Kaiju is a taxonomic classification method applied to sequencing reads; the cross-project digest names the specific configuration in use as “Kaiju NCBI-NR read classification”, i.e. read classification against the NCBI non-redundant protein database 1. Within this corpus it appears as the taxonomic profiler for the Kuehl_WGS (UC Davis) whole-genome-shotgun cohort of the IBD phage-targeting project, where it supplies both ecotype-projection features and pathobiont presence calls 2. Because the reference embedding it is projected onto was built from a different profiler, Kaiju is also the pivot case for classifier-database-compatibility-in-taxonomic-quantification and cross-cohort-microbiome-portability.
Aliases and identifiers
- “Kaiju NCBI-NR read classification” — the fuller form used in the cross-project digest for the configuration projected onto the reference embedding 1
- Contrasted throughout with metaphlan3, the marker-gene relative-abundance profiler used for the reference/training data 1
- No external accession or version identifier for the tool is recorded in the sources 2
Use in the Kuehl_WGS (UC Davis) cohort
All 26 Kuehl_WGS samples (23 unique patients) were projected onto the K = 4 reference via the synonymy layer, with 262 unique Kaiju-classified species normalized to 97 canonical species in the training feature space; the resulting UC Davis distribution was 2:
- E0 — diverse commensal: 7 samples (27 %)
- E1 — Bacteroides2 transitional: 11 samples (42 %)
- E2 — prevotella copri enterotype: 0 samples
- E3 — severe Bacteroides-expanded: 8 samples (31 %) 2
These calls feed gut-microbiome-ecotypes-as-patient-strata 2.
For per-patient target selection, Kuehl_WGS Kaiju Tier-A pathobiont presence was called at a threshold of ≥0.001 relative abundance, alongside ecotype, demographics (Montreal, calprotectin, medication), a Tier-A score, a mechanism profile, and phage availability 2.
Technical reliability
Kaiju repeatability was checked directly rather than assumed: patient 1112 contributed two resequencing replicates of the same biological sample, used as a technical Kaiju reliability validation and explicitly not as a longitudinal comparison 2. Those replicates gave a spearman correlation ρ = 1.000 on Tier-A, reported as evidence that Kaiju is reliable across reseq replicates and as a validation of the technical-noise floor 2. This rests on a single replicate pair, so it is best read as a case-specific rank-repeatability check for those two libraries rather than a cohort-wide bound on technical noise, and it carries no information about agreement with other classifiers 2.
Classifier mismatch with MetaPhlAn3
The structural problem is a mismatch of profilers across cohorts: a held-out cohort processed with Kaiju NCBI-NR read classification was projected onto a reference embedding trained on MetaPhlAn3 marker-gene relative abundance, and under that mismatch the two ecotype assignment methods — LDA (latent Dirichlet allocation, implemented as sklearn LatentDirichletAllocation) on pseudo-counts, and GMM (Gaussian mixture model) on CLR (centered log-ratio) transformed data followed by principal-component-analysis — behave very differently 1.
The project report states the asymmetry plainly: LDA on pseudo-counts is robust, because the out-of-training-space sparsity is handled by treating absence as not-detected, whereas GMM on CLR + PCA is fragile, the same sparsity forcing all 26 Kuehl samples into a single Gaussian (E3) at confidence > 0.97 — “an artifact, not biology”; LDA was therefore taken as the primary Kuehl projection call and GMM as advisory 2. The same asymmetry is flagged as a transferable finding relevant to any multi-classifier microbiome pipeline, LDA robust and CLR+GMM fragile 2. This refines ecotype-clustering-validity narrowly, at the single point of cross-classifier projection: the failure documented here is the projection artifact produced when Kaiju features are pushed through a CLR + PCA representation built on MetaPhlAn3 data, and nothing broader about the reference clustering is established by it 2.
The 54 % sparsity figure should be quoted with care, because the two sources attach it to incompatible denominators. The project report describes “54 % of Kuehl feature rows outside the training feature space” as what LDA absorbs by treating absence as not-detected 2. The cross-project digest instead states that Kuehl detects only 54 % of training species, and separately that 70 % of training species have no Kuehl detection 1. These descriptions are not reconciled in the sources and are not combined here.
The consequence is a graded, not binary, trust level for Kaiju-derived calls. The Kaiju vs MetaPhlAn3 classifier mismatch limits confidence in the UC Davis ecotype calls, with LDA more trustworthy than GMM in this setting 2; and because Kuehl_WGS uses Kaiju rather than MetaPhlAn3, its Tier-A presence calls have lower confidence than the curatedmetagenomicdata (CMD) analyses 2.
Resolution limits in this cohort
The Kaiju-based calls used here are species-presence calls, and the cohort has no strain-level companion assay: there is no per-patient AIEC (adherent-invasive escherichia-coli) strain-resolution diagnostic in the current cohort, and while the 8 / 23 E. coli-positive patients carry detectable E. coli by Kaiju, the 5-phage AIEC cocktail recommendation (NB13) assumes AIEC-subset prevalence per Dogan 2014 / Dubinsky 2022; per-patient AIEC genotyping (pks-island + Yersiniabactin + Enterobactin) is listed as a near-term clinical-translation prerequisite 2. This is a gap in the cohort’s assay panel as described, not a measured statement about what the classifier can or cannot distinguish 2.