INTRODUCTION
Millions of people have received genome and exome sequencing to date, a collective effort that has illuminated for the first time the vast catalog of small genetic differences that distinguish us as individuals within our species. However, the effects of most of these genetic variants remain unknown, limiting their clinical utility and actionability. New approaches that can accurately discern disease-causing from benign mutations and interpret genetic variants on a genome-wide scale would constitute a meaningful initial step towards realizing the potential of personalized genomic medicine.
RATIONALE
As a result of the short evolutionary distance between humans and nonhuman primates, our proteins share near-perfect amino acid sequence identity. Hence, the effects of a protein-altering mutation found in one species are likely to be concordant in the other species. By systematically cataloging common variants of nonhuman primates, we aimed to annotate these variants as being unlikely to cause human disease as they are tolerated by natural selection in a closely related species. Once collected, the resulting resource may be applied to infer the effects of unobserved variants across the genome using machine learning.
RESULTS
Following the strategy outlined above we obtained whole-genome sequencing data for 809 individuals from 233 primate species and cataloged 4.3 million common missense variants. We confirmed that human missense variants seen in at least one nonhuman primate species were annotated as benign in the ClinVar clinical variant database in 99% of cases. By contrast, common variants from mammals and vertebrates outside the primate lineage were substantially less likely to be benign in the ClinVar database (71 to 87% benign), restricting this strategy to nonhuman primates. Overall, we reclassified more than 4 million human missense variants of previously unknown consequence as likely benign, resulting in a greater than 50-fold increase in the number of annotated missense variants compared to existing clinical databases.
To infer the pathogenicity of the remaining missense variants in the human genome, we constructed PrimateAI-3D, a semisupervised 3D-convolutional neural network that operates on voxelized protein structures. We trained PrimateAI-3D to separate common primate variants from matched control variants in 3D space as a semisupervised learning task. We evaluated the trained PrimateAI-3D model alongside 15 other published machine learning methods on their ability to distinguish between benign and pathogenic variants in six different clinical benchmarks and demonstrated that PrimateAI-3D outperformed all other classifiers in each of the tasks.
CONCLUSION
Our study addresses one of the key challenges in the variant interpretation field, namely, the lack of sufficient labeled data to effectively train large machine learning models. By generating the most comprehensive primate sequencing dataset to date and pairing this resource with a deep learning architecture that leverages 3D protein structures, we were able to achieve meaningful improvements in variant effect prediction across multiple clinical benchmarks.