The landscape of tolerated genetic variation in humans and primates

Abstract

Personalized genome sequencing has revealed millions of genetic differences between individuals, but our understanding of their clinical relevance remains largely incomplete. To systematically decipher the effects of human genetic variants, we obtained whole genome sequencing data for 809 individuals from 233 primate species, and identified 4.3 million common protein-altering variants with orthologs in human. We show that these variants can be inferred to have non-deleterious effects in human based on their presence at high allele frequencies in other primate populations. We use this resource to classify 6% of all possible human protein-altering variants as likely benign and impute the pathogenicity of the remaining 94% of variants with deep learning, achieving state-of-the-art accuracy for diagnosing pathogenic variants in patients with genetic diseases.

One sentence summary: Deep learning classifier trained on 4.3 million common primate missense variants predicts variant pathogenicity in humans.

Fig. 1.. Common primate variants are largely benign in human.

(A) Counts of missense (solid green) and synonymous (shaded grey) variants from primates compared to the gnomAD database. Missense : synonymous counts and ratios are displayed above each bar. (B) Fractions of all possible human synonymous (grey) and missense variants (green) observed in primates. (C) Counts of benign (grey) and pathogenic (red) missense variants with two-star review status or above in the overall ClinVar database (left pie chart), compared to ClinVar variants observed in gnomAD (middle), and compared to ClinVar variants observed in primates (right). Conflicting benign and pathogenic annotations and variants interpreted only with uncertain significance were excluded. (D) Observed gnomAD (green) or primate (blue) missense variants in each amino acid position in the CACNA1A gene. Red circles represent the positions of annotated ClinVar pathogenic missense variants. Bottom scatterplot shows PrimateAI-3D predicted pathogenicity scores for all possible missense substitutions along the gene. (E) Multiple sequence alignment showing the ClinVar pathogenic variant chr11:77181548 G>A (red arrow) creating a cryptic splice site in human sequence (extended splice motif, blue). This variant is tolerated in Cebus Albifrons and other species with a G>C synonymous change in the adjacent nucleotide that stops the splice motif from forming. (F) Pie charts showing the fraction of benign (grey) and pathogenic (red) missense variants with ClinVar two-star review status or above in great apes, old world monkeys, new world monkeys, lemurs/tarsiers, mammals, chicken, and zebrafish. (G) Missense : synonymous ratios across the human allele frequency spectrum, with MSR of human variants seen in primates shown for comparison. The blue dashed line represents the expected missense : synonymous ratio of de novo variants. Colors and legend are the same as (A).

Fig. 2.. Selective constraint of primate genes compared to human.

(A) Scatter plot of missense : synonymous ratios between primate and human genes. Each gene is colored by its pLI score, with darker points showing haploinsufficient genes. (B) Observed and expected counts of synonymous (top) and missense (bottom) variants per gene in gnomAD (left) and primates (right). Genes are colored by their pLI scores. (C) Distributions of observed/expected ratios of synonymous (dashed lines) and missense (solid lines) variants for all genes. Results for primate genes (orange) and gnomAD genes (blue) are shown. (D) Scatter plot of missense : synonymous ratios between primate and human genes. Highlighted points are genes that are under significantly stronger (blue) or weaker (red) constraint in humans compared to non-human primates under both methods (Benjamini-Hochberg FDR < 0.05), while grey points show non-significant genes. The top 10 genes with the largest effect sizes in either direction are labeled.

Fig. 3.. PrimateAI-3D architecture and variant classification performance.

(A) PrimateAI-3D workflow. Human protein structures and multiple sequence alignments are voxelized (left) as input to a 3D convolutional neural network that predicts pathogenicity of all possible point mutations of a target residue (middle). The network is trained using a loss function with three components (right): common human and primate variants; fill-in-the-blank of a protein structure; score ranks from language models. (B) Protein structure of the STK11 gene, colored by PrimateAI-3D pathogenicity prediction scores (blue: benign; red: pathogenic). Spheres indicate residues with common human and primate variants (left) or residues with pathogenic mutations from ClinVar (right). For spheres, the color corresponds to the pathogenicity score of only the variant. For other residues, pathogenicity scores are averaged over all variants at that site. (C) Scatterplot shows performance of methods that predict missense variant pathogenicity in two clinical benchmarks (DDD and UKBB). Datasets are a subset of variants for which all methods have predictions. (D) Six barplots show method performance for six testing datasets (DMS assays, UKBB, ClinVar, DDD, ASD, and CHD).

Fig. 4.. Impact of training dataset size on classification accuracy.

(A) Improved performance of PrimateAI-3D with increasing number of common human and primate variants in the training dataset (x-axis). Performance of each dataset (y-axis) was divided by the maximum performance observed across all training dataset sizes. (B) Cumulative fractions of all possible human synonymous (grey) and missense (green) variants observed as common variants in 234 primate species, including human (allele frequency > 0.1%). Each point shows the average of ten permutations, calculated with a different random ordering of the list of primate species each time.

Fig. 5.. Enrichment of de novo mutations in the neurodevelopmental disorder cohort over expectation.

(A) Enrichment of DNMs from Kaplanis et al. (87) across all genes. Enrichment ratios are given for synonymous, all missense, and protein-truncating variants (PTV), along with missense split by PrimateAI-3D score into benign (<0.821) and pathogenic (>0.821). (B) Enrichment of benign and pathogenic missense above expectation at varying PrimateAI-3D thresholds for classifying pathogenic missense.