Motivation. Ensembl VEP is the de facto standard for variant annotation, yet its Perl codebase and plugin architecture (CADD, gnomAD, SpliceAI via tabix) scale poorly: whole-genome annotation of a single WGS VCF takes tens of minutes to hours, and biobank-scale cohorts multiply the cost linearly. There are faster alternatives each address one piece ( bcftools/csq for consequence calling, Illumina Connected Annotations as a C# batch binary), but none combines Rust-native performance, a composable query-engine interface, pluggable cache format and full Ensemble VEP concordance.

Results. We present vepyr, a Rust/Python variant annotation engine built on Apache DataFusion and Apache Arrow, integrated with the polars-bio ecosystem. vepyr exposes annotation as DataFusion table-valued functions that can be integrated with any SQL query plan: (i) lookup_variants() for co-located variant and population-frequency lookup, and (ii) annotate_variants() for consequence prediction, HGVS, and IMPACT classification. The novel cache layer offers two interchangeable paths: a full Parquet cache consumed through any DataFusion TableProvider for scan-heavy workloads, and an optimized fjall LSM-tree store tuned for fast point lookups. Predicate and projection pushdown with columnar file format ensure only requested annotation columns are read from disk, eliminating VEP's full-bundle-per-variant overhead, and streaming execution enables out-of-core annotation of population-scale cohorts. A plugin framework for sources such as SpliceAI, AlphaMissense, and CADD is in progress, as well as multi-threaded execution. On the GIAB HG002 v4.2.1 benchmark (4.2M variants, GRCh38), preliminary results show up to 50× speedup over Ensembl VEP when run with --everything and --hgvsc flags.

Availability. Open source (Apache 2.0); https://biodatageeks.org/vepyr/

Co-authors: Marek Wiewiórka