Community-state typing asks whether gut metagenomes fall into a small number of reproducible compositional states and — the translational question — whether those states are stable enough to serve as patient strata for targeted intervention. A K = 4 consensus on 8,489 CMD (curatedmetagenomicdata) metaphlan3 samples — 5,333 HC (healthy controls) plus 3,156 IBD (inflammatory bowel disease)/other — yields four ecotypes that the digest reports as aligning with published gut-microbiome enterotype literature (Vandeputte 2017, Lloyd-Price 2019) and as cleanly separating disease from healthy 1. How much weight a chosen K can bear is the general problem treated in ecotype-clustering-validity.
The consensus is not a single clustering. Training ran two independent methods over K ∈ {2..8}: LDA (latent Dirichlet allocation, a topic decomposition) on pseudo-counts and GMM (Gaussian mixture model) on CLR (centred log-ratio) + PCA-20, with their assignments combined. Per-method fit measures (LDA held-out perplexity, GMM BIC) decrease monotonically with K, so the discriminating signal was the cross-method ARI (adjusted Rand index, the agreement between two clusterings) between LDA and GMM, which has a local maximum at K = 4 (ARI = 0.131) and a second peak at K = 7 (0.140); a parsimony rule — smallest K within 0.02 ARI of the maximum — selects K = 4. Per-sample agreement between the two methods at K = 4 is 48.9 % 2.
The four states differ in dominant taxa and in which diagnoses they collect. CD is Crohn’s disease, UC ulcerative colitis, T1D and T2D type 1 and type 2 diabetes, CDI Clostridioides difficile infection, and nonIBD a non-IBD comparison group 1.
| Ecotype | Samples | Dominant taxa | Diagnosis distribution | Source |
|---|---|---|---|---|
| E0 — Diverse commensal | 3,604 | F. prausnitzii / R. bromii / B. uniformis / Bifidobacterium balanced | 66.8 % of HC | src: [[summaries/discoveries |
| E1 — Bacteroides2 transitional | 2,601 | P. vulgatus / B. uniformis / B. dorei | 48 % CD, 58 % UC, 100 % T1D, 97 % T2D, 67 % nonIBD | src: [[summaries/discoveries |
| E2 — Prevotella copri enterotype | 920 | P. copri 28 % | 16.9 % of HC, ~0 % disease (non-Western healthy) | src: [[summaries/discoveries |
| E3 — Severe Bacteroides-expanded | 1,364 | P. vulgatus 14 % + B. fragilis 3.6 % | 50 % CD, 40 % UC, 67 % IBD acute flare, 38 % CDI | src: [[summaries/discoveries |
The stratification signal is that disease does not collapse onto one state: CD and UC patients distribute across E1 and E3 rather than concentrating in any single ecotype, which is what downstream within-ecotype analyses are meant to exploit, while healthy samples concentrate in E0 + E2 (84 % combined); the assignments are saved as the consumable artifact data/ecotype_assignments.tsv 1. The co-occurrence of the Bacteroides2 state with metabolic disease (T1D, T2D) and IBD is reported as an echo of the broader “low-diversity disease state” hypothesis, and so stands as a hypothesis rather than a demonstrated mechanism 1.
The project report behind this digest entry is the same analysis at finer resolution, not an independent replication of it: the same four states and the same sample counts appear with per-taxon abundances — E0 diverse commensal, 3,604 samples, F. prausnitzii 6.8 %, R. bromii 4.5 %, B. uniformis 4.6 %, P. vulgatus 4.4 %, 66.8 % of HC; E1 Bacteroides2 transitional, 2,601 samples, P. vulgatus 9.8 %, B. uniformis 7.2 %, Phocaeicola dorei 3.5 %, 48 % CD, 58 % UC, 100 % T1D, 97 % T2D, 67 % nonIBD; E2 prevotella copri enterotype, 920 samples, P. copri 28 %, F. prausnitzii 6 %, 16.9 % HC, ~0 % disease (non-Western healthy); E3 severe Bacteroides-expanded, 1,364 samples, P. vulgatus 14.2 %, B. fragilis 3.6 %, 50 % CD, 40 % UC, 67 % IBD acute, 38 % CDI, donor 2708 2.
The report treats ”≥ 3 reproducible ecotypes” as directional support for its hypothesis H1a, and maps E0 / E1 / E2 / E3 recognizably onto the original Bacteroides / Prevotella / Ruminococcus enterotype framework (Arumugam 2011, Costea 2018), with E1 / E3 read as the Bacteroides2 (Bact2) low-cell-count dysbiosis signature documented in CD by Vandeputte et al. 2017 2. That refines the strata from bare statistical clusters into states with a prior literature identity, but it does not establish robust reproducibility: the same report records per-sample cross-method agreement of only 48.9 % at K = 4 and real cross-study variance in the ecotype framework itself, with a leave-one-study-out ARI of 0.113 2. Resemblance to published enterotypes and a large training sample therefore weaken rather than settle the “reproducible” label — a large n does not make a 48.9 %-agreement partition stable 2.
