Addressing dereplication crisis: Taxonomy-free reduction of massive genome collections using embeddings of protein content

Many recent microbial genome collections curate hundreds of thousands of genomes. This volume complicates many genomic analyses such as taxon assignment because the associated computational burden is substantial. However, the number of representatives of each species is highly skewed towards human pathogens and model organisms. Thus many genomes contain little additional information and could be removed. We created a frugal dereplication method that can reduce massive genome collections based on genome sequence alone, without the need for manual curation nor taxonomic information. We recently created a genome representation for bacteria and archaea called "nanotext". This method embeds each genome in a low-dimensional vector of numbers. Extending nanotext, our proposed algorithm called "thinspace" uses these vectors to group and dereplicate similar genomes. We dereplicated the Genome Taxonomy Database (GTDB) from about 150 thousand genomes to less than 22 thousand. The resulting collection increases the percent of classified reads in a metagenomic dataset by a factor of 5 compared to NCBI RefSeq and performs equal to both a larger as well as a manually curated GTDB subset. With thinspace, massive genome collections can be dereplicated on regular hardware, without affecting downstream results. It is released under a BSD-3 license (github.com/phiweger/thinspace).