UC Riverside

Approximately 8.7 million eukaryotic species exist on Earth, yet only a small fraction have been fully sequenced. Although sequencing costs continue to decline, de novo genome assembly remains challenging, requiring substantial expertise, computational resources, and time. From a computational standpoint, these challenges arise from the highly repetitive nature of eukaryotic genomes, short read lengths, uneven or ultra-deep sequencing coverage, sequencing errors, and chimeric reads. Despite decades of algorithmic advances, modern assemblers still struggle with highly repetitive regions that often harbor important functional and regulatory elements. The difficulty of producing complete assemblies is underscored by the 20-year effort required to resolve the final 8% of the human genome.In this dissertation, we leverage single-copy k-mers (hereafter unikmers) for accurate read anchoring, overlap detection, and read partitioning in de novo genome assembly using highly accurate PacBio HiFi reads. Unikmers are defined as k-mers whose frequencies are close to the average sequencing depth, reflecting sequences that occur only once in the genome rather than sequencing errors. First, we introduce a statistical framework to identify unikmers directly from sequencing reads without requiring a reference genome or a prior assembly. These unikmers are next applied to two separate genome assembly tasks.As our first application, we introduce RAmbler, a genome assembler specialized for resolving complex repetitive regions from PacBio HiFi reads. RAmbler detects repetitive regions, barcodes the reads using unikmers, and uses these barcoded reads to generate a consensus assembly. We evaluate RAmbler against state-of-the-art assemblers including hifiasm, LJA, HiCANU, and Verkko on over 250 synthetic HiFi datasets spanning diverse repeat structures and sequencing conditions. RAmbler outperforms competing methods on most datasets and successfully reconstructs centromeres and other complex repeats in the human genome, achieving assembly quality comparable to the manually curated telomere-to-telomere human genome without manual intervention.Finally, we address the largely unexplored problem of optimal read selection under ultra-deep sequencing conditions. We show that this problem is related to the NP-hard minimum tiling path problem and propose AWinK, a unikmer-based heuristic that selects informative read subsets. Across synthetic and real datasets, AWinK consistently produces highly contiguous and accurate assemblies, outperforming six alternative read-selection strategies.