Approximate nearest neighbor graph provides fast and efficient embedding with applications for large-scale biological data.

Dimension reduction (DR or embedding) algorithms such as t-SNE and UMAP have many applications in big data visualization but remain slow for large datasets. Here, we further improve the UMAP-like algorithms by (i) combining several aspects of t-SNE and UMAP to create a new DR algorithm; (ii) replacing its rate-limiting step, the K-nearest neighbor graph (K-NNG), with a Hierarchical Navigable Small World (HNSW) graph; and (iii) extending the functionality to DNA/RNA sequence data by combining HNSW with locality sensitive hashing algorithms (e.g. MinHash) for distance estimations among sequences. We also provide additional features including computation of local intrinsic dimension and hubness, which can reflect structures and properties of the underlying data that strongly affect the K-NNG accuracy, and thus the quality of the resulting embeddings. Our library, called annembed, is implemented, and fully parallelized in Rust and shows competitive accuracy compared to the popular UMAP-like algorithms. Additionally, we showcase the usefulness and scalability of our library with three real-world examples: visualizing a large-scale microbial genomic database, visualizing single-cell RNA sequencing data and metagenomic contig (or population) binning. Therefore, annembed can facilitate DR for several tasks for biological data analysis where distance computation is expensive or when there are millions to billions of data points to process.