Research

Structural variants (SVs) collectively alter more nucleotides in the human genome than all other variant classes combined. They play important roles in human evolution and are associated with a wide range of common and rare diseases. Yet, despite their significance, SVs remain less well-studied than other classes of genetic variation, such as common SNPs and rare SNVs and indels. This gap is particularly pronounced in large-scale genome and phenome-wide association studies, where longstanding technical and analytical challenges have hindered their routine inclusion, especially in ancestrally diverse populations.

To address these challenges, we are identifying SVs across multiple sequencing platforms and computational methods to enable high-resolution SV discovery at scale. We are currently applying this approach to the Mount Sinai Million Health Discoveries Program, a health system–embedded biobank of one million individuals with electronic health record (EHR)–linked genetic data, where we lead the SV analysis working group. To support this effort, we are systematically evaluating and building new statistical frameworks to model the full spectrum of SVs in large, ancestrally diverse cohorts and biobanks.

Understanding cis-regulation remains a central challenge in human genomics and is essential for predicting the disease relevance of noncoding variation. Most sequence-based machine learning models focus on short-range effects, typically within a few hundred kilobases of a gene. They also rely on a single linear reference genome, limiting their ability to capture three-dimensional genome organization and inter-individual genomic variation. As a result, these approaches struggle to link personal noncoding variation to changes in gene expression and disease risk. Even as some models extend to megabase-scale predictions, they do not explicitly model structural variation and are limited in their ability to predict long-range regulatory effects.

To address these gaps, we are developing models trained on large-scale genomic variant datasets and functional annotations to predict the regulatory impact of noncoding structural variants. We are taking a multi-pronged approach in which population genetic principles and patterns of human genetic variation inform model training, while in silico mutagenesis screens enable fine-grained dissection of regulatory loci that cannot be achieved using naturally occurring variation alone. We are using these models to prioritize novel noncoding SVs that contribute to developmental disorder risk and are validating their regulatory effects in patient-derived induced pluripotent stem cells (iPSCs).

Despite their clinical relevance, structural variants still lack a robust and scalable framework for classification. Predicting the clinical impact of SVs remains difficult across all variant classes and is especially challenging for variants that are complex, occur in noncoding regions, or have uncertain effects on gene expression and protein function. These challenges are further compounded in populations that are rarely sequenced and under-represented in variant databases used for classification. Although computational SV predictors can aid variant prioritization, few are clinically calibrated or routinely integrated into standardized laboratory workflows.

To address these limitations, we are systematically evaluating computational SV interpretation tools and calibrating their outputs within established clinical interpretation frameworks. As a member of the ACMG, AMP, CAP, and ClinGen Copy Number Variant (CNV)Technical Standards working group, we are drawing on these results to inform the next release of the CNV clinical interpretation guidelines. Harnessing our health system-embedded biobank, we are also curating diagnostic SVs from diverse populations to improve representation in variant databases and identifying systematic gaps in the detection of pathogenic and likely pathogenic SVs within our health system.