The research, led by Leonardo Lazzaro, Enrico Rossignolo, and Matteo Comin, focused on Kraken 2, one of the most widely adopted k-mer-based taxonomic classifiers in metagenomics. Rather than aligning each read against every reference genome, a strategy that becomes computationally infeasible as databases swell, Kraken 2 represents each genome by a compressed set of short sequence signatures called minimizers. These minimizers are stored in a compact hash table organized along the tree of life, and each read is classified by matching its own minimizers against this structure and tracing the highest-weight path through the taxonomy. The method is fast and accurate, but its performance depends intimately on the reference database it is trained on.

That dependency became the central question of the study. Since its inaugural release in 2003, the NCBI Reference Sequence database, known as RefSeq, has grown from roughly 2,000 cataloged organisms to more than 170,000. The team systematically reconstructed Kraken 2 databases from historical RefSeq releases spanning September 2020 to December 2024, alongside databases with varying taxonomic scope, from a viral-only index containing fewer than 14,000 species to the comprehensive PlusPFP build covering archaea, bacteria, viruses, fungi, protozoa, and plants across nearly 28,000 species and roughly 195 gigabytes of data. They then tested classification performance on three datasets: a simple simulated community of 50 species, a complex simulated marine metagenome from the CAMI2 benchmark with more than 800 species, and a real human gut metagenomic sample from the Human Microbiome Project.

The first experiment varied the taxonomic breadth of the reference while holding the RefSeq release constant. Expanding the database to include domains actually present in a sample improved classification, with the Standard database, covering archaea, bacteria, viruses, and the human genome, achieving the highest F-scores across all datasets. But adding genomes from domains absent from the sample had the opposite effect. When the plant-inclusive PlusPFP database was applied to the marine dataset, the F-score dropped by roughly five percent compared to the Standard build, driven primarily by an increase in false positives. The authors suggest that plant genomes may introduce spurious matches in marine samples, illustrating a practical dilemma: the ideal reference should be restricted to the smallest set of domains that fully represents the taxa in the sample, yet the true composition of a real metagenomic sample is unknown in advance.

The second experiment examined capped databases, compact versions of the Standard and PlusPFP builds reduced to 8 or 16 gigabytes by downsampling k-mers during construction. These Lite versions can run on ordinary laptops, making them attractive for real-time, in-field sequencing applications where high-performance computing is unavailable. The trade-off, however, was clear. Across all three datasets, the full, non-capped databases consistently achieved the highest F-scores, while the 8-gigabyte versions produced the lowest. The degradation was most pronounced for the marine dataset, the most complex community tested, suggesting that complex microbial communities are especially sensitive to reductions in reference resolution. The authors caution that results obtained with capped databases should be validated by re-running the analysis with the full build whenever feasible.

The most striking findings emerged from the longitudinal analysis. As the Standard database grew from 16,674 species in September 2020 to 27,634 in December 2024, with its size nearly doubling from 47 to 84 gigabytes, classification accuracy on the complex datasets steadily declined. For the real gut metagenome, the F-score fell from 0.66 with the 2020 release to 0.47 with the 2024 release, a drop of nearly a third. Precision on the same dataset plummeted from 0.56 to 0.36, while sensitivity declined from 0.79 to 0.68. The marine dataset showed a parallel deterioration. Only the simple simulated community, with its 50 well-separated species, remained largely stable. The pattern held at the genus level as well, and extended to species detection: the number of correctly identified species in the marine sample fell from 797 with the 2020 release to 607 with the 2024 release, even as the reference itself expanded by more than 10,000 species.

The mechanism behind this decline lies in how Kraken 2 assigns its minimizers. When a minimizer is unique to a single reference genome, it is associated with that organism’s taxonomic node. But when a minimizer appears in multiple genomes, it is reassigned to their lowest common ancestor in the taxonomic tree. As RefSeq accumulates ever more closely related genomes, k-mers that were once species-specific become shared across lineages and migrate upward in the hierarchy, eroding the discriminative power of species-level classification. The researchers traced this effect directly by examining minimizer counts for five bacterial species, including Escherichia coli, Salmonella enterica, and Pseudomonas fluorescens. In every case, the number of minimizers reassigned to higher taxonomic levels was greater in the 2024 release than in 2020, and the relative proportion of species-specific minimizers within the total database shrank consistently over time. The team also suspects that Kraken 2’s default masking parameter, which treats minimizers as equivalent even when they differ at certain positions, may compound the problem as the total number of stored minimizers grows, increasing the probability that unrelated sequences are treated as matches.

Database growth also carries a computational price. Classification runtime for the gut metagenome rose from about 119 seconds with the 2020 release to roughly 161 seconds with the 2024 release, an increase of approximately 35 percent, with similar trends across the other datasets. Memory requirements for the Standard index climbed from 49.5 to 86.8 gigabytes over the same period. For laboratories without access to servers equipped with a terabyte of RAM, the practical burden of keeping pace with RefSeq is substantial, and the accuracy gains that might justify that burden are, according to these results, largely absent.

The study’s conclusions challenge a deeply held assumption in genomics: that more reference data automatically yields better science. The authors argue that the accumulation of closely related genomes introduces increasing taxonomic ambiguity, reducing the resolution of k-mer-based classifiers while simultaneously inflating computational demands. They point toward smarter strategies as the way forward, including improved k-mer selection methods, taxonomic-aware indexing, and dynamic reference database construction tailored to specific datasets. Such approaches, they suggest, could preserve classification accuracy while retaining the undeniable benefits of expanding genomic resources.

For the broader metagenomics community, the message is a call for caution and intentionality. The same sample analyzed with the same tool can yield substantially different conclusions depending on which RefSeq release was used, an effect particularly pronounced for taxa that are inherently difficult to resolve. As sequencing costs continue to fall and genome repositories expand at an unprecedented pace, researchers may need to think less about simply enlarging their references and more about curating them with the same rigor they apply to the experiments themselves. In the race to catalog life on Earth, it turns out that bigger is not always better.

Subject of Research: The impact of RefSeq reference database growth on Kraken 2 metagenomic taxonomic classification accuracy

Article Title: When bigger is not better: the impact of RefSeq growth on Kraken2 classification accuracy

Article References: Lazzaro, L., Rossignolo, E., & Comin, M. (2026). When bigger is not better: the impact of RefSeq growth on Kraken2 classification accuracy. Frontiers in High Performance Computing, 4, Article 1860848. https://doi.org/10.3389/fhpcp.2026.1860848

Image Credits: AI Generated

DOI: 10.3389/fhpcp.2026.1860848

Keywords: metagenomics, Kraken 2, RefSeq, taxonomic classification, k-mers, minimizers, reference database, microbiome, bioinformatics, high-performance computing, genome databases, false positives