Metagenome-Prioritized Phage Cocktails for Crohn's Disease and IBD

Raw report: ibd_phage_targeting Feeds into: adaptive-versus-housekeeping-functional-differentiation, adversarial-research-quality-assurance, callability-limited-comparative-inference, carriage-prevalence-versus-within-carrier-function, classifier-database-compatibility-in-taxonomic-quantification, community-metabolic-interdependence, compositional-robustness-of-differential-abundance-calls, condition-specific-fitness, confirmatory-exploratory-ecological-association-discordance, cross-cohort-microbiome-portability, cross-tenant-data-bridging, cultivation-collection-bias-in-ecological-genomics, ecological-cost-of-microbiome-target-depletion, ecotype-clustering-validity, ecotype-environment-gene-content, evidence-triangulation-for-functional-annotation, fitness-matched-null-models, functional-marker-validation, gene-function-acquisition-depth, genome-wide-versus-locus-specific-ecological-adaptation, genomic-under-representation, gut-microbiome-ecotypes-as-patient-strata, iron-acquisition-and-pathobiont-expansion, metabolite-class-signatures-of-gut-inflammation, multi-omics-integration, occurrence-versus-catabolic-activity, ontology-and-category-schema-sensitivity, pathobiont-biosynthetic-gene-cluster-repertoires, phage-therapy-evidence-translation, phenotype-database-coverage-bias, pooled-run-pseudoreplication-and-metadata-label-noise, provenance-aware-resource-discovery, proxy-assays-for-community-state-assignment, selection-on-outcome-leakage, species-cooccurrence-modules-as-intervention-units, study-batch-confounding-of-environmental-associations, taxonomic-nomenclature-reconciliation, taxonomic-resolution-dependent-functional-inference

Overview

This report closes five project pillars by combining four microbiome ecotypes, rigor-controlled pathobiont prioritization, pathway/BGC/metabolite/strain/serology analyses, three phage-evidence layers, and UC Davis per-patient cocktail design. Its central thesis is that Crohn’s disease in joint species–metabolite space is dominated by one canonical correlation axis (r = 0.96), within which six actionable Tier-A pathobionts and two cross-corroborated mechanism narratives support state-dependent hybrid-cocktail design. The report’s own closing statement of that thesis is that CD at the gut-microbiome level is a single principal-direction phenomenon (NB07d CC1 r = 0.96) within which 6 actionable Tier-A pathobionts and 2 cross-corroborated mechanism narratives — iron plus bile-acid 7α-dehydroxylation — define a state-dependent, hybrid-cocktail design framework with concrete per-patient drafts for 14 of 23 UC Davis CD patients. ₁

Key Findings

Ecotypes and patient stratification

Four ecotypes were selected from 8,489 curatedMetagenomicData MetaPhlAn3 samples (5,333 healthy controls and 3,156 IBD/other) using consensus K = 4 from the cross-method adjusted Rand index (ARI, a chance-corrected measure of agreement between two clusterings) between LDA and GMM, a criterion chosen because per-method fit measures (perplexity, BIC) monotonically prefer larger K. The ecotypes were E0 diverse-commensal (3,604 samples; 66.8 % of healthy controls), E1 Bacteroides2-transitional (2,601; dominant in CD/UC/T1D/T2D), E2 Prevotella copri enterotype (920; non-Western healthy), and E3 severe Bacteroides-expanded (1,364; IBD flare / CDI / donor 2708). The consensus framework had 48.9 % per-sample LDA/GMM agreement; K = 4 had cross-method ARI = 0.131, while K = 7 had ARI = 0.140, and the smallest-K-within-0.02-of-maximum parsimony rule selected K = 4. ₁

The two clustering methods ran on the same 8,489 MetaPhlAn3 samples drawn from fact_taxon_abundance across the CMD_HEALTHY and CMD_IBD cohorts, with LDA fit on pseudo-counts and GMM fit on CLR plus PCA-20, over K ∈ {2..8}; LDA held-out perplexity and GMM BIC decrease monotonically with K, as expected for flexible latent-factor models, so the discriminating signal is the cross-method ARI between LDA and GMM, which has a local maximum of 0.131 at K = 4 and a second peak of 0.140 at K = 7, and the parsimony rule — smallest K within 0.02 ARI of the maximum — selects K = 4, at which per-sample method agreement is 48.9 %. ₁

Each ecotype is recorded with its sample count, its defining species at their mean per-ecotype abundances, and the share of each diagnosis group routed into it — these diagnosis percentages are the fraction of a diagnosis group assigned to that ecotype, not the diagnostic make-up within the ecotype: E0 diverse commensal, 3,604 samples, F. prausnitzii 6.8 %, R. bromii 4.5 %, B. uniformis 4.6 % and P. vulgatus 4.4 %, holding 66.8 % of healthy controls; E1 Bacteroides2 transitional, 2,601 samples, P. vulgatus 9.8 %, B. uniformis 7.2 % and Phocaeicola dorei 3.5 %, holding 48 % of CD, 58 % of UC, 100 % of T1D, 97 % of T2D and 67 % of nonIBD samples; E2 Prevotella copri enterotype, 920 samples, P. copri 28 % and F. prausnitzii 6 %, holding 16.9 % of healthy controls and approximately 0 % of disease samples (non-Western healthy); and E3 severe Bacteroides-expanded, 1,364 samples, P. vulgatus 14.2 % and B. fragilis 3.6 %, holding 50 % of CD, 40 % of UC, 67 % of IBD acute and 38 % of CDI samples, plus donor 2708. ₁

The report reads this as H1a directionally supported — at least 3 reproducible ecotypes — with E0, E1, E2 and E3 mapping recognizably onto the original Bacteroides / Prevotella / Ruminococcus enterotype framework (Arumugam 2011, Costea 2018) and with E1 and E3 reflecting the Bacteroides2 (Bact2) low-cell-count dysbiosis signature documented in CD by Vandeputte et al. 2017; both mappings are external frameworks invoked for interpretation rather than project measurements. ₁

All 26 Kuehl_WGS samples, from 23 unique UC Davis patients, were projected onto the four-ecotype embedding trained on the 8,489 curatedMetagenomicData samples via the synonymy layer, with 262 unique Kaiju-classified species normalized to 97 canonical species in the training feature space: E0 contained 7 samples (27 %), E1 11 samples (42 %), E2 0 samples, and E3 8 samples (31 %). The distribution differed from uniformity, χ²(3) = 10.0, p = 0.019. The executive summary describes the same result as all 23 patients projected, occupying E0 (27 %), E1 (42 %), and E3 (31 %) with zero in E2 — non-random and consistent with an active-disease Western cohort. A clinical-covariate classifier achieved macro one-vs-rest (OvR) AUC = 0.799 with {is_ibd, sex, age} and 0.810 with extended severity covariates — described in the executive summary as macro AUC 0.80 on pooled cross-validation — but agreement with metagenomic calls was only 41 % (9/22) and 36 % (8/22), respectively; the classifier collapses to ‘IBD → E1’ because is_ibd is the dominant feature and becomes constant on an all-CD test cohort, so metagenomics remained necessary for patient-level ecotype assignment in this all-CD cohort. The report records that OvR-AUC-versus-per-patient-agreement gap as a generalizable diagnostic for classifier-utility overstatement when a cohort-axis variable dominates the feature set. Training data in the extended classifier subset was also sparse for E0 and E2 — 3 and 11 samples in total, respectively, these being ecotype counts in that subset rather than case-control contrast counts — so the extended classifier is effectively an E1-versus-E3 binary in practice. ₁

The two clinical-covariate classifiers are specified as minimal, {is_ibd, sex, age} at n = 8,489 with macro OvR AUC = 0.799, and extended, adding {hbi_max, sccai_max, calp_max} on an n = 1,675 subset with macro AUC = 0.810; against the NB02 metagenomic calls the minimal classifier agreed on 41 % (9/22) and the extended on 36 % (8/22), with 12 of 22 patients disagreeing under both. The minimal classifier predicts E1 for 19 of 22 UC Davis patients, which the report attributes to its dominant learned rule: in the training cohort IBD samples split approximately 58 % E1, approximately 40 % E3, approximately 2 % E0 and approximately 0 % E2, so the rule “is_ibd = 1 → E1” collapses to the marginal mode on an all-CD cohort where is_ibd is constant, and the extended classifier’s training subset of 702 E1, 959 E3, 3 E0 and 11 E2 is effectively an E1-versus-E3 binary problem whose severity markers do not separate the two reliably. ₁

The UC Davis ecotype distribution is non-random, χ²(3) versus uniform = 10.0, p = 0.019, with 73 % of samples in E1 or E3 and no E2 Prevotella-dominant patients — read as a Western, active-disease cohort and as directional support for H1b, patients distributing across multiple ecotypes rather than concentrating in one. Among repeat-sampled patients, 1112 projected to E3 at both timepoints, p1/p1reseq to E3 at both and p2 to E1 at both, while patient 6967 flips between E1 and E3 between two samples, which the report calls the first direct observation of intra-patient ecosystem instability and ties to the Pillar-5 dosing-schedule question. Patient 1460 appears here at calprotectin 7,280 μg/g with an E1 call, which sits against the report’s own quiescent-class listing of patient 1460 at calprotectin 7; the report does not reconcile the two figures. ₁

Projecting the Kaiju-classified Kuehl profiles onto a MetaPhlAn3-trained embedding exposes an asymmetric robustness between the two ecotype methods: 54 % of Kuehl feature rows fall outside the training feature space, which LDA on pseudo-counts handles by treating absence as not-detected, whereas GMM on CLR plus PCA is fragile under the same sparsity and forces all 26 Kuehl samples into a single Gaussian (E3) at confidence > 0.97 — an artifact rather than biology, committed as a project-level finding — so LDA is the primary Kuehl projection call and GMM advisory. ₁

The ecotype framework showed real cross-study variance: leave-one-substudy-out (LOSO) ARI had mean 0.113 and range 0.000–0.282, while pathway-feature refitting gave ARI = 0.113 and 50.6 % overall agreement. Nevertheless, external replication in HMP_2019_ibdmdb — pulled live via curatedMetagenomicData v3.18 and explicitly not part of the cMD_IBD training set — included 1,627 samples from 130 subjects, 80.4 % of samples with projection confidence > 0.70, subject-level ecotype × diagnosis χ² = 15.61, p = 0.016, and 88.2 % sign concordance for the E1 Tier-A list (45/51 candidates). The report states in one sentence that this replication included every top-10 E1 candidate (M. gnavus, S. salivarius, S. thermophilus, E. innocuum, S. parasanguinis, E. asparagiformis, I. bartlettii, H. symbiosa, G. pamelaeae, E. ramosum) and in the next that two top-20 candidates failed to replicate — S. thermophilus on effect sign, possibly cohort-specific dairy exposure, and B. stercoris non-significant; the two statements about S. thermophilus contradict each other and the report does not reconcile them. The report’s own framing is that the ecotype framework itself has real cross-study variance (LOSO ARI 0.113) while the operational Tier-A list is externally validated. Bootstrap 80 %-subsample ARI on this data was reported as 0.13–0.17, while LOSO ARI across 8 independent sub-studies was reported as 0.00–0.28 with mean 0.113, per-substudy variation the report says bootstrap masks; it credits NB04f and NB04g with showing real cross-study and cross-feature-basis ecotype variance and NB04h with showing the operational Tier-A list replicating at 88.2 % sign concordance on HMP2 with high projection confidence. The generalization it draws is that framework stability and operational-claim replication are distinct properties that need separate tests. Per-substudy alignment varies widely: LifeLinesDeep aligns well at ARI 0.21 and 85 % agreement and HansenLBS at 65 % agreement, while AsnicarF reaches 38 % and VilaAV 17 %, and because that substudy-to-substudy variance is real the report says its ‘four reproducible ecotypes’ framing must be qualified, noting that the NB01b consensus itself has only 48.9 % cross-method agreement. The pathway-feature comparison is separately quantified as ARI 0.113 between pathway-based K = 4 LDA and the taxon-based consensus ecotype on 3,145 CMD_IBD samples, with E1 most recoverable at 65.3 % agreement and E3 weakest at 30.7 %. ₁

