold met any levels of evidence to support benignity. However, separate calibration specifically in transcripts with multiple MCRs found that MCR missense OE ≥ 1.56 met moderate and OE ≥ 0.97 met supporting evidence for BP4, indicating that in transcripts where we are powered to characterize regional constraint, MCRs with OEs close to one harbor an indication of benignity (Supplementary Fig. 9). Calibration of two additional regional constraint metrics, Constrained Coding Regions (CCRs⁹) and COntact Set MISsense tolerance (COSMIS¹²), which incorporates predicted 3D structure information, revealed that these metrics also reach moderate support for pathogenicity (PM1), and COSMIS only reaches supporting levels for benignity (Supplementary Fig. 9; Supplementary Table 2).

We transformed our regional missense constraint measure into a variant-level predictor of missense deleteriousness named MPC (Missense deleteriousness Prediction by Constraint) that additionally incorporates information about amino acid substitution type and local context. The logistic regression-based model integrates regional missense constraint-derived metrics together with BLOSUM³⁸, Grantham³⁹, and PolyPhen-2⁴⁰ and is trained on ClinVar pathogenic and gnomAD common (AF > 0.1%) variants in 2,987 genes defined as haploinsufficient in Collins et al.⁴¹ and 366 genes with DD associations in G2P through non-LoF mechanisms. Higher scores predict greater deleteriousness (Supplementary Fig. 10, 11). We assessed the utility of MPC in prioritizing potentially disease-causing variation by evaluating its ability to stratify case and control rare and de novo missense variation. Consistent with the regional constraint results, the de novo missense variants from DD and AUT cases are enriched for high MPC scores compared to controls (Supplementary Fig. 12). We further stratified by presence in 373 genes previously associated with NDD¹⁶ and three bins of MPC scores (< 1.6, 1.6-2.6, ≥ 2.6; see Supplementary Note for calibration of these bins), and found a very strong enrichment of de novo missense variants in the two most deleterious bins among both the DD (Fig. 5a) and AUT cases (Fig. 5b) compared to unaffected individuals. However, while the enrichment in the 373 NDD-associated genes was significant for missense variants with MPC ≥ 2.6 (RR in DD cases = 22.7, p < 10⁻⁵⁰; RR in AUT cases = 6.9, p < 10⁻²¹), as well as missense variants with MPC between 1.6-2.6 (RR in DD cases = 4.5, p < 10⁻³⁵; RR in AUT cases = 1.9, p = 3.0x10⁻⁵), it was only significant in NDD-unassociated genes for missense variants with MPC ≥ 2.6 (RR in DD cases = 3.1, p < 10⁻²⁸; RR in AUT cases = 1.5, p = 5.9x10⁻⁴). This suggests that while there is a sizable reservoir of potentially causal variants in genes yet to be associated with NDDs, they will be more difficult to find as they must reach stricter deleteriousness criteria. For autism, we additionally assessed inheritance rates of rare missense variants (AF < 0.1%) from parents to 13,384 probands and case-control rates for an additional 5,591 cases and 8,597 controls without de novo information. While we did not find substantial enrichment in inheritance rates in any missense category, we discovered substantial enrichment in the case-control analysis for variants in the 373 NDD-associated genes with MPC ≥ 2.6 (RR = 1.6, p < 10⁻¹²), which we infer is from de novo variants that are not recognizable as such due to lack of parental information.

Fig. 5:

MPC effectively stratifies case and control variation.

a, The difference relative to controls of missense variants stratified by MPC score and localization to genes associated with neurodevelopmental disorders (NDDs) for a, individuals with DD and b, autistic individuals (AUT). Relative difference is calculated as: for de novo variants, the average rate of variants in probands divided by that in sibling controls; for case-control, the average rate of variants in cases divided by that in controls from case-control data; for inherited, the average rate in probands of transmitted variants divided by that of untransmitted variants. Error bars represent 95% confidence intervals calculated from a binomial test. c, The odds ratio of case to control de novo missense variants in the top 10% vs. bottom 90% of respective rankings. De novo missense variants from each case cohort are ranked against those in the 5,492 controls for each predictor. DD: developmental disorders, EE: epileptic encephalopathy, AUT: autism, OFC: orofacial cleft, CHD: congenital heart disease, CDH: congenital diaphragmatic hernia. Error bars represent 95% confidence intervals. Only variants scored by all predictors are included. Points are solid colored if the difference from 1 is statistically significant (binomial or Fisher exact p < 0.05).

