Quik 2.0: Efficient Large-Scale DNA Barcode Calling

Abstract

DNA barcodes are used as unique identifiers in high-throughput sequencing technologies with applications in areas such as single cell analysis, spatial transcriptomics and DNA data storage. Given a set of barcodes and a set of reads, each containing a barcode, the task of barcode calling is to assign each read to its respective barcode. This is challenging in applications involving large barcode sets and high rates of base insertion, deletion and substitution errors. Naive solutions require the calculation of pairwise distances between each barcode and read. As this is infeasible for modern applications with millions of barcodes and billions of reads, much work has been done during the previous years in accelerating this task. In 2026, Uphoff et al. introduced the barcode calling tool Quik based on k-mer filtering and pseudo-distances. They demonstrated that it is faster than state-of-the-art tools by several orders of magnitude. Here, we present Quik 2.0, which is faster than the original release by a factor of up to 56 and scales well to multiple GPUs. We discuss several algorithmic design choices that led to this speedup. In large-scale experiments with 10⁶ barcodes, we can now process approximately 300 million reads per hour on a GPU server equipped with four GPUs. In addition, we show that unfiltered barcode calling approaches can only slightly improve the accuracy at the cost of a vastly increased running time. Finally, we introduce more fine-grained assignment rejection criteria to achieve a better trade-off between precision and acceptance rate. To help users select suitable rejection parameters for real-world experiments, we propose an automatic calibration procedure that optimizes parameters for specific barcode and read sets.