A systematic taxonomy synonymy layer covering 2,417 aliases mapped onto 1,848 canonical species, grounded in NCBI taxid with GTDB r214+ genus renames (Bacteroides → Phocaeicola and similar), is described as the project-wide reconciliation backbone and the tool that made both curatedMetagenomicData pooling and the Kaiju → MetaPhlAn3 projection tractable; it is committed as the artifact data/species_synonymy.tsv. ₁

Rigor-controlled target prioritization

The report retracts the original NB04 within-ecotype interpretation of the C. scindens result, the original Jaccard framing, and the original 33-species Tier-A list — the retraction of NB04 claims 1 and 2, with the superseded original within-ecotype CD-versus-HC CLR differential-abundance output kept for audit. Under a confound-free within-IBD-substudy CD-versus-nonIBD meta-analysis, C. scindens was CD-up with pooled CLR-Δ (centered log-ratio difference, a compositionally aware effect size) = +1.18, FDR (false-discovery rate) = 1e-8, and 4/4 sign concordance across sub-studies; the earlier within-ecotype null call was attributed to a compound artifact of feature leakage — clustering samples on taxon abundances and then testing the same taxa within cluster — together with cMD substudy–diagnosis nesting, the two generalizable pitfalls the report records, and under a leave-one-species-out refit C. scindens is CD-up in both E1 and E3, so the report concludes there is no paradox to resolve. The corrected E1-versus-E3 top-30 Jaccard was 0.104 against a permutation-null mean of 0.785 ± 0.054 over 200 permutations, empirical p = 0.000, which the report treats as high-confidence support for ecotype-specific targeting; the earlier ‘Jaccard = 0.14 supports H2b’ framing was retracted because that value sat near the random-overlap baseline — approximately 0.10 for a top-30 of roughly 300 filtered species — with no null distribution attached. Compositional handling mattered more broadly: raw Mann-Whitney testing mis-directs 4+ of 8 protective species, and centered-log-ratio correction recovers them. ₁

The starting observation behind that repair was a preliminary pooled Mann-Whitney differential-abundance call of Clostridium scindens as CD-enriched at log₂FC +2.67, contradicting its established role as a bile-acid-producing protective species at approximately 79 % prevalence in healthy individuals; three explanations were entertained — compositional artifact, strain heterogeneity, and ecotype mixing in the pooled analysis. ₁

On the curated battery, compositional correction recovers a depletion signal for protective species that raw Mann-Whitney misses: F. prausnitzii, A. muciniphila, R. hominis, L. eligens and A. rectalis all move from non-significant under raw testing to CD-down under CLR, and Roseburia intestinalis shows an outright sign flip from raw CD-up to CLR CD-down, reproducing the C. scindens artifact pattern on a second species. The top reported pathobionts agree across methods — M. gnavus, reported in the source under the alias R. gnavus, together with E. bolteae, E. lenta and H. hathewayi, are CD-enriched under both raw and CLR testing. ₁

C. scindens itself remained CD-enriched under both methods — raw log₂FC +5.66 and CLR Δ +1.25 — so compositional correction alone was insufficient and the strain-heterogeneity and ecotype-mixing explanations stayed live at that stage. The report’s reading is that pooled Mann-Whitney on relative abundance is systematically under-sensitive for protective-species depletion, an artifact of compositional bias plus pooled-cohort heterogeneity, and that C. scindens is not resolvable at that level; under the later confound-free within-IBD-substudy design C. scindens is CD-up, which it treats not as a paradox but as the expected behaviour of a species more prevalent in the IBD-cohort source studies than in the healthy-cohort source studies within curatedMetagenomicData. ₁

The retracted within-ecotype Tier-A list held 33 species, 18 from E1 and 15 from E3; held-out-species sensitivity gave Jaccard 0.230 for E1 and 0.064 for E3 against the > 0.5 leakage-bounded threshold, and when tested against an independent confound-free CD-versus-control contrast 14 of the 18 E1 candidates had negative within-substudy effects, making them ecotype markers rather than CD drivers. The replacement analysis ran as a staged rigor-repair pipeline — NB04b (bootstrap CIs, leakage-bound sensitivity, leave-one-out refit, Jaccard permutation null, ecotype stability), NB04c (confound-free within-IBD-substudy CD-versus-nonIBD meta plus LinDA in pure Python), NB04d (rigor-controlled stopping rule) and NB04e (within-ecotype × within-substudy CD-versus-nonIBD meta) — with the full arc recorded in FAILURE_ANALYSIS.md and the two generalizable pitfalls, curatedMetagenomicData substudy-nesting unidentifiability and feature leakage in cluster-stratified differential abundance, committed to docs/pitfalls.md. ₁

The report quantifies both pitfalls and the design that answers them. curatedMetagenomicData pools samples across approximately 51 source studies. In the ecotype-assigned slice of 8,489 samples, 45 sub-studies have ≥ 10 HC samples and 5 have ≥ 10 CD samples, but zero have both — the healthy-cohort samples come from HC-only studies (LifeLinesDeep_2016, AsnicarF_2021, YachidaS_2019 and others) and the CD samples from IBD-cohort studies (HallAB_2017, VilaAV_2018, LiJ_2014, IjazUZ_2017, NielsenHB_2014) — so a pooled CD-versus-HC linear mixed-effects (LME) model with a substudy random effect is structurally unidentifiable because the random effect perfectly predicts diagnosis, empirically verified by statsmodels.mixedlm silently failing to converge on every battery species, and the confound-free alternative is the within-IBD-substudy CD-versus-nonIBD contrast, for which 4 cMD studies carry both groups at ≥ 10 CD and ≥ 10 nonIBD: HallAB_2017 89/73, LiJ_2014 76/10, IjazUZ_2017 56/38 and NielsenHB_2014 21/248, with CLR-Δ computed within each sub-study on samples processed the same way by the same group and combined by inverse-variance weighted meta-analysis. Clustering samples by taxon abundance and then running differential abundance on the same taxa within cluster is selection-on-outcome confounding, under which within-cluster effect sizes are mechanically inflated for cluster-defining taxa; it is detectable via held-out-species sensitivity (bound: Jaccard > 0.5 = leakage bounded) or a leave-one-species-out refit, and the NB04b measurements of E1 Jaccard 0.230 and E3 Jaccard 0.064 confirmed the original within-ecotype Tier-A list was substantially leakage-driven. The NB04e analysis form is credited with simultaneously eliminating feature leakage — the clustering axis and the differential-abundance axis are disjoint, with samples partitioned by ecotype and CLR-Δ computed across sub-studies within a partition — and eliminating study confounding, since a within-substudy contrast has no study-level variation; the design fails gracefully when substudy × ecotype × diagnosis cells are too small, reporting explicitly which ecotypes are meta-viable, single-study-only, or not viable. The report further records that bootstrap-stable within-ecotype differential abundance shares the feature-leakage bias with CLR-Mann-Whitney, because CLR-MW and LinDA — LinDA implemented in pure Python for this project — operate on the same ecotype-defined subsamples and therefore share the selection-on-outcome bias; only the within-substudy CD-versus-nonIBD contrast is an independent evidence source, which is why the Tier-A gating requires within-substudy concordance. ₁

Across the four IBD sub-studies (242 CD / 369 nonIBD pooled), the report’s 14-species curated battery — whose displayed table lists fewer species rows than that label, an inconsistency it does not address — produces the canonical CD signature of pathobionts up and protective commensals down with strong sign concordance: Mediterraneibacter gnavus +5.13 (FDR reported as ~0, 4/4), Eggerthella lenta +2.30 (FDR 4e-9, 4/4), Escherichia coli +1.43 (2e-4, 3/4), Clostridium scindens +1.18 (1e-8, 4/4), Enterocloster bolteae +1.09 (3e-6, 4/4), Hungatella hathewayi +0.92 (5e-4, 3/4), Bilophila wadsworthia +0.07 (not significant, 3/4), Lachnospira eligens −1.01 (4e-3, 3/4), Roseburia intestinalis −1.14 (2e-3, 4/4), Akkermansia muciniphila −1.30 (3e-3, 3/4), Faecalibacterium prausnitzii −1.67 (9e-8, 4/4), Roseburia hominis −1.77 (4e-7, 4/4) and Coprococcus eutactus −3.09 (4e-15, 4/4); B. wadsworthia is the one battery species with no significant pooled effect despite its 3/4 sign concordance. ₁

That battery flatly contradicts NB04’s within-ecotype calls for several species: NB04 called F. prausnitzii, R. hominis and L. eligens CD-up within both E1 and E3 — the original section’s “Simpson’s paradox” — where the confound-free analysis shows them CD-down, consistent with their classical protective-commensal role, and NB04’s C. scindens “n.s.” within both ecotypes is likewise contradicted. The report attributes those within-ecotype direction reversals to a compound artifact of feature leakage plus the pooled-cohort substudy × diagnosis confound, and notes that both compositional-bias-aware methods tried on the within-ecotype subsets — CLR-Mann-Whitney and LinDA — share the bias, so an evidence count assembled from within-ecotype methods alone does not resolve it. ₁

The E1 Tier-A list contains 51 candidates from a meta-analysis across two sub-studies (HallAB_2017 and NielsenHB_2014; 82 CD / 280 nonIBD), reported as 100 % sign-concordant and led by M. gnavus (+4.85), S. salivarius (+3.26), S. thermophilus (+2.69), Erysipelatoclostridium innocuum (+2.65), S. parasanguinis (+2.44), and Enterocloster asparagiformis (+2.41) — classical CD pathobionts rather than the ecotype-marker commensals the earlier within-ecotype differential-abundance analysis produced. The E3 list of 40 candidates comes from the single study HallAB_2017 (22 CD / 31 nonIBD), is flagged as provisional pending replication, and is led by H. symbiosa (+4.64), M. gnavus (+4.46), B. coccoides (+4.22), R. faecis (+4.14), C. spiroforme (+4.11), S. salivarius (+4.04) and E. innocuum (+3.68), with F. plautii replicating at +2.26 (FDR 0.02). Blautia wexlerae does not replicate (+0.25, FDR 0.80), so the earlier “rock-solid E3 triad” that included B. wexlerae — a set that had rested on NB04c’s cohort-level within-substudy evidence rather than the E3-restricted within-substudy evidence — loses that member under the stricter E3 × HallAB_2017 test. E3-specific replication requires a second cMD-IBD study that populates E3 with both CD and nonIBD samples at ≥ 10 each, the candidate being HMP2 once PENDING_HMP2_RAW is resolved, and is blocked until HMP2 raw data are ingested. The HMP2 projection cannot supply that second study: it places only 10 of 130 subjects in E3 against 106 in E1, so HMP2 cannot provide a second-substudy E3 Tier-A meta-analysis, and the 40 E3 candidates remain derived from HallAB_2017 alone. ₁

E0 and E2 are the healthy-cohort ecotypes — E0 is 66.8 % of healthy controls and E2 is the P. copri enterotype, almost entirely non-Western healthy — and no IBD sub-study in curatedMetagenomicData populates them with nonIBD controls, so for those two ecotypes the confound-free CD-versus-nonIBD contrast is not computable. For the E0 patients, 27 % of the UC Davis cohort, the report therefore proposes drawing Tier-A targets either from cross-ecotype cohort-level evidence (the 5 engraftment-confirmed pathobionts) or from an ecotype-agnostic within-substudy CD-versus-nonIBD contrast — a conditional design proposal rather than a computed E0 target list. ₁

At report time HMP2 MetaPhlAn3 had not been ingested, recorded as PENDING_HMP2_RAW in lineage.yaml; the report projects that HMP2 reingestion would add a second IBD sub-study to the E3 stratification and so unblock Tier-A replication, expand curatedMetagenomicData plus HMP2 to about 11.5 K samples for a stability-improved ecotype refit, and enable HMP2 serology × ecotype integration — projected work rather than completed analysis. ₁