We extended our assessment of case-control de novo stratification for a comparison of our model against several other missense deleteriousness predictors: AlphaMissense⁴², CCRs⁹, M-CAP⁴³, REVEL⁴⁴, PrimateAI-3D⁴⁵, MVP⁴⁶, Polyphen-2⁴⁰, CADD^47,48, mammalian conservation phyloP³⁵, and SIFT⁴⁹. For this assessment, we evaluated four additional early-onset development-related phenotypes: epileptic encephalopathy (EE), orofacial cleft (OFC), congenital heart disease (CHD), and congenital diaphragmatic hernia (CDH). To compare across predictors with different score distributions, we used a ranking-based performance assessment. For each predictor, we ranked the de novo missense variants from each case cohort against those in the 5,492 controls and computed the OR of case vs. control variants in the top percentiles of these rankings (Fig. 5c). At the top 10% of variants, MPC displays the highest OR for DD (OR = 5.2, Fisher’s exact p < 10⁻⁴⁸), EE (OR = 3.1, p = 2.2x10⁻⁷), AUT (OR = 1.7, p = 8.9x10⁻⁹), and OFC (OR = 1.5, p = 0.025), although there is substantial confidence interval overlap with other predictors. This indicates that MPC effectively ranks high-impact de novo variants in the most deleterious prediction regimes. Of the other predictors, AlphaMissense also performs consistently well across all phenotypes. In particular, in CHD and CDH, which have the least de novo enrichment across predictors, we observe MPC lagging in performance, while AlphaMissense is one of the top performers. This may suggest that causal de novo variants in these phenotypes may occur at a narrow set of sites where 3D structure is important, which AlphaMissense can more deftly capture through integration of protein structure prediction. These observations are more or less consistent over a range of thresholds used to define the ranking top percentiles (Supplementary Fig. 13).

We have developed a method to identify sub-genic regions with differential intolerance to missense variation at base-level resolution. We demonstrate that coding regions depleted for missense variation in the general population are enriched for established disease-associated variation, de novo variants from individuals with NDDs, and heritability for 268 complex traits from the UK Biobank and other large GWAS. Additionally, we have calibrated these constraint scores to establish that regions with less than 20% of their expected variation can achieve moderate evidence for association to disease following ACMG/AMP guidelines. Finally, we incorporated regional missense intolerance information into the missense deleteriousness metric, MPC, and show that MPC effectively separates potentially risk-carrying variants identified in various developmental disorder cases from those seen in controls.

At current sample sizes, we are unable to characterize constraint at single amino acid resolution. Furthermore, because our approach relies on variant presence or absence in a large reference dataset, many of the constrained regions we find are linked to variants that cause severe, early-onset disease. However, the true nature of the variation we capture is more accurately linked to reproductive fitness and the strength of selection acting on heterozygotes⁵⁰. Our methodology specifically searches for linear sub-genic regions in canonical transcripts that are depleted of missense variants compared to a null mutational model. This means that our model is unable to find depleted sequences that are clustered specifically in 3D space and is also currently ignorant of coding sequences not present in the Ensembl canonical transcript. However, we note that our linear metric achieves similar evidence for both pathogenicity and benignity as the structural constraint-based COSMIS model¹² (Supplementary Fig. 9).

In summary, we identify 28% of canonical transcripts with variable levels of missense constraint and demonstrate that coding regions specifically depleted of missense variation in the general population are enriched for disease-associated variation. Additionally, we show that this depletion of missense variation can be used as moderate evidence when classifying variants according to ACMG/AMP guidelines and that incorporation of regional missense constraint into an in silico predictor effectively prioritizes a subset of de novo missense variation in individuals with developmental phenotypes for association testing. We have publicly released these data for use in both research and clinical settings. We anticipate refined resolution of these metrics as datasets grow, both in size and in ancestral diversity, and with the incorporation of complementary structural or functional data.

