Integrating sequence-to-omics predictions and personal multi-omics for rare variant prioritization

Sequence-to-omics (S2O) predictions and molecular outliers provide complementary evidence to prioritize functional rare variants and facilitate rare disease diagnoses. However, current approaches do not integrate both lines of evidence. We present Watershed-DL, a framework that integrates S2O predictions of variant effects from DNA and protein sequence with multi-omic outliers to identify large-effect rare variants. We applied Watershed-DL to 1000 Genomes samples with RNA-seq (n=1,030) to prioritize rare variants underlying expression, splicing, and allele-specific expression outliers, and, using a separate peak-level model, to samples with ATAC-seq (n=174) for chromatin accessibility outliers. Compared to a model without S2O predictions, Watershed-DL increased the number of prioritized variants by 61.4%, and S2O features were weighted highly in the model across all molecular signals. Prioritized variants were over 60-fold enriched for ClinVar pathogenic classifications. Matching prioritized variants to known pathogenic variants in the same gene with a concordant molecular effect identified 331 additional candidate recessive disease alleles. Applying this framework to the GREGoR rare disease cohort with RNA-seq (n=677) prioritized previously reported causal variants across 10 solved cases and identified 451 candidate variants in unsolved cases, including a diagnostic variant in PPP1R12A and a novel candidate disease gene, ATP5PF.