Precise and scalable metagenomic profiling with sample-tailored minimizer libraries

Reference-based metagenomic profiling requires large genome libraries to maximize detection and minimize false positives. However, as libraries grow, classification accuracy suffers, particularly in k-mer-based tools, as the growing overlap in genomic regions among organisms results in more high-level taxonomic assignments, blunting precision. To address this, we propose sample-tailored minimizer libraries, which improve on the minimizer-lowest common ancestor classification algorithm from the widely used Kraken 2. In this method, an initial filtering step using a large library removes non-resemblance genomes, followed by a refined classification step using a dynamically built smaller minimizer library. This 2-step classification method shows significant performance improvements compared to the state of the art. We develop a new computational tool called Slacken, a distributed and highly scalable platform based on Apache Spark, to implement the 2-step classification method, which improves speed while keeping the cost per sample comparable to Kraken 2. Specifically, in the CAMI2 'strain madness' samples, the fraction of reads classified at species level increased by 3.5×, while for <i>in silico</i> samples, it increased by 2.2×. The 2-step method achieves the sensitivity of large genomic libraries and the specificity of smaller ones, unlocking the true potential of large reference libraries for metagenomic read profiling.