hvantk: an agentic multi-omics framework for variant annotation and analysis

Abstract

Large-scale genetic variant analysis involves multiple computational steps, data formats, gene and variant annotation sources and compute requirements. Some of the steps like joint calling for large cohorts are run on large compute clusters, while gene/variant classification algorithms can be run in a desktop computer. At the same time, variant annotations such as gene expression and gene constraints metrics from gnomAD are provided in different formats, by resources that frequently change their versions. Here, we present hvantk, a framework that provides a common data model for large-scale annotation and analysis of genetic variants, by extending popular data formats such as Hail MatrixTable and AnnData. hvantk agentically integrates 21 public multi-omic data providers as plugins that a developer, bioinformatician or an AI agent can author without modifying the framework's core, drives analysis workflows using agent skills, and scales from single desktop computers to high-performance computing clusters. We used hvantk for reanalysing three case-control cohorts, comprising congenital heart disease (CHD), developmental and epileptic encephalopathy (Epi25) and schizophrenia (SCHEMA). Re-ranking of gene-based burden testing results using multi-omic annotations improved the recovery of curated disease genes from 26% to 49% in CHD, 40% to 48% in Epi25 and 51% to 54% in SCHEMA, among the top decile re-ranked genes. Expression at bulk and single-cell levels were the features contributing the most over a baseline of burden, gene constraint and publication count.