Applying the A3–A6 criteria to 71 unique rigor-controlled candidates produced six actionable Tier-A targets with total score ≥ 2.5: Hungatella hathewayi (4.0, top-scoring), Mediterraneibacter gnavus (3.8), Escherichia coli (3.6), Eggerthella lenta (3.3), Flavonifractor plautii (3.3), and Enterocloster bolteae (2.8); the rigor-controlled set also contained 5 cross-ecotype engraftment-confirmed pathobionts, and 9 Tier-B candidates in the 2.2–2.4 range (including S. salivarius with Salivaricin MIBiG matches, E. asparagiformis, V. parvula, and S. parasanguinis) remain candidates subject to phage-availability promotion. Five of six donor-2708-engraftment pathobionts passed the confound-free CD-up test at FDR < 0.10 with ≥ 66 % sign concordance: M. gnavus (+5.13, FDR given in this table as 0, 4/4 sign concordance), E. lenta (+2.30, FDR 4e-9, 4/4), E. coli (+1.43, FDR 2e-4, 3/4), E. bolteae (+1.09, FDR 3e-6, 4/4), and H. hathewayi (+0.92, FDR 5e-4, 3/4); K. oxytoca was below the prevalence filter and was not tested, so 5 of 6 engraftment pathobionts pass the confound-free CD-up test. The M. gnavus FDR is written as 0 in this engraftment table and as ~0 in the 14-species battery table recorded above; both source formulations are kept as written rather than reconciled. The same table sets those confound-free effects against the retracted within-ecotype calls, which mostly ran the other way: M. gnavus E1 CD-down −2.7 and E3 CD-up +1.6; E. lenta E1 CD-down −2.4 and E3 CD-down −0.8; E. coli E1 CD-down −1.8 and E3 CD-down −0.9; E. bolteae E1 CD-down −1.6 and E3 non-significant +0.6; and H. hathewayi E1 CD-down −1.4 and E3 non-significant −0.1. ₁

The NB05 scoring table records each actionable species with its ecotype membership and evidence basis: Flavonifractor plautii ranked fifth at 3.3 with E1 | E3_prov membership on Kumbhari strain-competition evidence plus a BGC with CD-enriched CB-ORFs, and Enterocloster bolteae ranked sixth at 2.8 on engraftment with “no BGC hit in catalog”, which that table flags as a potential blind spot rather than as demonstrated absence. That NB05 row contradicts the project’s own BGC-repertoire result recorded below, which gives E. bolteae 18 BGCs (with zero iron and zero genotoxin MIBiG matches); both statements are kept as written and the report does not reconcile them. ₁

The nine sub-threshold Tier-B candidates scoring 2.2–2.4 are named as Enterocloster asparagiformis, Streptococcus salivarius, E. citroniae, E. clostridioformis, Blautia coccoides, Veillonella atypica, S. parasanguinis, Actinomyces oris and V. parvula; each carries BGC plus A4-pass plus A3 = 1–2 signals but lacks direct engraftment or strain-adaptation evidence, and the report notes that Pillar-4 phage-targetability scoring may promote any of them on B-tier phage-availability evidence. ₁

CLR-Spearman-Louvain co-occurrence networks contained 3–7 modules per subnet, and every E1_all, E1_CD, E3_all, and E3_CD subnet placed 4–5 of the 6 actionable targets in one pathobiont module of 57–84 nodes; the report’s own synthesis row states the same result as 5 of 6 actionable Tier-A species forming a single pathobiont module per subnet, which it reads as making multi-target cocktails appropriate. The modules contained 28,730, 15,354, 30,453, and 19,909 edges, respectively, across 6, 4, 3, and 7 modules. F. plautii and E. coli showed ecotype-specific module membership, which the report treats as relevant to per-patient cocktail specificity. The result supports multi-target ecological intervention, but the originally proposed A. caccae cross-feeding interpretation was later reframed as shared-environment co-occurrence after paired metabolomics testing, and the module-anchor commensals Butyricicoccus pullicaecorum, Anaerostipes caccae, and Lactococcus lactis (the last in E3) are offered only as possible functional-driver context, not as demonstrated drivers. ₁

Those networks were built via CLR transform plus rank-based Pearson correlation (equivalent to Spearman rho), with per-edge Benjamini-Hochberg false-discovery-rate correction and edges retained only at |rho| > 0.3 and FDR < 0.05, followed by Louvain community detection (networkx.community.louvain_communities) weighted by |rho|; networkx 3.5’s built-in Louvain was judged sufficient and FastSpar / SpiecEasi installation was held back as unnecessary for the H2d question given the module structure already visible at CLR-Spearman. ₁

The four subnets rest on 2,601 samples (E1_all), 581 (E1_CD), 1,364 (E3_all) and 605 (E3_CD), with 318, 255, 296 and 252 nodes and 28,730, 15,354, 30,453 and 19,909 edges across 6, 4, 3 and 7 modules respectively. ₁

The raw mean number of actionable candidates per module is 1.38 on E1_all plus E3_all, below the ≥ 2 bar stated in the plan, but the signal is not uniformly distributed across modules: in every subnet a single module contains 4–5 of the 6 actionable Tier-A candidates, the pattern the report reports in place of the per-module mean. ₁

The module structure is weaker in E3 than in E1: the two oral Actinomyces anchors of the E3_CD module couple only weakly with module pathobionts at about ρ = 0.17–0.19, which the report reads as consistent with E3 being characterized by inflammation-driven oral-gut ectopic colonization rather than by metabolic structure, so the module-anchor cocktail-design implication is E1-specific and may not transfer to E3 patients — 31 % of the UC Davis cohort — who may need a different module-anchor design framework. ₁

Mechanism and multi-omics findings

The H3 framework produced 5 SUPPORTED, 1 PARTIALLY SUPPORTED, 2 PARTIAL, and 1 NOT SUPPORTED verdicts; the report states that the two PARTIALs (H3e and H3a-new) reflect single-cohort data-scope structural caveats rather than biological refutation. The pathway contrast was a within-IBD-substudy CD-versus-nonIBD differential-abundance analysis: class-hierarchy pathway analysis identified 52 CD-up MetaCyc pathways at FDR < 0.10 and |effect| > 0.5 against a permutation-null mean of 0.077, together with 137 pathway–pathobiont attribution pairs at |ρ_meta| > 0.4 including heme biosynthesis (PWY-5920) paired with E. coli at ρ = 0.640, which the report calls its top attribution hit although the attribution table recorded below ranks the glyoxylate cycle (GLYOXYLATE-BYPASS) first at ρ_meta = 0.797 and places heme biosynthesis from glycine at rank 19 — a ranking conflict the report does not reconcile, with iron/heme acquisition enriched among the CD-up pathways at odds ratio (OR) = 8.11 and FDR = 7.4e-6, including 15 of 52 CD-up pathways; H. hathewayi had species-level enrichment for purine/pyrimidine recycling (OR = 4.86, FDR = 0.048) and TMA/choline (OR = 9.33, FDR = 0.048). The report grounds that TMA/choline enrichment in the canonical bacterial choline-degradation pathway, in which the CutC glycyl radical enzyme breaks down choline to TMA in human gut microbes (Timsina 2025, PMID 40853000) and the host converts TMA to TMAO with established cardiovascular and renal disease implications (Cheng 2025, PMID 39709651; Jiang 2024 on the phosphatidylcholine → TMA → TMAO axis) — external mechanism rather than project measurement. On that basis it reads H. hathewayi’s CD-up phosphatidylcholine acyl editing together with its choline-degradation enrichment as making the species functionally a TMA-producer pathobiont in CD samples, contributing to the systemic-inflammation and cardiovascular-comorbidity axis documented in IBD patients — an inference from pathway-level findings, not a measurement of its TMA production. ₁

That verdict was ontology-dependent on the same data: regex matching on pathway names (plan v1.7) gave a FAIL-degenerate result on clause (b) of the three-clause H3a falsifiability test, with clauses (a) and (c) passing, while the ModelSEEDDatabase MetaCyc class hierarchy (v1.8), covering 90 % of 575 HUMAnN3 pathways, gave H3a (b) SUPPORTED with iron/heme acquisition as the dominant CD-up theme (OR = 8.1, FDR 7e-6; 15 of 52 CD-up pathways) plus H. hathewayi purine/pyrimidine recycling (OR = 4.9) and TMA/choline (OR = 9.3) — same data, opposite verdict, driven by ontology choice, a lesson the report encodes as plan norm N17. The two ontologies differ in coverage as well as verdict: the regex-on-pathway-names scheme categorized 44 of 409 pathways into 7 themes, the curator-validated MetaCyc class hierarchy from ModelSEEDDatabase categorized 514 of 575 pathways into 12 themes, and the report attributes the opposite verdicts entirely to ontology choice rather than to any change in data or pipeline. ₁

NB07a’s committed outputs record its analysis scope rather than its headline counts: data/nb07a_pathway_meta.tsv holds 409 pathways × meta statistics, data/nb07a_pathway_pathobiont_pairs.tsv holds 2,454 pathway × Tier-A-core species pairs with per-substudy plus meta ρ, and data/nb07a_h3a_verdict.json holds the formal H3a v1.7 verdict with all permutation-null statistics — the 2,454 tested pairs being a different denominator from the 137 pairs that passed |ρ_meta| > 0.4. ₁

Within that attribution layer the report states that the top 25 pathway–pathobiont pairs are all Escherichia coli pathways: glyoxylate cycle (GLYOXYLATE-BYPASS) at rank 1 with ρ_meta = 0.797, PWY-6803 phosphatidylcholine acyl editing at rank 5 with 0.725, PWY-7385 1,3-propanediol biosynthesis at rank 6 with 0.707, allantoin degradation to glyoxylate at rank 12 with 0.685, 2-methylcitrate cycle I at rank 13 with 0.683, heme biosynthesis from glycine at rank 19 with 0.640 and L-arginine degradation II (AST pathway) at rank 25 with 0.617. The AIEC mechanistic labels attached to those rows — fat utilization in the fasting or inflamed gut, TMA precursor, eut/pdu virulence, purine recycling under inflammation, propionate detox, iron metabolism and amino-acid decarboxylation — are the report’s interpretations of correlations, not measurements. ₁

The pair counts above |ρ_meta| > 0.4 are concentrated on one species: E. coli 105 pairs (76 % of total signal), H. hathewayi 16, M. gnavus 8, F. plautii 7, E. lenta 1 and E. bolteae 0. The report attributes E. coli’s dominance to three factors — a highly specialized AIEC functional repertoire well matched to MetaCyc, higher relative-abundance variance, and E. coli genome content being especially well represented in HUMAnN3 relative to obligate anaerobes such as E. lenta and E. bolteae — and reads E. bolteae’s zero pathway attribution despite actionable status as a signal that its CD-up evidence may sit at the BGC level or the Kumbhari strain-frequency level rather than the HUMAnN3-pathway level. ₁

A separately named figure, NB07a_H3a_falsifiability.png, presents the H3a falsifiability test as a 2×2 panel: a top CD-up pathways forest plot, the count permutation null, MetaCyc category enrichment, and the pathway–pathobiont attribution null. ₁

Clause (b) of the three-clause H3a test failed in its v1.7 form because the test was structurally underpowered: only 3 of the 52 CD-up pathways landed in the 7 a-priori categories, so the top-3-concentration statistic is trivially 100 % under both the observed and the random-allocation null. The report’s reading is that cMD pathway differential abundance at the unstratified MetaCyc level does not preferentially load on the classical IBD-themed categories (bile-acid 7α-dehydroxylation, mucin degradation, sulfidogenesis, TMA/TMAO, eut/pdu, polyamine, amino-acid decarboxylation) but instead captures broader bacterial-fitness-in-inflamed-gut themes — heme/iron, glyoxylate, fat metabolism, allantoin/purines — that were not in the prior-literature category set as constructed. ₁

