A comprehensive benchmarking of adaptive sampling tools for nanopore sequencing

Adaptive sampling is an emerging technology to enrich target reads while depleting unwanted reads during real-time nanopore sequencing. The application of different algorithms has spawned various tools for the determination of read rejection. However, an evaluation in conjunction with identifying the optimal enrichment performance for a specific task has yet to be conducted. This study aimed to evaluate the performance of six widely used tools for nanopore adaptive sampling. Three distinct types of tasks were selected for testing, including the intraspecies enrichment of COSMIC genes, the interspecies enrichment of Saccharomyces cerevisiae, and the depletion of human host DNA. All the tools show increases in coverage depths of targets varying from 1.50- to 4.86-fold. The combination of Guppy for base calling and minimap2 for read alignment emerged as the optimal read classification strategy with the highest accuracy. MinKNOW, Readfish, and BOSS-RUNS using this strategy show generally excellent enrichment or depletion performance. The deep learning method utilizing raw signals demonstrates higher accuracy and quicker read ejection compared to the conventional signal-based approach, also achieving top-class performance in host depletion. Our benchmarking study conducted a thorough comparison of current tools on various adaptive sampling occasions. The nucleotide-alignment-based approach is capable of handling diverse target references with broad application. The tools employing this strategy, especially MinKNOW, could be considered as a prior option for most adaptive sampling scenarios. The deep learning technique utilizing raw signals demonstrates remarkable classification efficiency and accuracy, warranting greater emphasis and exploration in future software development endeavors.