Methods

Transcripts

This study analyzed only canonical, coding transcripts as defined by GENCODE v19/Ensembl v74. We excluded the same set of transcripts from this analysis that were excluded in the previous gnomAD v2.1.1 genic constraint estimates¹³. Briefly, we excluded transcripts that had outlier counts of variants expected under neutrality (zero expected pLoF, missense, or synonymous variants; too many observed pLoF, missense, or synonymous variants compared to expectation; or too few observed synonymous variants compared to expectation). In total, this study analyzed 18,629 transcripts.

gnomAD variants

All analyses in this paper were conducted using the 125,748 gnomAD v2.1.1 exomes¹³ on GRCh37. Median coverage was calculated on a random subset of the gnomAD exomes as described previously¹³. We defined the set of sites with possible missense variants using a synthetic Hail Table (HT) containing all possible single nucleotide variants in the exome. We annotated this HT with the Variant Effect Predictor (VEP, version 85) against GENCODE version 19, and filtered to variants with the consequence "missense_variant" in the canonical, coding transcripts as defined in Transcripts. We then further filtered to variants that fit one of the following criteria: (1) allele count (AC) > 0 and AF < 0.001, variant QC PASS, and median coverage > 0 in gnomAD v2.1.1 exomes; or (2) AC = 0, i.e. variants not seen in gnomAD v2.1.1 exomes.

ClinVar variants

We annotated functional consequences for ClinVar¹⁴ (v.20230305) variants using the VEP table described in gnomAD variants. Missense ClinVar variants with non-conflicting P, LP, B, LB classification and a review status of at least one star were selected for analysis.

Rare and de novo variants from developmental cohorts

Case de novo mutations for association analyses were obtained from studies of developmental disorders¹⁵ (DD), autism¹⁶ (AUT), congenital heart disease⁵¹ (CHD), orofacial cleft⁵² (OFC), congenital diaphragmatic hernia⁵³ (CDH), and epileptic encephalopathy⁵⁴ (EE). Control de novo mutations were obtained from neurotypical siblings of the autistic probands¹⁶. Variants from the autism study were lifted over from GRCh38 to GRCh37 using the “liftover” function in Hail. Variant functional consequences were re-annotated using the VEP table described in gnomAD variants. Variants transmitted and not transmitted from parents to autistic probands were procured from previously published ASC-SSC and SPARK cohorts, and case-control variants for autism were procured from previously published iPSYCH and Swedish cohorts¹⁶. Both the inherited/uninherited and case-control variant sets were filtered to AF < 0.1%.

Training, validation, and test datasets

To generate independent training and test sets, we selected 80% (14,894 transcripts) of the 18,629 canonical coding transcripts to comprise the training set and the remaining 20% (3,735 transcripts) to the test set. To ensure the training and test transcripts have similar distributions of features that may impact constraint estimates, we used stratified randomization to match the training and test transcripts on s_het coefficients (as a measure of selection) and number of potential missense sites (as a measure of power to detect transcript-wide constraint changes). The training set was used for MPC model training and MCR model selection, and the test set was used for MPC model evaluation. No similar hold-outs of data were performed for training of the mutational model used to compute expected variant counts (see Modeling of mutation rates and expected neutral missense variation).

Modeling of mutation rates and expected neutral missense variation

Expected missense variant counts were determined as described previously¹³. Briefly, we created a model using the 15,708 gnomAD v2.1.1 genomes that estimated the mutation rate for each single nucleotide substitution with one base of context (e.g., ACT > AGT) in non-coding regions of the genome. We then calibrated this mutation rate against the proportion observed of each context at synonymous sites to adjust for the larger size of the gnomAD v2 exomes, adjusting for low coverage regions (median coverage < 40x) and methylation levels at CpG sites using methylation data from the Roadmap Epigenomics Consortium⁵⁵. We created three separate models (referred to as "plateau" models moving forwards): one for autosomal and pseudoautosomal sites, one for chromosome X sites, and one for chromosome Y sites. Each of these models contains mutation rate estimates for each substitution, context, and methylation level. We then applied the plateau models to the proportion observed of each substitution and its context, exome coverage, and methylation level. We counted all possible variants in our synthetic Hail Table (HT) that passed the following criteria: (1) Median coverage > 0; (2) no low-quality variant observed in gnomAD v2 exomes; (3) no variants above 0.1% AF observed in gnomAD v2 exomes. We then correlated this proportion observed value with the mutation rate calculated using the appropriate model above. For low coverage sites (median coverage below 40x), we calculated a scaling factor as described previously¹³: briefly, we computed the total number of observed synonymous variants in the gnomAD v2 exomes divided by the total number of possible synonymous variants in the synthetic HT multiplied by the mutation rate aggregated across all possible substitutions and their contexts and methylation levels. We used this scaling factor to create a model to adjust the proportion of expected variation for low coverage sites (coverage model).