Two non-mutually-exclusive explanations are offered for that failure: category-set choice, where broadening to alternative-electron-acceptor metabolism, iron acquisition and fat-utilization-in-inflammation categories may recover the signal; and stratified-pathway resolution, where the unstratified level may be dominated by housekeeping pathways shared across many genera so that per-species stratification is the appropriate next test. With roughly 42K stratified pathway–species combinations (PWY-XXX|g__species) available to query directly and the clause-(c) attribution signal already strong, the report treats H3a (b) as best resolved as a follow-up question rather than as a current refutation. ₁

NB07b ran that follow-up, asking whether per-species CD-up pathways at the HUMAnN3 stratified-pathway level concentrate in the same 7 a-priori categories. The verdict was NOT SUPPORTED for all 6 Tier-A core species, and the report calls that verdict structural — the 7 a-priori categories are sparse in per-species pathway repertoires — rather than biological. ₁

The per-species counts behind that verdict are H. hathewayi 16 CD-up and 17 CD-down (underpowered: only 1 of 16 CD-up pathways in the 7-category set), M. gnavus 0 and 0 (untestable: background only 4 pathways in the 7-category set, no CD-up), E. coli 2 and 18 (underpowered: 0 of 2 CD-up in-theme), E. lenta 0 and 0 (untestable: background 3 pathways), F. plautii 0 and 2 (underpowered: 0 of 0 CD-up in-theme) and E. bolteae 1 and 3 (untestable: background 3 pathways). The F. plautii entry of 0 CD-up pathways contradicts the report’s own bile-acid narrative recorded above, which counts F. plautii within-carrier CD-up pathways including F420-dependent bile-acid metabolism as one of the three lines supporting that axis; both statements are kept as written and the report does not reconcile them. ₁

Combining NB07a and NB07b, the report concludes that the 7 a-priori IBD-mechanism categories are too narrow to capture HUMAnN3 MetaCyc CD signal at either the unstratified or the species-resolved level, CD-associated pathways being dominated instead by biosynthesis and niche-shift signals; it is explicit that the refutation is real for the v1.7-stated category set but does not mean that no compositional themes exist, only that a broader, programmatically derived schema is needed. ₁

NB07b_stratified_H3a_b.png carries that re-test as a four-panel diagnostic: the per-species × category CD-up heatmap, total-versus-in-7-category bars, a top species-stratified pathways forest plot, and a per-species verdict panel. ₁

The stratified output is nonetheless informative per species. H. hathewayi had 33 passing pathways describing a coherent within-carrier biosynthesis-versus-degradation shift CD-versus-nonIBD: CD-up pentose phosphate pathway +1.18 (FDR 1.6e-5), glycolysis IV (plant cytosol) +0.94 (1.0e-3), chorismate biosynthesis I +0.86 (1.6e-6) and purine nucleobases degradation +0.89 (2.6e-4); CD-down pyrimidine deoxyribonucleosides salvage −1.51 (2.4e-7), 1,3-propanediol biosynthesis −1.27 (4.3e-6), lactose / galactose degradation −1.18 (7.0e-5) and O-antigen biosynthesis (GDP-mannose) −1.07 (3.4e-4). ₁

The report’s interpretation is that under CD H. hathewayi shifts to a biosynthetic / catabolic-stress state (pentose phosphate, anabolic amino-acid biosynthesis, purine salvage) and away from sugar utilization (lactose, galactose, propanediol, inositol), with the within-species propanediol-degradation CD-down read as consistent with niche partitioning in which E. coli may dominate that niche under CD — a possible explanation rather than a demonstrated interaction. ₁

E. coli had 20 passing pathways and shows a within-carrier CD-down per-pathway abundance, opposite to its cohort-level CD-up direction; the stratified counts above record 2 of those as CD-up and 18 as CD-down, and the CD-down members listed are CMP-legionaminate biosynthesis −0.95 (FDR 3.0e-5), L-1,2-propanediol degradation −0.81 (2.9e-4), allantoin degradation to glyoxylate −0.81 (1.8e-6), phospholipid remodeling (PE) −0.84 (7.3e-6), octane oxidation −0.72 (2.4e-5) and L-histidine degradation I −0.62 (2.8e-3). ₁

The report does not treat that as contradicting the cohort-level direction: at cohort level E. coli relative abundance is CD-up so total pathway flux scales up, while within carriers the per-cell pathway repertoire is CD-down, which it reads as E. coli in CD samples showing less metabolic versatility per cell. Two non-mutually-exclusive interpretations are offered — CD’s E. coli being an AIEC-specialized subset that has shed peripheral metabolic capabilities, or CD’s E. coli facing metabolic competition from co-abundant Klebsiella and other Enterobacteriaceae so that per-cell read-mapping share is lower across pathways — and neither is resolved here. ₁

Deeper strain-level analysis on cMD raw reads was dropped under the plan’s no-raw-reads scope, with the stated alternative a genome-content survey on kbase_genomes plus the Kumbhari fact_strain_competition table. ₁

The other four Tier-A core species — M. gnavus, E. lenta, F. plautii and E. bolteae — showed only small within-carrier shifts, at most 4 passing pathways each, so the report reads their CD signal as dominated by carriage prevalence rather than within-carrier metabolic shift and judges the pathway level the wrong resolution for them, pointing instead to BGC-level and Kumbhari strain-frequency evidence. ₁

The v1.7 clause-(b) FAIL is attributed to regex-on-pathway-names limitations: only 44 of 409 background pathways matched the 7 a-priori category patterns, and pathways such as PWY-5920 (superpathway of heme biosynthesis from glycine) were silently categorized as “0_other” because “heme” and “iron” were not in the regex set. The v1.8 re-test replaced the regex with structured MetaCyc class assignments taken from the ModelSEEDDatabase MetaCyc_Pathways.tbl provenance table and expanded to 12 IBD-relevant themes including iron/heme acquisition, fat metabolism / glyoxylate, anaerobic respiration, purine/pyrimidine recycling, and aromatic amino acid / chorismate / indole. That experience is encoded as plan norm N17: prefer ontology or class hierarchy over name-pattern regex for pathway, gene or metabolite categorization wherever feasible, with regex retained as a sensitivity check. ₁

The v1.8 cohort-level H3a (b) verdict is SUPPORTED. The test was a per-theme Fisher’s exact enrichment of CD-up × in-theme on the NB07a passing pathways across the 12 themes with BH-FDR across themes, a theme counting as supported at FDR < 0.10 and odds ratio > 1.5; iron/heme acquisition passed with a background of 32 of 409 pathways, 15 of 52 CD-up pathways (29 %) against 4.07 expected, odds ratio 8.11, p = 5.8e-7 and FDR = 7.4e-6. ₁

Seven further themes did not meet that criterion, each at FDR = 1.0: anaerobic respiration (background 50, 8 CD-up, 6.36 expected, OR 1.36, p 0.29), mucin/glycan/host (83, 11, 10.55, OR 1.06, p 0.50), aromatic AA / chorismate / indole (22, 3, 2.80, OR 1.09, p 0.55), fat metabolism / glyoxylate (111, 13, 14.11, OR 0.88, p 0.70), amino-acid decarboxylation (62, 7, 7.88, OR 0.85, p 0.71), purine/pyrimidine recycling (78, 7, 9.92, OR 0.63, p 0.91) and polyamine/urea (90, 6, 11.44, OR 0.42, p 0.99). ₁

Iron/heme acquisition is therefore the dominant CD-up biochemical theme at 8.1× over background — 29 % of CD-up pathways against 8 % expected — and its 15 CD-up members include heme biosynthesis (PWY-5920, ρ = 0.64 with E. coli in the attribution layer), heme degradation and siderophore-related pathways. The report states that this completely reverses the v1.7 clause-(b) FAIL verdict, driven entirely by category-schema choice and not by any change in the underlying data: the v1.7 regex matched “iron” / “heme” only via PWY-5920’s descriptive name and dropped it into “0_other”, whereas the v1.8 ModelSEED class hierarchy assigns PWY-5920 to HEME-SYN, Heme-b-Biosynthesis, Cofactor-Biosynthesis and Tetrapyrrole-Biosynthesis. ₁

At species level the v1.8 test supported two themes for H. hathewayi out of its 16 CD-up pathways — purine/pyrimidine recycling at OR = 4.86, FDR = 0.048 (7 of 16) and TMA/choline at OR = 9.33, FDR = 0.048 (4 of 16) — while the other Tier-A core species had insufficient species-level CD-up pathway counts for per-theme power, M. gnavus, E. lenta and F. plautii having 0 CD-up pathways (the F. plautii zero being the same count its own bile-acid narrative contradicts) and E. coli and E. bolteae 1-2. ₁

The iron reading is assembled from four named lines: the NB05 E. coli MIBiG matches to Yersiniabactin and Enterobactin, both iron siderophores, plus Colibactin from the same pks pathogenicity island; the top pathway–pathobiont attribution of heme biosynthesis to E. coli at ρ = 0.640; the v1.8 cohort finding that iron/heme is the dominant CD-up theme at OR = 8.11, FDR = 7e-6; and AIEC literature — Dalmasso 2021 on yersiniabactin, Prudent 2021 on LF82 IBC formation via yersiniabactin, Dogan 2014 on AIEC iron-pathway enrichment — all of which flag iron acquisition as a central AIEC fitness mechanism. ₁

The H. hathewayi themes are tied back to its own CD-up pathway list — pentose phosphate as a precursor for nucleotide biosynthesis, purine nucleobases degradation, and PWY-6803 phosphatidylcholine acyl editing — which the report calls exactly the building-block pathways of the two enriched themes, and it identifies H. hathewayi as a known TMA producer through CutC/D activity so that the choline-metabolism enrichment is mechanistically coherent. ₁

The methodological lesson recorded from this is that the v1.7 → v1.8 move is a major scientific reversal driven entirely by category-schema choice — “no compositional themes” (FAIL) becoming “iron/heme is the dominant theme” (SUPPORTED, OR 8.1, FDR 7e-6) on the same data and the same differential abundance — and that regex-on-pathway-names is a poor substitute for a curator-validated ontology or class hierarchy when one is available, the ModelSEEDDatabase MetaCyc class hierarchy covering 90 % of HUMAnN3 outputs and being proposed as the default for any pathway category-enrichment test. ₁

The NB07c coupling test followed from the co-occurrence result that 4–5 of the 6 actionable Tier-A pathobionts co-cluster in a single CD-specific module per ecotype, asking whether the non-pathobiont anchors of those modules show metabolic coupling with the pathobionts beside them. Its substrate was E1_CD module 0 (75 nodes), whose anchor commensals are Clostridiales bacterium 1_7_47FAA, Anaerostipes caccae — the only genuine butyrate producer among the module-anchor commensals — and Bacteroides nordii, alongside that subnet’s 5 actionable Tier-A pathobionts H. hathewayi, F. plautii, E. bolteae, E. lenta and M. gnavus; and E3_CD module 1 (57 nodes), whose anchors are Actinomyces sp. oral-taxon-181 and Actinomyces sp. HMSC035G02, both oral-cavity ectopic colonizers, plus the lactate utilizer Lactonifactor longoviformis, alongside its 4 module pathobionts E. lenta, H. hathewayi, E. coli and M. gnavus. ₁

Per anchor–pathobiont pair the method was a within-IBD-substudy Spearman ρ across CMD_IBD samples, a Fisher z-meta across the 3 robust substudies ZellerG_2014, NielsenHB_2014 and IjazUZ_2017, and sign concordance, with an added iron-context layer computing triple correlations of anchor × pathobiont × iron pathway over the 15 v1.8 iron/heme pathways to test whether the iron theme is a community-wide signature or pathobiont-specific. ₁

A. caccae showed what the report calls clean strong-positive coupling with all 5 E1_CD module pathobionts at sign concordance 1.0 across all 3 substudies, the meta values being E. bolteae +0.39 (strongest pair), H. hathewayi +0.33, M. gnavus +0.31, F. plautii +0.29 and E. lenta +0.08 (the weakest, so the strong-positive label does not hold for that pair), while B. nordii ran negative against M. gnavus (−0.21) and F. plautii (−0.20). ₁

