PacBio HiFi amplicons get a reproducible taxonomy pipeline
Long-read amplicon sequencing has changed what microbial profiling can resolve, but the gain in sequence length comes with a familiar problem: taxonomy is only as reliable as the reference database, the classifier, and the settings used to run them. HiFiTaxa is built to bring those moving parts under one reproducible workflow for PacBio HiFi 16S rRNA gene and fungal ITS data. It packages the analysis into a single-command Nextflow pipeline and builds versioned references from current GTDB and UNITE releases, aiming to reduce the ad hoc parameter drift that often makes two analyses impossible to compare directly.
The design choice matters because long-read profiling is no longer just about detecting a genus name. Full-length amplicons can, in principle, support species-level calls, but only when the reference sequence is present and the classifier handles ambiguity carefully. HiFiTaxa exposes that tradeoff directly by supporting both Bayesian lowest common ancestor and naive Bayes classification, while also allowing read-level profilers for 16S rRNA gene and ITS data. One workflow, several decision points, and a clear place for users to choose between sensitivity and caution.
Mock community results split sensitivity from restraint
The 16S rRNA gene mock community put the pipeline’s options under a bright light. In the 20-species mock, BLCA detected all 20 species and returned two false positives. NB found 17 species and produced no false positives, while Emu also detected all 20 species but reported eight false positives. Those are not interchangeable outcomes. A classifier that sweeps up every true species while adding extra names may be useful for exploratory profiling, but in a clinical or environmental setting where false assignments carry downstream consequences, the extra calls can distort interpretation quickly.
In this mock setting, the contrast between methods reflects a basic statistical tension in taxonomic assignment. A more permissive classifier can recover low-abundance or borderline reads, but at the cost of inflation in species calls. A more conservative method can leave a few organisms undiscovered while avoiding incorrect labels. BLCA sat between those poles in the 16S rRNA gene benchmark: complete species recovery, but not without noise.
HiFiTaxa does not try to erase that tension. It makes it visible and operational. For laboratories that need a reproducible sequence-to-name workflow, the practical value lies in being able to run the same data through distinct algorithms without reconstructing the entire computational stack from scratch each time. The output is not one universal answer; it is a controlled comparison of answers generated under the same pipeline architecture.
Holdout tests exposed the cost of missing reference sequences
The holdout analyses are where the biological realism sharpens. When 10% of the GTDB or UNITE database sequences were withheld, species-level performance changed in ways that depend on whether a taxon was represented in the reference or absent from it. In the GTDB holdout, among taxa represented in the reference, BLCA reached 80.0% species-level accuracy, compared with 48.1% for the exact-matching species step of NB. For taxa absent from the reference, NB produced fewer incorrect assignments.
That pattern is clinically familiar even outside microbiology: better resolution on known entities often comes with more overconfident guesses when the true target is missing. In taxonomy, the penalty for absence from the database is especially sharp because the classifier can only map reads onto what it has seen before. BLCA’s stronger performance on represented taxa suggests greater utility when the relevant lineages are already captured by the database. NB’s greater restraint on novel taxa makes it easier to trust when the sample may contain organisms that are underrepresented or absent from the current reference set.
The fungal ITS benchmark landed in the same general territory, though with a narrower margin. BLCA achieved slightly higher holdout species-level accuracy at 54.7% versus 52.2% for NB. NB remained more conservative for novel taxa, while EMITS detected the most mock species. The fungal results reinforce a central point: classifier choice is not a cosmetic setting. It shifts both the number of species recovered and the probability of incorrect assignment, and those effects change with the marker gene and the state of the reference database.
Reproducibility is the main product, not just taxonomy calls
HiFiTaxa is less a single algorithm than a reproducible analysis frame for long-read amplicon work. That matters because the field has been split by fragmented tooling, inconsistent reference snapshots, and analysis scripts that age quickly once the database is updated. By versioning references and wrapping the workflow in Nextflow, the pipeline gives users a way to document what was used, rerun it later, and compare outputs across studies without guessing which hidden parameter changed the result.
The inclusion of both GTDB and UNITE also broadens the pipeline’s reach across bacterial and fungal profiling. For investigators working with complex communities, the value is not only improved speed or convenience. It is the ability to choose a classifier based on the question at hand: maximum species recovery in a constrained reference set, or lower risk of false assignment when the community may include undercataloged organisms. Those are different analytical goals, and HiFiTaxa treats them as such.
A short final point is easy to miss but important. The authors built the workflow around current releases of the reference databases, which means the taxonomy is intentionally tied to a moving target. That improves currency, but it also means results remain dependent on database updates and on the extent to which future lineages are represented. For now, HiFiTaxa offers a structured way to confront that limitation rather than hide it behind a single black-box call.