Artificial intelligence enables comprehensive genome interpretation and nomination of candidate diagnoses for rare genetic diseases.

Abstract

Clinical interpretation of genetic variants in the context of the patient's phenotype is becoming the largest component of cost and time expenditure for genome-based diagnosis of rare genetic diseases. Artificial intelligence (AI) holds promise to greatly simplify and speed interpretation by comprehensively evaluating genetic variants for pathogenicity in the context of the growing knowledge of genetic disease. We assess the diagnostic performance of GEM, a new, AI-based, clinical decision support tool, compared with expert manual interpretation. We benchmarked GEM in a retrospective cohort of 119 probands, mostly NICU infants, diagnosed with rare genetic diseases, who received whole genome sequencing (WGS) at Rady Children's Hospital. We also performed a replication study in a separate cohort of 60 cases diagnosed at five additional academic medical centers. For comparison, we also analyzed these cases with commonly used variant prioritization tools (Phevor, Exomiser, and VAAST). Included in the comparisons were WGS and whole exome sequencing (WES) as trios, duos, and singletons. Variants underpinning diagnoses spanned diverse modes of inheritance and types, including structural variants (SVs). Patient phenotypes were extracted either manually or by automated clinical natural language processing (CNLP) from clinical notes. Finally, 14 previously unsolved cases were re-analyzed. GEM ranked >90% of causal genes among the top or second candidate, using manually curated or CNLP derived phenotypes, and prioritized a median of 3 genes for review per case. Ranking of trios and duos was unchanged when analyzed as singletons. In 17 of 20 cases with diagnostic SVs, GEM identified the causal SVs as the top or second candidate irrespective of whether SV calls where provided or inferred ab initio by GEM when absent. Analysis of 14 previously unsolved cases provided novel findings in one, candidates ultimately not advanced in 3, and no new findings in 10, demonstrating the utility of GEM for reanalysis. GEM enables automated diagnostic interpretation of WES and WGS for all types of variants, including SVs, nominating a very short list of candidate genes and disorders for final review and reporting. In combination with deep phenotyping by CNLP, GEM enables substantial automation of genetic disease diagnosis, potentially decreasing the cost and speeding case review.

Causal and Candidate Gene Variants in a Large Cohort of Women with Primary Ovarian Insufficiency

“…After an initial broad search for damaging variants, we used an AI-based eCDSS tool, GEM, that employs variant impact (VAAST and VVP), patient phenotypes (Phevor), known Mendelian and pathogenic variants (OMIM, ClinVar) and ancestry to identify disease-causing genotypes 15,24,25,42,54 . Using GEM, we supplemented our data with previously analyzed WESs and have replicated genetic findings in 11 out of 12 subjects with primary amenorrhea, demonstrating the utility of the new software.…”

“…We also used GEM to identify gene variants in each subject that were most likely to be pathogenic 24 . GEM is an Electronic Clinical Decision Support System (eCDSS) framework that aggregates and adjudicates data from multiple algorithms and clinical datasets to provide rapid and accurate diagnosis of individual genomes 24 .…”

“…We also used GEM to identify gene variants in each subject that were most likely to be pathogenic 24 . GEM is an Electronic Clinical Decision Support System (eCDSS) framework that aggregates and adjudicates data from multiple algorithms and clinical datasets to provide rapid and accurate diagnosis of individual genomes 24 . GEM generates a Bayes Factor-based score that calculates the degree of support for and against a given model (a gene allele is pathogenic vs. benign) considering multiple lines of evidence from the following variant analysis tools and data sources: VVP, VAAST, Phevor, mode of inheritance for disease genes from Online Mendelian Inheritance in Man (OMIM), pathogenicity of variants (ClinVar), population specific allele frequencies (gnomAD), quality of variants and the overall genome (data) and quality of the genomic location (gnomAD) 25,26,42,44, 54–56 .…”

“…Control subjects for category-wide association using GEM 24 included 96 unrelated, unaffected subjects recruited for health in old age and 137 CEU, FIN and GBR samples from the 1000 Genomes Project (total n=233 controls) 30,31 . The majority of CEU samples are from Utah families recruited for large family size (n=47 of 61) 32 .…”

“…However, a variant-centric approach has identified novel POI candidate genes 19,23 . We used an unbiased approach and a new prioritization algorithm (GEM) 24–26 to identify damaging gene variants in known POI genes. We then used a category-wide association approach to test the hypothesis that additional candidate mutations could be found in clustered gene sets created using known genes and gene candidates from model organisms 27,28 .…”