The report read the A. caccae pattern as consistent with butyrate-producer cross-feeding embedded in the CD pathobiont module — pathobiont-released substrates such as M. gnavus glucorhamnan and mucin sugars, F. plautii bile-acid metabolites and H. hathewayi lactate feeding A. caccae’s butyrogenic fermentation — while its own iron layer ruled against an iron-cross-feeding mechanism, A. caccae’s mean correlation across the 15 iron pathways being +0.13 against E. coli’s +0.45, so the coupling was judged substrate-, sugar- or lactate-mediated rather than iron-mediated; cross-feeding versus shared-environment disambiguation was deferred to the paired metabolite test, which later rejected cross-feeding. ₁

The negative B. nordii coupling with M. gnavus and F. plautii is interpreted as niche competition for similar polysaccharide substrates, B. nordii being a generalist Bacteroides competing with the same mucin- and glycan-degrading pathobionts with CD selecting one over the other — an interpretation of the correlation sign rather than a measured competition. ₁

In E3_CD only Lactonifactor × E. lenta was called strong, at ρ_meta = +0.27; the two oral Actinomyces anchors coupled weakly at about ρ = 0.17–0.19 with M. gnavus and E. lenta, which the report reads as co-trafficked ectopic colonization under CD inflammation rather than metabolic partnership, so their module membership reflects shared inflammation-driven colonization. ₁

Mean correlation with the 15 iron pathways ranks the Tier-A core as E. coli +0.45, H. hathewayi +0.20, F. plautii +0.17, M. gnavus +0.16, E. bolteae +0.09 and E. lenta +0.03. Those 15 pathways include ENTBACSYN-PWY (enterobactin biosynthesis, E. coli-canonical), HEMESYN2-PWY (heme biosynthesis II) and 8 menaquinol-biosynthesis pathways, and E. coli’s ρ = +0.45 is computed across 2,674 CMD_IBD samples, which the report reads as the iron-pathway signal being dominantly carried by E. coli while the other Tier-A core species show weak iron coupling consistent with other CD specializations — TMA/choline for H. hathewayi, bile-acid 7α-dehydroxylation for F. plautii, glucorhamnan/mucin for M. gnavus. ₁

The report therefore narrows its own v1.8 iron-theme interpretation from “all CD pathobionts have iron specialization” to the framing that CD’s E. coli (AIEC subset) drives the iron-acquisition theme while the other Tier-A pathobionts have non-iron specializations, calling that coherent with E. coli being the only actionable Tier-A species carrying iron-siderophore MIBiG matches and with its within-carrier CD-down per-pathway pattern. ₁

The H3a-new verdict at this stage is PARTIALLY SUPPORTED: A. caccae × pathobiont coupling is clean in E1_CD with 4 strong-positive, sign-concordant pairs, while the E3_CD anchors lack metabolic-coupling structure because oral-gut co-trafficking dominates, and the cross-feeding-versus-shared-environment disambiguation was deferred to the metabolite-level test. ₁

The cocktail-design consequence drawn at this stage was a metabolic-coupling cost: a phage cocktail depleting M. gnavus, F. plautii, H. hathewayi and E. bolteae may incidentally reduce A. caccae abundance through loss of substrate, and because A. caccae is the only genuine butyrate producer among the module-anchor commensals and butyrate is anti-inflammatory, that incidental depletion could partially offset the cocktail’s therapeutic benefit, so the report called for a per-target “metabolic-coupling cost” annotation naming which beneficial commensals depend on each target as a substrate source and what the predicted net inflammatory balance is. This was a hypothesis about an untested intervention; the later paired-metabolite result reframed the coupling as shared-environment co-occurrence and the annotation as a bile-acid coupling cost. ₁

NB08a was then posed as the genomic-content level of the same question, using ref_bgc_catalog (Elmassry 2025; 10,060 BGCs across 6,221 species-annotated entries) to ask whether Tier-A pathobiont genomes carry an over-represented iron-siderophore / genotoxin biosynthetic gene-cluster signature and whether that signature is uniformly distributed across the actionable Tier-A set or concentrated on E. coli. ₁

The iron-acquisition narrative was concentrated on E. coli across five evidence streams the report describes as independent: per-actionable MIBiG lookup; pathway × pathobiont attribution, with heme biosynthesis correlated with E. coli at ρ = 0.640; cohort pathway-class enrichment at OR = 8.1; sample-level co-variation, with E. coli at mean ρ = +0.45 with 15 iron pathways against mean ρ = 0.13 for A. caccae in the same sample-level species × pathway test; and genomic BGC (biosynthetic gene cluster) content at OR = 44.4 driven by E. coli’s 54 iron BGCs. Within the actionable Tier-A core, E. coli alone carries the iron-plus-genotoxin MIBiG signature — 146 BGCs in total, of which 54 are iron and 25 genotoxin, with the iron set comprising 19 Yersiniabactin, 16 Enterobactin and 19 BGCs of class siderophore (some redundant per-strain assemblies) and the genotoxin set comprising 8 Colibactin, 15 Microcin B17 and 2 Microcin J25 — while E. lenta (41 BGCs), M. gnavus (58), E. bolteae (18), H. hathewayi (13), and F. plautii (10) each carry zero iron and zero genotoxin MIBiG matches. The report calls those five species MIBiG dark matter: they carry substantial BGC content but no MIBiG-annotated iron or genotoxin clusters, so the zeros record absent annotation rather than demonstrated absence of such clusters. The underlying BGC catalog, attributed to Elmassry 2025, holds 10,060 BGCs across 6,221 species-annotated entries, and within it E. coli alone among the six actionable Tier-A core species carries the canonical AIEC virulence biosynthetic-gene-cluster repertoire of 19 Yersiniabactin, 16 Enterobactin and 19 siderophore-class BGCs plus 8 Colibactin and 15 Microcin B17 BGCs. Of M. gnavus’s 58 BGCs, 31 are lanthipeptides. In the same place the report states that these five MIBiG-dark species’ read-level CB-ORFs are CD-versus-HC RPKM-enriched; that blanket statement contradicts its own read-level table reported below, in which E. lenta has 0 % CD-up and 54 % CD-down CB-ORFs, and the report does not reconcile the two. Its reading of the converging streams — described as five independent evidence streams across three distinct analytical granularities — is that iron is not just a host nutrient but a pathobiont-selective resource and that CD pathobionts have systematically over-invested in iron-acquisition machinery, meaning siderophores, heme uptake, and heme biosynthesis as a precursor for haem-containing iron-uptake and regulation systems; that is an interpretation of correlational and genomic-content evidence, not a measured selection experiment. ₁

The narrowing of the iron theme to E. coli began as correlation-based evidence derived from species × pathway Spearman ρ across the curatedMetagenomicData IBD samples, and the report notes that 8 of the 15 iron pathways are menaquinol biosynthesis — canonical bacterial respiratory quinones, not strictly E. coli-specific — while the strongest signature pathways, ENTBACSYN-PWY and HEMESYN2-PWY, are E. coli-canonical. The direct genomic-content evidence that E. coli alone among the Tier-A core carries iron-siderophore MIBiG matches (54 BGCs against 0 for the other 5 actionable species) is therefore what the report treats as its fine-grained attribution path, because HUMAnN3 species-stratified per-pathway attribution from gene-family files requires raw reads and was dropped under plan v1.9. ₁

Fisher’s exact tests of Tier-A core BGCs against background BGCs, with BH-FDR across four IBD-relevant themes, gave iron_siderophore 54/286 versus 51/9,774 (OR = 44.4, FDR = 6.5e-56), genotoxin_microcin 25/286 versus 4/9,774 (OR = 234.0, FDR = 3.3e-35), and NRPS_PKS_hybrid 32/286 versus 739/9,774 (OR = 1.54, FDR = 0.042) as supported, while bacteriocin_RiPP at 154/286 versus 5,217/9,774 (OR = 1.02, FDR = 0.90) sat at background rate. Within the Tier-A core, E. coli alone carries the iron-plus-genotoxin BGC signature, 5 of 6 Tier-A core species have CD-up CB-ORF rates above background, and the report states ebf/ecf RPKM is CD-up across 4 cohorts at p < 1e-31 — a blanket threshold for both families that does not hold for its own tabulated ebf meta p = 1.1e-31, an inconsistency it leaves unreconciled — which it records as replicating Elmassry 2025. The report argues that the iron-siderophore OR of 44.4 complements rather than restates the pathway-class iron OR of 8.1, because the two come from different evidence streams — pathway-cohort co-occurrence versus genomic gene-cluster content — and converge on iron acquisition as a CD pathobiont-defining genomic capability. Figure NB08a presents three panels: the Tier-A core BGC repertoire as a Grouped Class stacked bar, theme enrichment Fisher odds ratios with iron, genotoxin and NRPS-PKS supported, and ebf/ecf RPKM CD-vs-HC by cohort as boxplots. ₁

Two committed mechanism tables record the scope of these results without restating them: the per-species per-theme Fisher enrichment output notes 2 themes supported for H. hathewayi without identifying those themes in that row, and the Tier-A core BGC repertoire table holds one row for each of the 6 Tier-A core species with columns for BGC count, MIBiG compounds and grouped class. ₁

The H3c verdict is PARTIALLY SUPPORTED: localization of iron-plus-genotoxin BGCs to a minority of Tier-A pathobionts is confirmed (uniquely E. coli, 1 of 6), and the beyond-species-abundance claim is partially confirmed — read-level CB-ORF CD-versus-HC is independently CD-up enriched for 5 of 6 Tier-A core species, and ebf/ecf is independently CD-up at p < 1e-31 across 4 cohorts — but the strict species × BGC interaction-term test was not performed, because species-stratified per-sample BGC abundance was absent from the pre-computed mart slice. M. gnavus, E. lenta, E. bolteae, H. hathewayi, and F. plautii therefore remain BGC-mechanism dark, with design resting on Tier-A scoring, module membership, and within-carrier metabolic signatures rather than BGC presence, and ebf/ecf RPKM — which has no per-species attribution — is positioned as a sample-level treatment-response biomarker rather than a cocktail target-selection feature. ₁

The read-level CB-ORF test matched Tier-A core species against a reference of 5,157 CB-ORFs: E. bolteae had 11 CB-ORFs matched with 9 (82 %) CD-up at FDR < 0.10, 0 CD-down and mean effect +2.81; F. plautii 5 with 2 (40 %) CD-up, 0 CD-down and +0.77; E. coli 51 with 14 (27 %) CD-up, 0 CD-down and +1.56; M. gnavus 19 with 5 (26 %) CD-up, 0 CD-down and +1.43; H. hathewayi 7 with 1 (14 %) CD-up, 0 CD-down and +0.52; and E. lenta 13 with 0 CD-up but 7 (54 %) CD-down and mean effect −0.28, against a background catalog of 5,052 CB-ORFs at 2.5 % CD-up and mean effect −0.18. Five of six Tier-A core species therefore have CD-up rates of 14–82 %, above the 2.5 % catalog background, and E. lenta is the exception with its CB-ORFs CD-down at 54 %. The report reads the E. lenta CD-down pattern as consistent with its per-pathway abundance being mostly carriage-prevalence-driven rather than within-carrier abundance-shifted, and as aligning with a drug-metabolism rather than BGC-encoded mechanism for Eggerthella. It ties that 54 % CD-down versus 0 % CD-up pattern to Koppel et al. 2018 (PMID 29760174), which established that Eggerthella’s CD association is driven by drug-metabolism activity — cardiac glycoside reductase Cgr2 and arginine-2,3-aminomutase — rather than secondary-metabolite biosynthesis. Separately, E. coli’s per-pathway abundance is CD-down within carriers, the opposite of its cohort-level CD-up direction, across allantoin degradation, propanediol degradation, octane oxidation, histidine degradation and phospholipid remodeling, which the report groups as alternative-electron-acceptor and niche-specialization pathways and reads as consistent with CD-associated E. coli being an AIEC-specialized subset that has invested in iron and adherent-invasion machinery at the cost of generalist metabolic capabilities — an interpretation of the within-carrier depletion, not a demonstrated trade-off. ₁

