Systematic benchmarks reveal DNA polymerase combinations for high-quality genome assembly

Abstract

Background: A fundamental challenge in biodiversity genomics is generating long-read sequencing data and high-quality genome assemblies from samples lacking gold-standard flash-frozen tissue, including ethanol-preserved specimens, minimally invasive biopsies, and small-bodied organisms. Amplification-based library preparation enables long-read sequencing of such samples, but assembly quality critically depends on the polymerase used for amplification. Results: We present a systematic benchmark of eleven long-range DNA polymerases for PacBio HiFi sequencing and genome assembly across three taxonomically diverse species: a springtail, a sea slug, and a bat. We show that polymerase choice substantially affects assembly contiguity and gene completeness, primarily caused by polymerase-intrinsic GC-dependent amplification bias and read coverage heterogeneity. Five polymerases consistently produced assemblies with comparatively high contiguity and completeness, with some enzymes outperforming the polymerase currently used in the PacBio Ampli-Fi protocol. Importantly, combining reads amplified with different polymerases frequently resulted in superior assemblies, because polymerases complement each other by covering regions where others drop out. Combining reads amplified with different polymerases, we generated haplotype-resolved assemblies of the ~1 Gb giant springtail and ~2 Gb bat genome and improved the sea slug reference genome. Using the human GIAB benchmark, we further show that optimized amplification strategies improve variant-calling accuracy, highlighting their utility in medical and clinical research. To facilitate polymerase benchmarking across additional taxa, we provide the polymeval pipeline (https://github.com/casparbein/polymeval), which automates the computational workflow. Conclusion: Our study demonstrates that optimized amplification-based HiFi sequencing generates assemblies with >1 Mb contig N50 and >90% gene completeness, providing a toolkit to produce long-read genome assemblies from challenging samples.