Benchmarking UMI clustering tools for accurate detection of low-frequency variants from deep sequencing

Accurate detection of low-frequency variants utilizing next-generation sequencing (NGS) is of paramount importance in both biomedical research and clinical diagnosis. However, its analytical sensitivity is impeded by NGS's inherently high error rates. An efficient solution employs unique molecular identifiers (UMIs) to tag individual DNA molecules before amplification to mitigate NGS errors. Nevertheless, UMIs are prone to collisions and amplification- or sequencing-induced errors, leading to the inaccurate clustering of UMI-tagged data. Multiple UMI clustering tools have been proposed to address these challenges; however, a systematic evaluation of their impact on low-frequency variant detection remains lacking. Here, we conducted a comprehensive benchmarking of eight UMI clustering tools-AmpUMI, Calib, CD-HIT, Du Novo, Rainbow, Starcode, UMICollapse, and UMI-Tools-utilizing simulated, reference, and sample datasets to evaluate their clustering efficiency, low-frequency variant detection accuracy, and computational performance. UMI utilization and read family counts are largely consistent across tools, while data loss differs markedly among clustering algorithms. The sensitivities of all tools, except AmpUMI, are influenced by both variant allele frequencies (VAFs) and sequencing depths. Conversely, the precisions and F1 scores of most tools-excluding AmpUMI, CD-HIT, and UMICollapse-exhibit a stronger dependence on sequencing depth as VAFs decreased. Furthermore, UMI clustering tools demonstrate a substantial reduction in false-positive (FP) calls across datasets. Concerning computational efficiency, AmpUMI achieves the fastest execution, Rainbow exhibits the lowest memory consumption, and Calib performs robustly in both aspects, particularly on small datasets. Overall, Calib exhibits the most balanced performance and is recommended for UMI clustering in low-frequency variant calling. These findings may provide valuable insights for improving variant detection and advancing the development of UMI clustering algorithms.