The ebf/ecf cohort meta-analysis across 1,349 samples × 4 cohorts, by Mann-Whitney CD-versus-HC per cohort with Stouffer’s z-meta, gave RPKM (ebf) meta z = 11.71, meta p = 1.1e-31 and RPKM (ecf) meta z = 11.97, meta p = 5.1e-33. All four cohorts — HMP2-IBDMDB, MetaHIT, LLDEEP-NLIBD and PRISM — showed CD > HC for both ebf and ecf with cliff-deltas 0.17–0.73, which the report treats as a clean replication of a previously reported immunoactive fatty-acid-amide BGC-family CD-up finding and as the largest single effect in the project so far. The report describes the ebf and ecf families as producing immunoactive fatty acid amide signaling lipids structurally similar to host endocannabinoids, originally implicated by Elmassry et al. as a CD-defining functional signature, and summarizes the 1,349-sample HMP2-IBDMDB / MetaHIT / LLDEEP / PRISM cohort architecture as giving meta z = 11.71 (ebf) and meta z = 11.97 (ecf), among the largest effects in the project; elsewhere it compresses both families into the blanket phrase “CD-up across 4 cohorts at p < 1e-31”, a threshold that does not hold for the ebf meta p = 1.1e-31 recorded here, and it does not reconcile the two statements. ₁

The bile-acid 7α-dehydroxylation narrative identified F. plautii, E. lenta, and E. bolteae as active network members in 468 paired HMP2 samples. F. plautii correlated with cholate at ρ = −0.26 and lithocholate at ρ = +0.15; E. bolteae correlated with deoxycholate at ρ = +0.17 and lithocholate at ρ = +0.18; and E. lenta correlated with ketodeoxycholate at ρ = −0.14. The full bile-acid panel, in the source’s column order of A. caccae, B. nordii, H. hathewayi, F. plautii, E. bolteae, E. lenta, M. gnavus and E. coli (significance marks omitted here), reports tauro-α/β-muricholate (primary tauro-conjugated) at −0.03, −0.15, 0.00, −0.04, −0.04, −0.11, +0.20, +0.06; cholate (primary unconjugated) at −0.13, −0.18, −0.10, −0.26, −0.10, −0.13, +0.11, −0.02; deoxycholate (secondary, from cholate) at −0.07, +0.05, −0.12, +0.06, +0.17, +0.03, −0.10, −0.05; lithocholate (secondary, from CDCA) at −0.05, +0.06, −0.12, +0.15, +0.18, +0.07, −0.20, −0.13; and ketodeoxycholate (secondary, oxidized) at −0.12, −0.15, −0.01, −0.19, −0.05, −0.14, +0.14, +0.24. ₁

F. plautii, E. lenta, and E. bolteae — the canonical bile-acid 7α-dehydroxylating bacteria — show the predicted substrate-product signature of negative correlation with primary tauro-conjugated substrates and positive correlation with secondary unconjugated products (deoxycholate, lithocholate), which the report calls direct sample-level confirmation that the 7α-dehydroxylation network operates in HMP2 samples; the evidence is correlational, not an intervention. M. gnavus and E. coli show the opposite pattern — positive with primary tauro-bile acids and negative with secondary ones — and are not placed in the 7α-dehydroxylation network. The panel behind this reading is 468 paired HMP2 CSM* samples × 8 species × 583 metabolites by Spearman ρ with per-species FDR correction, within which E. coli dominates the cohort-level correlations (cadaverine +0.45, choline +0.25, tryptophan +0.25). The report positions this paired-correlation stream as independent of and stronger than the within-carrier F. plautii F420-dependent bile-acid pathway differential abundance — pathway abundance confirming the genomic capability, the paired correlation the active in-vivo enzymatic activity — a weighting of correlational evidence that does not by itself establish enzymatic causation. ₁

The pattern is converted into a bile-acid coupling cost for targeting these species, with F. plautii assigned the highest cost from its strong negative correlation with primary cholate (−0.26) plus positive correlation with secondary lithocholate (+0.15), on the prediction — not a measurement — that phage depletion would shift the bile-acid pool toward primary inflammatory forms. E. bolteae carries a moderate cost as a secondary 7α-dehydroxylation contributor (deoxycholate +0.17, lithocholate +0.18); E. lenta a moderate cost from a partial pattern (cholate −0.13, ketodeoxycholate −0.14) consistent with Eggerthellaceae bile-acid metabolism; and H. hathewayi, M. gnavus and E. coli a low bile-acid coupling cost, being outside the network in this dataset. Three distinct lines support the bile-acid axis: F. plautii within-carrier CD-up pathways including F420-dependent bile-acid metabolism, subject-level CD-up tauro-α/β-muricholate plus free taurine consistent with reduced microbial 7α-dehydroxylation in CD, and the paired sample-level substrate-product signature. ₁

The project’s first metabolomics analysis used precomputed HMP2 metabolomics plus curatedMetagenomicData HMP2 sample metadata rather than raw-FASTQ reprocessing: 468 of 546 metabolomics samples (86 %) matched HMP2 metagenomics samples directly by sample-ID code, covering 106 subjects (50 CD + 30 UC + 26 nonIBD), analysed one sample per subject at first-occurrence visit with UC excluded from the primary contrast. Of 592 named, HMDB-annotated metabolites out of 81,867 total peaks, 579 were testable (≥ 5 samples in each group), and 52 passed FDR < 0.10 with |cliff delta| > 0.20 — 50 CD-up and 2 CD-down. ₁

Across 11 IBD-relevant chemical-class themes, two were enriched: polyamines (5 of 9 CD-up; OR = 14.6, FDR = 0.008) and long-chain polyunsaturated fatty acids (PUFAs; 6 of 15 CD-up; OR = 7.9, FDR = 0.009). The tryptophan/indole, fatty-acid-amide, aromatic-amino-acid and oxidative themes each had 0 CD-up members on backgrounds the report marks only as small, with FDR = 1.0 and no odds ratio reported; tma_choline (2 of 29; OR = 0.78) and short_chain_fatty_acids (0 of 14) were likewise unsupported at FDR = 1.0. The CD-up polyamines were putrescine (cliff = +0.45, FDR = 0.08), N1-acetylspermine (+0.56, 0.06), N-acetylputrescine (+0.46, 0.06), anserine (+0.43, 0.08), and diacetylspermine (+0.43, 0.08); the named CD-up PUFAs were adrenate (C22:4, +0.55, FDR = 0.027), arachidonate (C20:4, +0.48, 0.05), docosapentaenoate (C22:5, +0.45, 0.06), docosahexaenoate (C22:6, +0.41, 0.08), and eicosapentaenoate (C20:5, +0.40, 0.09), spanning both n-6 and n-3 classes. Only the polyamine theme runs opposite to its pathway-level counterpart — the polyamine/urea pathway theme was CD-down at OR = 0.42 against the metabolite-level OR = 14.6 — whereas the related fat-metabolism/glyoxylate pathway theme was simply not Fisher-significant (OR = 0.88) rather than oppositely directed; the report’s readings of those gaps (pool accumulation from increased catabolism and reduced microbial clearance without elevated biosynthetic flux for polyamines; impaired host uptake, dietary fat mobilization, or reduced microbial biohydrogenation for PUFAs) are offered as candidate explanations, not as measured flux. The H3d-DA verdict is SUPPORTED: 52 metabolites pass differential abundance and 2 themes pass Fisher enrichment, both falsifiability gates are met, and the report adds polyamines and long-chain PUFAs as two mechanism axes complementing the iron, bile-acid and TMA findings. Because no per-species attribution is available, the polyamine and PUFA elevations are positioned as sample-level CD biomarkers usable for clinical follow-up such as treatment-response monitoring, not as per-species cocktail-design targets. Across the modality as a whole, 52 of 579 named HMP2 metabolites passed differential abundance at subject level in 50 CD versus 26 nonIBD subjects, and the CD-up tauro-α/β-muricholate plus free taurine corroborate the F. plautii 7α-dehydroxylation deficit at metabolite level. The report’s general statement on the polyamine inversion is that the metabolite pool reflects the difference between production and consumption rates plus dietary and host inputs, so pool and flux measurements are not interchangeable and both belong in the analysis when both are available; the polyamine metabolite class CD-up at OR = 14.6 is called the largest theme-level effect in the metabolomics analysis, against the CD-down pathway-level polyamine/urea biosynthesis theme at OR = 0.42. It also states that the CD-up PUFA signal is mechanistically connected to the fat-metabolism/glyoxylate theme, which was sub-threshold at pathway level (OR = 0.88) while theme-significant at metabolite level — a proposed connection rather than a measured one. ₁

Only 3 of 21 bile-acid-theme metabolites were CD-up (theme OR = 0.56, not significant): free taurine (cliff = +0.47) and tauro-α- and tauro-β-muricholate (+0.40). The report groups all three as tauro-conjugated forms, but free taurine is the conjugating amino acid rather than a conjugated bile acid, so that grouping is the report’s own and should not be read as three bile-acid conjugates. The report reads the pattern as consistent with reduced microbial bile-acid 7α-dehydroxylation and as corroborating F. plautii’s mechanistic role. Two of 22 acyl-carnitines were CD-up — C16 carnitine (+0.41, FDR = 0.08) and C18:1 carnitine (+0.39, 0.09) — while the acyl-carnitine theme itself was not enriched (OR = 1.06), attributed to a 22-carnitine background dominated by short- and medium-chain forms that do not shift. Nineteen lipid-class metabolites were CD-up (12 cholesteryl esters, 4 sphingomyelins, 1 ceramide, 2 TAGs), but the theme did not reach significance (OR = 1.64, FDR = 0.76), which the report attributes to dilution across a 163-metabolite background in which most lipid classes do not shift. It reads the CD-up tauro-α/β-muricholate plus free taurine as the metabolite-level confirmation of reduced microbial bile-acid 7α-dehydroxylation and as giving the actionable F. plautii phage target a metabolite-level pre-treatment biomarker. Elsewhere it describes the lipid-class signal as cholesteryl-ester C20:4, C20:5 and C18:2 elevation with “19 CEs CD-up”, corroborating cholesteryl esters as the storage form of esterified PUFAs — a count that sits against its own breakdown of 19 CD-up lipid-class metabolites of which 12 are cholesteryl esters, and the report does not reconcile the two. ₁

No short-chain fatty acid (SCFA) was differentially abundant: 0 of 14 reached FDR < 0.10, including acetate, propionate, butyrate, valerate, and hexanoate. The report lists three non-exclusive explanations — untargeted LC-MS (liquid chromatography–mass spectrometry) under-samples volatile polar SCFAs that typically require GC-MS, the subject-averaged cohort contrast masks within-subject variation, or gross SCFA pool differences are smaller than within-subject variability — and treats the null as decisive only in the negative: the A. caccae × pathobiont butyrate cross-feeding hypothesis cannot be confirmed from cohort-level butyrate differential abundance. The report draws the metabolic-coupling-cost consequence that depleting A. caccae may not produce a measurable cohort-level butyrate change, while the within-sample paired test remains the right test for the cross-feeding causal claim. ₁

Urobilin was CD-down (cliff = −0.38, FDR = 0.09), which the report interprets as reduced gut-bacterial bilirubin reduction consistent with loss of urobilinoid-producing commensals; the urobilin/porphyrin theme reached OR = 143 but FDR = 0.13 on a single CD-down member. Two mechanism axes were not resolvable in this modality: the iron/heme axis was not measured at all, because LC-MS under-samples siderophores, leaving metabolomics neutral on an axis corroborated instead by BGC content; and TMA/choline was only partially corroborated, with TMAO not theme-significant (tma_choline OR = 0.78) and support coming from the C16 and C18:1 carnitine elevations. Figure NB09a presents the HMP2 metabolomics CD-versus-nonIBD three-panel summary: a volcano of cliff_delta against −log10 FDR with key metabolites annotated, an 11-theme Fisher odds-ratio bar on which polyamines and PUFAs are supported, and the top 15 CD-up plus top 15 CD-down metabolites. ₁