Identifying breakpoints within transcripts of regional missense constraint

Observed missense variant counts were calculated using sites from the 125,748 gnomAD exomes that passed all the following criteria: (1) Allele count (AC) > 0; (2) allele frequency (AF) < 0.001; (3) variant QC PASS (passed gnomAD variant QC filters, including random forest filters); (4) median coverage > 0. We filtered the gnomAD v2 exomes Hail Table (HT) to the sites that matched the above criteria and then annotated the synthetic HT with whether that variant (chromosome/locus plus reference and alternate alleles) was observed in the gnomAD exomes. We then aggregated the total number of observed variant counts per locus by summing the number of observed variants for each possible substitution (reference and alternate allele) at each locus. Finally, we grouped the synthetic HT by transcript annotation to sum the total number of observed missense variants per transcript.

As previously described¹³, we applied the two models (plateau and coverage) described in Modeling of mutation rates and expected neutral missense variation to calculate the total proportion of expected missense variation. Briefly, we summed the mutation rate (mu_agg) for each substitution, context, and methylation level across the exome. We then applied the appropriate plateau model (autosomal/pseudoautosomal, chromosome X, chromosome Y) and adjusted CpG vs. non-CpG sites separately. After applying the appropriate plateau model, we applied the coverage model to low coverage (median coverage < 40x) sites to create the final adjusted mutation rate (mu_adj). We then aggregated the raw mutation rate sum (mu_agg) and mutation rate (mu_adj) per transcript to get the total mutation rate sum and proportion of expected missense variation per transcript.

We implemented a minimum number of expected missense variants to prevent finding breakpoint positions that would create very small (i.e., a handful of base pairs in size) transcript subsections (see Supplementary Note).

We applied a likelihood ratio test to determine whether the missense observed/expected (OE) ratio was uniform along a transcript or whether a transcript had evidence of distinct sections of missense constraint. We used the observed and expected missense counts to search for positions that would divide a transcript into two or more regions with varying levels of missense depletion. For our analyses, we assume that the observed missense counts should follow a Poisson distribution around the expected missense counts. We defined our null model as transcripts not having any evidence of regional variability in missense depletion (where the expectation, the OE ratio, is consistent across the length of the transcript). Our alternative model was that transcripts exhibited evidence of distinct sections of missense depletion (OE ratio calculated per transcript subsection). Because the alternative model should always have a better fit than the null model, we require a chi square value above a given threshold (p = 0.001) to establish significance. We used the following formulas to determine the significance of a breakpoint that would split a transcript into two sections, A and B:

• $p_0=Pois(ob{s}_A,ex{p}_A\ast OE)\ast Pois(ob{s}_B,ex{p}_B\ast OE)$

• $p_1=Pois(ob{s}_A,ex{p}_A\ast O{E}_A)\ast Pois(ob{s}_B,ex{p}_B\ast O{E}_B)$

• ${\chi }^2=2(ln(p_1)-ln(p_0))$

where OE is the missense observed/expected ratio across the entire transcript, obs_A is the number of observed missense variants in transcript section A, exp_A is the number of expected missense variants in transcript section A, OE_A is the OE ratio across transcript section A, obs_B is the number of observed missense variants in transcript section B, exp_B is the number of expected missense variants in transcript section A, OE_B is the OE ratio across transcript section B, and Pois is the Poisson likelihood.

We used the following formulas to determine the significance of a breakpoint that would split a transcript into three sections, A, B, and C:

• $p_0=Pois(ob{s}_A,ex{p}_A\ast OE)\ast Pois(ob{s}_B,ex{p}_B\ast OE)\ast Pois(ob{s}_C,ex{p}_C\ast OE)$

• $p_1=Pois(ob{s}_A,ex{p}_A\ast O{E}_A)\ast Pois(ob{s}_B,ex{p}_B\ast O{E}_B)\ast Pois(ob{s}_C,ex{p}_C\ast O{E}_C)$

• ${\chi }^2=2(ln(p_1)-ln(p_0))$

where OE is the missense observed/expected ratio across the entire transcript, obs_A is the number of observed missense variants in transcript section A, exp_A is the number of expected missense variants in transcript section A, OE_A is the OE ratio across transcript section A, obs_B is the number of observed missense variants in transcript section B, exp_B is the number of expected missense variants in transcript section A, OE_B is the OE ratio across transcript section B, obs_C is the number of observed missense variants in transcript section C, exp_C is the number of expected missense variants in section C, and Pois is the Poisson likelihood.