Transfer to clinical material was tested on the UC Davis Kuehl_WGS material: all 26 samples (23 unique patients) were projected onto the K = 4 reference via a synonymy layer, with 262 unique kaiju-classified species normalized to 97 canonical species in the training feature space 2. The cohort percentages below are therefore fractions of 26 samples, not of 23 patients, and the sample set includes repeat sequencing of the same individuals 2.
The projection also exposes an asymmetry between the two consensus methods that qualifies the transfer. Projecting Kaiju data onto a MetaPhlAn3-trained embedding leaves 54 % of Kuehl feature rows outside the training feature space; LDA on pseudo-counts tolerates this by treating absence as not-detected, whereas GMM on CLR + PCA forces all 26 Kuehl samples into a single Gaussian (E3) at confidence > 0.97 — an artifact, not biology — so LDA supplies the primary projection calls and GMM is advisory 2. Only one of the two methods that defined the strata survives the cross-classifier move 2, which is the portability problem examined in cross-cohort-microbiome-portability and classifier-database-compatibility-in-taxonomic-quantification.
- E0, 7 samples (27 % of UC Davis): minimal pathobiont burden — the report judges that it possibly does not need the cocktail 2
- E1, 11 samples (42 % of UC Davis): the dominant CD ecotype, for which the full Tier-1 cocktail (escherichia-coli AIEC — adherent-invasive E. coli — plus Tier-2 enterocloster bolteae / eggerthella lenta and bile-acid monitoring) is called appropriate 2
- E3, 8 samples (31 % of UC Davis): severe Bacteroides-expanded, with E. coli and M. gnavus (mediterraneibacter gnavus) prominent, so cocktail plus ecological-cost annotation is more critical given the higher pathobiont burden 2
The distribution across the four states is non-random: χ²(3) versus uniform = 10.0, p = 0.019, computed over those 26 samples. UC Davis looks Western — no E2 Prevotella-dominant patients — with active disease dominating (73 % E1 or E3), which the report records as directional support for H1b, that patients distribute across multiple ecotypes rather than concentrating in one, validating the stratified-targeting premise 2. The per-stratum cost of acting on such a call is treated in ecological-cost-of-microbiome-target-depletion and phage-therapy-evidence-translation.
Repeat samples bear on stability, but most of them test the assay rather than the patient: 1112 → E3 at both timepoints; p1 / p1reseq → E3 both; p2 → E1 both; 1460 (calprotectin 7,280 μg/g) → E1; and patient 6967 flips E1 ↔ E3 between two samples, described as the first direct observation of intra-patient ecosystem instability and flagged as relevant to dosing-schedule implications 2. The 1112 pair is a technical replicate (Spearman ρ = 1.000 on Tier-A, taken as evidence that Kaiju is reliable across reseq replicates), and 6967 is the only multi-timepoint UC Davis patient with biological-replicate samples 2. The stable repeats are therefore technical reproducibility, not demonstrated biological stability, and the one patient sampled biologically twice changed state — which weakens any reading of ecotype as a fixed patient attribute and makes a stratum assignment a time-stamped measurement rather than a baseline label 2.
Evidence grade. The four-state partition rests on a large cross-study sample and is reported as a consensus 1, but its internal reproducibility is weak on the report’s own measures (48.9 % cross-method agreement, leave-one-study-out ARI 0.113) 2. The clinical application rests on 26 samples from 23 patients, one usable biological time course, a single p = 0.019 χ² against uniform, and two hypotheses labelled only directionally supported, so community-state typing as a patient-stratification device should be read as a working hypothesis, not an established finding 2. Whether a cheap assay could assign a patient to a stratum at the bedside is the separate question tracked in proxy-assays-for-community-state-assignment.
Open Directions
- Settle whether K = 4 states are reproducible at all before using them as strata: re-select K with resampling and report cross-method agreement and leave-one-study-out ARI per candidate K, since the chosen K = 4 rests on a cross-method ARI of 0.131 against a 0.140 peak at K = 7 and a 48.9 % per-sample agreement 2.
- Quantify intra-patient state stability with a cohort designed for it: 6967 is the only UC Davis patient with biological multi-timepoint samples, so the single observed E1 ↔ E3 flip cannot bound a per-patient switching rate or a dosing interval 2.
- Re-project the clinical cohort with a MetaPhlAn3-based profile of the same samples, which would test whether the E0/E1/E3 split is a property of the communities or of the Kaiju-to-MetaPhlAn3 feature mapping that already collapses the GMM arm onto E3 at confidence > 0.97 2.
- Join ecotype assignments from
data/ecotype_assignments.tsvto available inflammatory markers to test whether E1 versus E3 membership predicts anything clinical beyond composition — for instance whether the E1 sample with calprotectin 7,280 μg/g is typical of E1 or an outlier 1,2. - Probe the absence of E2 as a portability limit: E2 holds 16.9 % of healthy controls and ~0 % disease in the cross-study consensus, and 0 of the 26 UC Davis samples fall in it, so this cohort cannot test whether a P. copri-dominant patient would need a different cocktail at all 1,2.
- Check whether the Bacteroides2-with-metabolic-disease pattern (100 % T1D, 97 % T2D in E1) survives per-study stratification, since those diagnosis fractions could reflect single contributing cohorts rather than a cross-population association 1.