Cross-cohort HMP2-to-FRANZOSA bridging produced 9 strict replications and theme-level sign concordance of 100 % for urobilin/porphyrin, 80 % for acyl-carnitines, and 75 % for long-chain PUFAs; polyamines had no m/z bridge. The replication cohort, FRANZOSA_2019, is an independent multi-site IBD metabolomics cohort of 220 participants × 88 CD + 76 UC + 56 Control, tested from precomputed metabolomics only under the no-raw-reads scope; a mouse ileal AKR/J tissue metabolomics dataset was dropped as not human IBD. The bridge itself, matching named HMP2 metabolites to Franzosa peaks within the same method at ±0.005 Da, mapped 122 of 592 named HMP2 metabolites to Franzosa peaks with 118 testable, and the three themes replicating at ≥ 75 % sign concordance were urobilin/porphyrin (100 %, n = 3, median Franzosa cliff −0.747, the strongest replicated effect), acyl-carnitines (80 %, n = 5, +0.531) and long-chain PUFAs (75 %, n = 12, +0.273). ₁

The bridge was built on mass alone: Franzosa metabolite IDs follow a [m/z]_[RT] format with no compound-name annotations in the mart, so HMP2 named metabolites (carrying HMDB ID, m/z and RT) were matched to Franzosa peaks within method by m/z at ±0.005 Da tolerance, with retention time deliberately excluded as a matching constraint because the two runs used different gradients and columns. 122 of 592 named HMP2 metabolites matched a Franzosa peak, and 118 of those 122 passed the Franzosa CD-versus-Control differential-abundance filters. ₁

Per theme, the matched count, CD-up counts in each cohort, sign-concordant count and median Franzosa cliff were: urobilin/porphyrin 3 matched, 0 CD-up in HMP2 and 0 in Franzosa, 3 concordant (100 %), median cliff −0.747 (CD-down); acyl-carnitines 5, 4, 5, 4 (80 %), +0.531; long-chain PUFA 12, 12, 9, 9 (75 %), +0.273; aromatic AA metabolites 3, 3, 2, 2 (67 %), +0.181; primary bile acids 7, 5, 4, 4 (57 %), +0.252; lipid classes 21, 14, 4, 11 (52 %), −0.069; tryptophan/indole 2, 1, 0, 1 (50 %), −0.199; secondary bile acids 5, 3, 2, 2 (40 %), −0.159; and polyamines 0 matched, with no m/z match at all, which the report marks a methodological gap. ₁

Across the 118 filtered matches, overall sign concordance was 64 % (76 of 118), three themes reached ≥ 75 %, and 9 strict cross-cohort replications passed the conjunction of FDR < 0.10, |cliff| > 0.20 and matching sign in both cohorts — which the report calls the strongest cross-cohort signal in its metabolomics analyses. ₁

Figure NB09b presents the cross-cohort cliff-δ scatter with themes colour-coded and key metabolites annotated, together with the per-theme sign-concordance bar chart on which urobilin, acyl-carnitines and long-chain PUFA are the themes at ≥ 75 % concordance. ₁

Urobilin CD-down replicated in 3 of 3 matches (100 %) at a median Franzosa cliff of −0.747 — the largest cross-cohort effect size in the panel and, in the report’s words, the strongest replicated metabolite-level finding in the project; the reading that it reflects loss of bilirubin-reducing commensals is a literature-anchored interpretation of a correlational contrast, not a measured activity. ₁

CD-up acyl-carnitines replicated in 4 of 5 matches (80 %) at a median Franzosa cliff of +0.531, with the C16 and C18:1 long-chain forms flagged CD-up in HMP2 independently elevated in Franzosa CD samples; the link to impaired β-oxidation in inflamed mucosa is mechanistic interpretation rather than measured flux. ₁

CD-up long-chain PUFAs replicated in 9 of 12 matches (75 %) at a median Franzosa cliff of +0.273, with adrenate, arachidonate, DHA and DPA replicating cleanly and eicosapentaenoate (EPA) showing the opposite sign in 1 of the 3 non-concordant cases; cohort-specific dietary fish-oil supplementation is offered only as a possible reason. ₁

Bile-acid signals were weaker cross-cohort than within-cohort: primary bile acids replicated at 57 % (4/7) and secondary bile acids at 40 % (2/5). The report attributes this to bile-acid pool sizes being more susceptible to dietary and antibiotic confounding than within-cohort species × metabolite correlation, and explicitly does not treat it as invalidating the paired sample-level substrate-product signature. ₁

The polyamine bridge failed outright: m/z-only matching recovered none of the 9 polyamines that drove the OR = 14.6 theme enrichment, because polyamines are low-mass HILIC-positive (hydrophilic-interaction chromatography, positive mode) compounds — putrescine 89.11, spermidine 146.17, spermine 203.22 — whose m/z tolerance plus Franzosa method-library overlap yields no matches. The HMP2 polyamine OR = 14.6 therefore stands as a single-cohort finding, a methodological gap rather than a failed biological replication, until external Franzosa supplementary tables with compound-name annotations are used, which sits outside the project’s no-raw-reads scope. Secondary bile acids, by contrast, did bridge but replicated at only 40 %, which the report reads as cohort-aggregate differential abundance being the weaker stream relative to the paired sample-level evidence. ₁

The report’s verdict is strong cross-cohort replication on 3 of 8 testable themes (urobilin, acyl-carnitines, long-chain PUFAs): 64 % overall sign concordance is moderately above the 50 % chance baseline and the 9 strict replications are firmly above chance, while polyamines could not be bridged and the bile-acid differential-abundance signal is partly cohort-specific. ₁

The bile-acid network finding consequently rests on two distinguishable evidence streams: the paired sample-level species × metabolite correlations within HMP2 (F. plautii × cholate ρ = −0.26, × lithocholate ρ = +0.15), which are not a CD-versus-Control aggregate contrast, and the cohort-aggregate cross-cohort differential abundance, which replicates at 57 % for primary and 40 % for secondary bile acids. ₁

The report anchors its six-line bile-acid narrative on the paired sample-level evidence plus literature mechanism rather than on cross-cohort aggregate replication, judging the within-HMP2 paired correlations the more robust stream and the cohort-aggregate metabolite differential abundance to be dominated by batch effects. It records as a methodological note that cohort-aggregate cross-cohort differential abundance and within-cohort paired-sample correlation are different evidence streams, so one can be weaker without invalidating the other. ₁

An earlier step had left a key disambiguation deferred, recorded as a PARTIAL H3a-new verdict on a coupling the report called clean in E1_CD in the same analysis where E. coli dominated iron-pathway co-variation: the A. caccae × pathobiont species-level coupling in E1_CD at +0.39 (E. bolteae), +0.33 (H. hathewayi), +0.31 (M. gnavus) and +0.29 (F. plautii) was consistent with either butyrogenic cross-feeding or shared-environment co-response to the same CD niche. The paired analysis that addressed it used 468 paired HMP2 samples carrying both metabolomics and curatedMetagenomicData MetaPhlAn3 metagenomics, under the project’s no-raw-reads scope, and rejected cross-feeding across its 8 species × 583 named metabolites: only 7 strict triangles appeared, and lactate × A. caccae ran opposite in sign to lactate × pathobionts, so the coupling was reframed as shared-environment co-occurrence in healthy/normobiotic samples. Before that reframing the report had read the pattern as pathobionts co-clustering in single CD-specific modules with the non-pathobiont module anchor A. caccae — the only genuine butyrate producer among the anchors — strongly positively coupled to the pathobiont set, consistent with cross-feeding and mapping onto known gut circuits that pair A. caccae with M. gnavus as a mucin-glucorhamnan producer (Henke 2019), F. plautii as bile-acid 7α-dehydroxylating, H. hathewayi as a lactate/TMA producer and E. bolteae as a sugar fermenter; that circuit mapping was an interpretation of correlation, never a demonstrated exchange. ₁

Under strict criteria — anchor and pathobiont same sign, both |ρ| > 0.20, both FDR < 0.10 — only 7 cross-feeding triangles appeared across 8 species × 583 named HMP2 metabolites, with top candidates caffeine (B. nordii / F. plautii), linoleoylethanolamide (B. nordii / E. lenta; A. caccae / E. lenta), urobilin (A. caccae / E. lenta; B. nordii / E. lenta) and cholate (A. caccae / F. plautii; B. nordii / F. plautii). None of the 7 candidates is a butyrogenic cross-feeding intermediate; the report reads the pattern as shared-niche health-direction co-occurrence among mostly commensal species that together correlate negatively with primary cholate and positively with the CD-down metabolite urobilin, i.e. co-occurrence in healthy rather than CD-dysbiotic samples. ₁

The curated species × metabolite panel, in the source’s column order of A. caccae, B. nordii, H. hathewayi, F. plautii, E. bolteae, E. lenta, M. gnavus and E. coli (significance marks omitted here), reports butyrate at +0.10, +0.13, −0.08, −0.02, −0.04, −0.05, −0.02, +0.04; lactate at +0.18, +0.10, +0.03, −0.23, −0.20, −0.07, +0.10, +0.24; choline at +0.04, +0.01, −0.07, −0.16, −0.15, −0.13, +0.11, +0.25; carnitine at −0.04, −0.02, −0.01, −0.18, −0.06, −0.04, +0.04, +0.19; putrescine at −0.04, −0.09, −0.07, −0.02, −0.15, −0.11, +0.13, +0.19; tryptophan at −0.09, −0.10, −0.10, −0.13, −0.10, −0.22, +0.14, +0.25; and cadaverine, marked an E. coli specialty, only for E. coli at +0.45. ₁

Butyrate × A. caccae at +0.10 is significant but very weak — consistent with A. caccae as a population-level butyrate producer, but weak partly for methodological reasons, since this analysis describes the HMP2 LC-MS run as undersampling SCFAs with only butyrate and propionate detected, a statement left unreconciled against the report’s own cohort contrast in which 14 SCFAs including acetate, valerate and hexanoate were tested for differential abundance; the detection-versus-testing-scope discrepancy is the report’s and it does not address it. The weakness is also attributed to the gut butyrate pool being dominated by other producers such as Faecalibacterium and Roseburia. The opposite-sign lactate pattern — A. caccae +0.18 against F. plautii −0.23 and E. bolteae −0.20 — is the strongest evidence against the cross-feeding hypothesis, placing A. caccae and the coupling-candidate pathobionts in different cohort-level metabolic niches; only 7 strict cross-feeding triangles emerged in the same paired test, so the report’s preferred reading of the species-level coupling is shared-environment co-occurrence in healthy/normobiotic samples rather than cross-feeding. E. coli dominates the cohort-level correlation signal with choline +0.25, carnitine +0.19, tryptophan +0.25, putrescine +0.19 and cadaverine +0.45, the strongest correlation in the entire panel. A. caccae itself is literature-characterized as a butyrate producer that grows on lactate and acetate co-substrates and on host-mucin-derived sugars, and as a primary lactate utilizer in the colon (Schwiertz 2002, Duncan 2004) — external characterization rather than a project measurement. The report also records that the coupling is not iron-mediated, A. caccae’s correlation with iron pathways being far below E. coli’s, consistent with A. caccae being a fermentative butyrate producer rather than an iron-respiring organism. ₁

The earlier H. hathewayi TMA/choline finding from cMD pathway-level analysis does not strongly replicate at HMP2 sample level: H. hathewayi × choline was ρ = −0.07 and non-significant while E. coli dominated, which narrows the signal to a combined Enterobacteriaceae-plus-Lachnospiraceae one rather than a specifically H. hathewayi one. The report offers as possible reasons H. hathewayi’s lower prevalence in HMP2 (25 % of paired samples) than in the cMD studies that drove the earlier finding, and E. coli’s high prevalence (50 %) and abundance variance dominating the cohort-correlation signal. ₁