For the purposes of our analyses, all transcript subsections with more observed variants than expected were capped at an OE of 1, as we were looking for areas of missense depletion and not missense enrichment. We also converted the expected counts for transcript subsections with zero expected variants from 0 to 10⁻⁹ to prevent nonfinite OE values.

To search for a single breakpoint that would divide a transcript into two subsections, we calculated chi square statistics (as discussed above) to conduct likelihood ratio tests simultaneously for every eligible position within a transcript. The positions we considered were positions with a possible missense variant substitution that had at least 16 expected missense counts in either direction (i.e., both transcript subsections created by dividing the transcript at this point would have at least 16 expected missense variants). We then aggregated chi square values across each transcript to find the maximum value per transcript, and we marked any positions as breakpoints if the chi square calculated at that position was equal to the maximum chi square value over all sites in the transcript and significant at p = 0.001.

Any transcripts that did not have a single significant breakpoint moved forwards into our two simultaneous breaks search flow. In this search flow, we again calculated chi square statistics to conduct likelihood ratio tests for every eligible position pair. For every position with a possible missense, we calculated the chi square statistic of that position paired with each possible position downstream as long as the two positions created transcript subsections with at least 16 expected missense variants (i.e., all three of the transcript subsections created would have at least 16 expected missense variants). Because of the large number of pairwise computations, this step is the most computationally intensive portion of our algorithm. After completing the single and two simultaneous break search workflows, we merged the results from both search types.

Our breakpoint search flow is recursive, and the steps are as follows: Search for a single significant breakpoint dividing a transcript into two subsections. If no single significant breakpoint was found in the transcript, search for two simultaneous breakpoints. Merge the results from the single and two simultaneous breakpoint searches. Repeat the steps above, treating each separate transcript subsection as if it were an independent transcript, until no more significant breakpoints are found.

Modeling deleteriousness of missense substitution classes with missense constraint

We incorporated two MCR OE-based metrics to measure the increased deleteriousness of amino acid substitution classes (e.g., Met to Tyr) in functionally important areas of proteins: the overall OE for each substitution and the second derivative of this OE value per OE bin of missense constraint (Supplementary Fig. 14). To calculate the first metric, the substitution overall OE, we divided the total number of rare, high quality variants (see gnomAD variants) causing that substitution by the total number of expected variants (see Modeling of mutation rates and expected neutral missense variation). To calculate the second metric, the substitution OE second derivative, we aggregated the OEs of each substitution by MCR OE bin in 10 bins from 0 to 1.0+ (i.e., for the 0-0.1 OE bin, we calculated all of the observed substitutions that occurred within regions with a OE between 0 and 0.1 and divided that number by the total number of expected substitutions occurring in those regions).

Modeling deleteriousness of individual missense variants

We designed a missense variant deleteriousness predictor to explicitly incorporate information on amino acid substitution class and position-specific variant effects. A logistic regression model was first trained to differentiate pathogenic from benign missense variants. The pathogenic training set consisted of high-quality ClinVar variants (see ClinVar variants) labeled as pathogenic or likely pathogenic in 2,987 likely-haploinsufficient genes, defined as having probability of haploinsufficiency (pHaplo) ≥ 0.86⁴¹, or in 366 genes with DD associations in G2P through non-LoF mechanisms. The latter gene set was created by filtering on the G2P DD panel (accessed October 6, 2023) to select genes where: 1. confidence_category is either definitive or strong evidence, 2. allelic_requirement was monoallelic, and 3. mutation_consequence included altered gene product structure or increased gene product level. The benign training set consisted of high-quality common variants as described in gnomAD variants. Variants matching criteria for both the benign and pathogenic sets were removed from the training data. We evaluated models with all possible combinations of the following complementary features: amino acid substitution overall OE and OE second derivative (see Modeling deleteriousness of missense substitution classes); BLOSUM³⁸ and Grantham³⁹ scores of amino acid substitution class severity; the local missense constraint level of a variant (missense OE across the MCR if applicable, else across the transcript); and PolyPhen-2⁴⁰. We selected BLOSUM, Grantham, and PolyPhen-2 because of the orthogonal information added on top of our OE-based (and therefore population allele frequency-dependent) metrics. For each model, only variants with all relevant annotations were used in training the regression model and the subsequent calculations to produce deleteriousness scores. The deleteriousness score prediction for any missense variant $i$ is given as:

