Assigning a sample to a community state (an “ecotype”) currently requires shotgun metagenomics, and one project tested two cheaper candidate substitutes for that assignment: clinical covariates already recorded in the chart, and a small targeted qPCR (quantitative PCR of selected taxa) panel 1. The covariate route returned an explicit negative result and the qPCR route remains an unvalidated proposal 1.
Clinical covariates alone do not reproduce metagenomic calls
The reference labels under test were four reproducible IBD ecotypes derived from 8,489 cMD (curatedmetagenomicdata) metaphlan3 samples; a classifier using clinical covariates passed AUC 0.80 in cross-validation but reached only 41 % agreement with metagenomics on UC Davis, and the report’s revised reading is that metagenomics remains required 1. For what those strata are see gut-microbiome-ecotypes-as-patient-strata, for how firmly they are delimited see ecotype-clustering-validity, and for the project itself see Metagenome-Prioritized Phage Cocktails for Crohn’s Disease and IBD.
The two headline numbers are not measured on the same scale, and the difference matters. The AUC figures are macro one-versus-rest (OvR: each ecotype scored against all others pooled) AUCs on pooled cross-validation — 0.799 for the minimal classifier on {is_ibd, sex, age} at n = 8,489 and 0.810 for the extended classifier on an n = 1,675 subset, both above the project’s 0.70 threshold — whereas the 41 % is patient-level agreement (9/22) of the minimal classifier with the metagenomic call in the UC Davis cohort, against 36 % (8/22) for the extended classifier, with 12 / 22 patients disagreeing under both 1. Figure NB03_h1c_auc.png reports the per-ecotype OvR AUCs against that 0.70 threshold 1.
The source attributes the failure to cohort structure, not to an unspecified modelling accident: the dominant learned rule is “is_ibd = 1 → E1”, and in UC Davis, which is all-CD so that is_ibd is constant, the rule collapses to the marginal mode and the minimal classifier predicts E1 for 19/22 patients 1. The project’s verdict is therefore an explicit null — clinical-covariate-only ecotype assignment was judged nonviable 1 — which contradicts any reading of the cross-validated AUC on its own as evidence that covariates suffice 1.
Figure NB03_feature_importance.png compares classifier feature importance between the minimal and extended classifiers 1. Read beside the revised interpretation — that clinical covariates distinguish HC vs IBD trivially, dominated by is_ibd, but do not separate the IBD ecotypes E1 and E3 — such a comparison is diagnostic of where the signal sat rather than a route to assignment 1.
Figure NB02_ucdavis_ecotype_x_clinical.png cross-tabulates UC Davis ecotype against Montreal location and against medication class 1. These are not the covariates used by the failed minimal classifier, which was trained on {is_ibd, sex, age}, so the 41 % patient-level agreement is not evidence about them; whether Montreal location or medication class carries per-patient ecotype information is not established by the figure description alone 1.
Targeted qPCR as a candidate cheaper proxy
The remaining proxy on the table is abundance-based rather than covariate-based: a proposed targeted qPCR ecotype panel of 4–6 species, named as F. prausnitzii, P. copri (prevotella copri), P. vulgatus, B. fragilis and M. gnavus (mediterraneibacter gnavus) 1. The panel is a design proposal in the report, not a validated assay 1.
Its supporting observation is a single patient trajectory: the 14× M. gnavus expansion in patient 6967’s E3 transition suggests the hypothesis that qPCR could substitute for full metagenomics in clinical follow-up, and validation requires a prospective cohort with paired qPCR + metagenomics across timepoints 1. Because the support is one patient and the stated scope is clinical follow-up, this is marked as a hypothesis rather than a finding 1.
Tensions
Within this one project the covariate classifier’s two performance measurements point in opposite directions: macro OvR AUC 0.80 on pooled cross-validation clears the 0.70 threshold, while 41 % patient-level agreement with the metagenomic call on UC Davis does not, and the report resolves the gap by revising toward “metagenomics remains required” rather than by splitting the difference 1. The source’s own account of the gap is that OvR AUC on a pooled cohort containing a strong cohort-axis feature overstates per-patient classifier usefulness 1. Both numbers are kept here exactly as stated, and because the discrepancy is internal to a single project it is recorded as a tension rather than promoted to a conflict page 1.
Open Directions
- Run the proposed 4–6-species qPCR panel (F. prausnitzii, P. copri, P. vulgatus, B. fragilis, M. gnavus) and metagenomics on the same specimens in a prospective cohort across timepoints, reporting per-patient agreement on the same footing as the 41 % (9/22) covariate figure — this closes the gap that the panel currently rests on a single patient’s 14× M. gnavus expansion 1.
- Refit the covariate classifier with an all-CD cohort held out before training rather than scored after pooled cross-validation, to test directly whether the AUC 0.80 / 41 % gap is fully accounted for by
is_ibdbecoming constant and predictions collapsing to E1 for 19/22 patients 1. - Test whether Montreal location and medication class — cross-tabulated in
NB02_ucdavis_ecotype_x_clinical.pngbut absent from the minimal classifier’s {is_ibd,sex,age} — carry per-patient ecotype signal within an all-CD cohort, and whether the covariates ranked inNB03_feature_importance.pngretain importance onceis_ibdis dropped 1. - Ask whether the individual panel species behave as per-patient markers under the criteria used in functional-marker-validation, since a marker that tracks ecotype only in group averages would reproduce the covariate failure mode in which an E1-vs-E3 distinction is what goes missing 1.