jvecfor

jvecfor accelerates k-nearest neighbor search for single-cell analysis by providing a drop-in replacement for BiocNeighbors::findKNN that leverages portable SIMD acceleration through the Java Vector API.

Key Features:

• HNSW-DiskANN approximate search: Implements HNSW-DiskANN algorithm for approximate nearest neighbor search.
• VP-tree exact search: Implements VP-tree algorithm for exact nearest neighbor search.
• Portable SIMD acceleration: Leverages the Java Vector API for SIMD acceleration across AVX2, AVX-512, and ARM NEON hardware.
• BiocNeighbors compatibility: Returns output structurally identical to BiocNeighbors for seamless integration into existing Bioconductor single-cell workflows.
• Performance improvement: Achieves approximately 2x speedup over Annoy-based search at n >= 50K cells.
• Graph construction delegation: Provides convenience wrappers that delegate shared nearest-neighbor (SNN) and k-nearest-neighbor (KNN) graph construction to the bluster package.

Scientific Applications:

• Single-cell RNA-seq analysis: Enables fast nearest neighbor search for dimensionality reduction, clustering, and trajectory inference in large single-cell datasets.
• Bioconductor workflow integration: Facilitates scalable analysis of datasets exceeding 50,000 cells within existing Bioconductor single-cell analysis pipelines.

Methodology:

jvecfor builds on the jvector library and implements HNSW-DiskANN approximate search and VP-tree exact search, with Java Vector API-based SIMD acceleration across AVX2, AVX-512, and ARM NEON hardware.