$$\begin{gathered} d_i=-lo{g}_{10}(m_i∕M) \\ {m}_i=max(0.83,f_i) \end{gathered}$$

where $d_i$ is the deleteriousness score prediction, $f_i$ is the number of common missense variants with a fitted value from the regression that is less than the fitted value for variant $i$, and $M$ is the number of common missense variants (equivalent to the number of benign training variants for the regression). $m_i$ is set to have a minimum value of 0.83 to avoid a mathematical error in the log when the fitted value for a given variant is less than those of all common variants. Larger values of $d_i$ indicate stronger predicted-deleteriousness. The best model was chosen to be the model featurized with all six possible features. The training set for this model consisted of 64,023 benign variants and 12,955 pathogenic variants. This model was then applied to produce MPC scores for the 68,576,965 possible exome-wide missense variants with all features, and the distribution of these MPC scores is given in Supplementary Figs. 10, 11, and 12.

Comparison of MPC to other predictors

We compared our model to the following missense deleteriousness predictors: AlphaMissense⁴², Constrained Coding Regions (CCRs)⁹, MVP⁴⁶, M-CAP⁴³, PrimateAI-3D⁴⁵, REVEL⁴⁴, CADD^47,48, PolyPhen-2⁴⁰, and SIFT⁴⁹. We annotated the case and control de novo missense variants described in Rare and de novo variants from developmental cohorts and ranked the variants based on their annotated scores. To assess each predictor's ability to stratify case and control variation, we assessed the proportion of case to control variants among the variants with the top 10% for each score and compared this number to the overall proportion of case to control variation using a Fisher exact test.

Data availability

The missense constraint regions (MCRs) are displayed on the gnomAD v2 browser (https://gnomad.broadinstitute.org) and available for download on the gnomAD website (https://gnomad.broadinstitute.org/downloads#v2) and in the gnomAD v2 public datasets on Google, Amazon, and Microsoft clouds. MPC scores for all possible variants in canonical transcripts is available in the gnomAD v2 public datasets on Google (gs://gcp-public-data--gnomad/release/2.1.1/regional_missense_constraint/gnomad_v2.1.1_mpc.ht). gnomAD v2 exome, genome, and coverage data and the table of all possible single nucleotide polymorphisms used to calculate mutational models and search for MCRs are also available in the gnomAD public buckets and are easily accessed using code in the gnomAD Hail utilities GitHub repository (https://github.com/broadinstitute/gnomad_methods/blob/7c0c994883f321492a48962674d5caeb289df4c7/gnomad/resources/grch37/gnomad.py#L107 and https://github.com/broadinstitute/gnomad_methods/blob/7c0c994883f321492a48962674d5caeb289df4c7/gnomad/utils/vep.py#L161).

ClinVar data were downloaded from ClinVar's FTP server (https://ftp.ncbi.nlm.nih.gov/pub/clinvar/). De novo variants were extracted from the supplemental files of the cited studies.

AlphaMissense scores were downloaded from https://github.com/google-deepmind/alphamissense. CCRs were downloaded from https://github.com/quinlan-lab/ccrhtml. MCAP scores were downloaded from http://bejerano.stanford.edu/MCAP/. REVEL scores were downloaded from https://sites.google.com/site/revelgenomics/. PrimateAI-3D scores were downloaded from https://primad.basespace.illumina.com/download. MVP scores were downloaded from https://figshare.com/articles/dataset/Predicting_pathogenicity_of_missense_variants_by_deep_l_earning/13204118. PolyPhen-2 and SIFT scores were obtained from VEP⁵⁶. CADD scores were downloaded from the CADD website (https://cadd.gs.washington.edu/download). phyloP scores were downloaded from the UCSC browser (https://genome.ucsc.edu/cgi-bin/hgTrackUi?db=hg38&g=cons241way).

Code availability

Code to determine missense constraint regions (MCRs) and calculate MPC is available at https://github.com/broadinstitute/regional_missense_constraint. Code used to generate the mutational models is available at https://github.com/broadinstitute/gnomad_lof and https://github.com/broadinstitute/gnomad-constraint. The Hail library is available at https://hail.is/.