The cross-feeding hypothesis is therefore not supported by sample-level paired evidence, and shared environment is the more parsimonious explanation for the A. caccae × pathobiont coupling. This reframes the earlier cocktail-design note: the inference that depleting pathobionts would incidentally reduce A. caccae is less load-bearing than originally framed, because the substrate-product relationship is not detectable at sample level, so phage targeting of pathobionts is unlikely to substantially deplete A. caccae through metabolic coupling — an inference, not a tested depletion. The therapeutic implication originally drawn from the coupling — that depleting the pathobiont cluster may incidentally remove an anti-inflammatory butyrate producer through loss of substrate — was the ecological-cost consideration behind the rubric’s Tier C ecological-durability criterion, and was intended as a per-target ecological-cost annotation sitting alongside the A3–A6 Tier-A scoring; it remains hypothetical rather than measured. Figure NB09c presents the curated 7-theme cross-feeding panel ρ heatmap across 8 species, showing the bile-acid 7α-dehydroxylation network, together with the top 25 cross-feeding-triangle candidates scatter on which only 7 strict triangles appear. ₁

The H3b strain-adaptation test was restricted to the Kumbhari/LSS-PRISM cohort, the only project data with strain-level resolution, and ran on precomputed tables under the no-raw-reads scope: a gene × species strain-inference table of 219,121 rows covering 59 species, cross-referenced with a strain-competition table of 15,520 strain × sample rows covering 100 UHGG genome IDs in 94 LSS-PRISM participants. Only F. plautii of the six actionable Tier-A core species appears in the 59-species panel — H. hathewayi, M. gnavus, E. lenta, E. bolteae and E. coli are absent because Kumbhari focused on commensal strain heterogeneity rather than pathobiont mechanism — so the test evaluates the methodology of strain-adaptation gene inference on 59 commensals, with only one actionable Tier-A species directly tested. ₁

Kumbhari strain-adaptation analysis identified 23,579 FDR < 0.10 genes across 59 species, comprising 12,159 IBD-biased and 11,420 health-biased genes; the committed outputs record IBD-biased and health-biased gene counts for each of the 59 species and a functional classification of all 23,579 FDR-passing genes. IBD-biased genes were enriched for adaptation at 2.13 % against a 1.59 % background (OR = 1.38, p = 2.4e-6) and depleted for housekeeping at 2.81 % against 4.36 % (OR = 0.62, p = 6.4e-20), with 95.06 % unclassified against a 94.05 % background, so the strict housekeeping-domination falsifiability bound was rejected at p = 6.4e-20. ₁

Within the classified IBD-biased adaptation genes, mucin/glycan utilization dominated with 94 genes, followed by two-component signaling (52), antibiotic resistance (35), membrane transport specialty (33), and virulence/secretion, bile acid and oxidative stress response (15 each). The recurring cross-species orthologs, each IBD-biased in 15–21 of the 59 species, were K03088 in 21 species (sigma factor RpoE/RpoS, environmental stress response), K01990 in 20 (ABC transporter ATP-binding protein), K06147 in 19 (ABC-2 type transport system), K02529 in 18 (LacI/GalR family transcriptional regulator), K07240 in 17 (TIM44-like protein), K03424 in 17 (TatD-related deoxyribonuclease), K08303 in 16 (SmpB, tmRNA-binding translation rescue) and K03466 in 15 (DNA topoisomerase); top gene symbols consistent across 9–10 species each were corC (Mg²⁺/Co²⁺ efflux), luxQ (autoinducer-2 quorum sensing), pglA and rfbC (capsule polysaccharide / LPS), bepC (efflux pump) and abgB (aminobenzoyl-glutamate utilization), a pattern dominated by transport, regulation, stress response and DNA repair. ₁

F. plautii had 0 strain-adaptation genes passing FDR < 0.10 in the Kumbhari cohort among 3,245 genes tested, an informative null the report reads as CD-association operating through species-level abundance rather than strain-level genomic adaptation; the supporting premise that the core bai-operon 7α-dehydroxylation genes are present in essentially all F. plautii strains is stated as a presumption rather than a measurement, and the design consequence drawn — that phage targeting of F. plautii would deplete bile-acid 7α-dehydroxylation activity proportionally to abundance reduction, with no strain-level escape route via gene-content selection — is an inference, not a measured phage intervention. The H3b verdict is SUPPORTED: the gene-content signal is biologically interpretable and statistically distinct from housekeeping artifact, and the shared cross-species IBD-adaptation set (sigma factor, ABC transport, quorum sensing, capsule LPS, efflux) is consistent with a real IBD-niche gene signature operating across multiple commensal species. Figure NB10a presents per-species IBD-biased and health-biased gene counts for the top 25 species, functional category × direction on a housekeeping-versus-adaptation log-scale bar, and an F. plautii top-genes panel that is empty — the informative null. ₁

The serology test (H3e) asked whether anti-microbial antibody titers correlate with Tier-A pathobiont abundance within IBD patients, across 67 subjects at three sites (CCHMC, Harvard, Emory), with site as a covariate and an effect threshold of |r| > 0.40 calibrated to n = 67 power 0.80 at α = 0.05; the single-cohort caveat was structural in the plan. Inputs were curatedMetagenomicData-fetched HMP2 MetaPhlAn3 abundances plus the mart’s fact_serology table of 2,520 measurements across 12 assays, and all 67 serology subjects (E-prefix Emory, H-prefix Harvard, M-prefix MGH/CCHMC) matched a MetaPhlAn3 metagenomics profile. ₁

The test matrix crossed six EU-deduplicated serology axes — ANCA, ASCA, CBir1, IgA-ASCA, IgG-ASCA and OmpC — with eight species (the 6 actionable Tier-A core plus the 2 module-anchor commensals), giving 48 tests of partial Pearson r residualized on site dummy variables with Benjamini-Hochberg FDR correction. ₁

A cohort sanity check in 32 CD, 14 UC and 21 nonIBD subjects found the canonical IBD-serology patterns intact. ₁

Axis means ran ANCA 17.8 in CD versus 42.6 in UC, the canonical UC-elevated pANCA direction, while ASCA 0.28 versus 0.0, CBir1 29.0 versus 13.7, IgA ASCA 14.2 versus 1.2 and IgG ASCA 15.0 versus 3.6 were all CD-elevated as expected; OmpC at 9.3 versus 7.6 was recorded as only a partial match to the expected slight CD elevation. ₁

The leading site-adjusted pairs, with raw ρ and FDR in parentheses, were ANCA × M. gnavus +0.31 (ρ = +0.24, FDR = 0.40; Harvard r = +0.456, p = 0.008, CCHMC r = −0.03 NS), OmpC × E. lenta +0.29 (+0.10, 0.40; CCHMC −0.02 NS, Harvard +0.14 NS), ANCA × H. hathewayi +0.23 (+0.16, 0.59; Harvard +0.345, p = 0.050, CCHMC +0.02 NS), IgA-ASCA × E. coli +0.23 (+0.22, 0.59; Harvard +0.30, p = 0.09, CCHMC +0.06 NS), CBir1 × E. bolteae +0.22 (+0.22, 0.59; CCHMC +0.381, p = 0.035, Harvard +0.04 NS) and ANCA × F. plautii +0.22 (+0.32, 0.59; CCHMC +0.443, p = 0.013, Harvard +0.25 NS) — the site-stratified columns being where the single-site |r| values come from. ₁

Serology did not meet the strict target-correlation threshold: across the 48 assay × species tests in 67 HMP2 subjects by site-adjusted partial Pearson r, no pair cleared the plan’s |r| > 0.40 plus FDR < 0.10 thresholds, and the largest site-adjusted association was ANCA × M. gnavus at r = +0.31, FDR = 0.40. Associations were concentrated in 1 of 2 productive sites and diluted when cohort-aggregated — Harvard 33 subjects, CCHMC 31, Emory 3 — with single-site partial r reaching |r| ≈ 0.40–0.46 individually while the cohort pull toward null kept the plan’s |r| > 0.40 plus FDR < 0.10 threshold unmet, so strict H3e is NOT SUPPORTED, the verdict recorded as PARTIAL but NOT SUPPORTED at that strict plan threshold; the site-stratified single-site Harvard partial r of +0.46 against that cohort-aggregation pull toward null is the shape of the result, and the single-cohort caveat was structural per the plan rather than an outcome of this analysis. All top pairs were positive, which the report reads as biologically plausible in direction — ANCA co-elevated with M. gnavus, H. hathewayi and F. plautii; CBir1 × E. bolteae coherent for an anti-flagellin antibody against a flagellated species; IgA-ASCA × E. coli at +0.23 as the most mucosal-immunity-relevant axis — all interpretations of correlations rather than demonstrated mechanisms. ₁

The design implication is that serology is not yet a quantitative per-target prioritization signal in this cohort: the canonical CD-versus-UC axes (ANCA-UC, ASCA-CD, CBir1-CD) are intact, so serology can be used as a CD-versus-UC stratifier, but the |r| ≈ 0.31 cohort-level ceiling means a single subject’s anti-M. gnavus ANCA titer cannot confidently predict that subject’s M. gnavus abundance, and multi-cohort meta-analysis — out of project scope — is the structural next step. The H3e verdict is recorded as PARTIAL but NOT SUPPORTED at the strict plan threshold, the PARTIAL framing resting on the cohort sanity check reproducing canonical IBD-serology patterns, the plausible direction of all top pairs, productive sites reaching |r| ≈ 0.40–0.46 individually, and the acknowledged structural single-cohort caveat. Figure NB11 presents the site-adjusted partial Pearson r heatmap across 8 species × 6 serology axes and scatterplots of the top 10 assay × species pairs with FDR annotations. ₁

A further test (H3d-clust) asked whether metabolite-derived clustering is more cross-cohort-stable than the taxonomic ecotype framework: if so, refitting ecotypes on metabolite features would give a more cross-cohort-portable framework for clinical translation. It used the m/z-bridge feature panel together with pooled HMP2 and FRANZOSA metabolomics under the no-raw-reads scope. ₁

The method pooled 326 subjects (106 HMP2 × 220 Franzosa) on the 111 unique m/z-bridge metabolite features, standardized per metabolite, reduced to 15 principal-component-analysis (PCA) components, and fit K-means with K = 4, comparing against the taxonomic-ecotype leave-one-substudy-out ARI baseline of 0.113 and the bootstrap baseline of 0.16. K = 4 was the plan-imposed value matching the taxonomic framework; the report notes that the natural cluster count on this feature space might instead be K = 2, separating HMP2 from Franzosa. ₁

All four clusters came out cohort-pure despite mixed diagnoses within each: cluster 0 had 86 subjects, 100 % Franzosa (46 Control + 30 UC + 10 CD); cluster 1 had 58, 100 % HMP2 (29 CD + 23 UC + 6 nonIBD); cluster 2 had 134, 100 % Franzosa (78 CD + 46 UC + 10 Control); and cluster 3 had 48, 100 % HMP2 (21 CD + 7 UC + 20 nonIBD). ₁

PC1 explained 79 % of total variance and separated HMP2 (PC1 ≈ +12) from Franzosa (PC1 ≈ −5) cleanly, while PCs 2–15 collectively explained about 10 % of variance and carried the diagnosis information; within each cohort the clusters did carry diagnosis information (cluster × diagnosis χ²(6) = 88.3, p = 7e-17), but cluster boundaries were mostly co-aligned with cohort boundaries. ₁

The pooled HMP2-plus-Franzosa metabolite-feature K = 4 K-means on the m/z-bridge panel (326 subjects × 111 features) produced cluster structure that is cohort-driven rather than diagnosis-driven, with PC1 explaining 79 % of variance and separating HMP2 from Franzosa cleanly — the result the report records as the H3d-clust verdict NOT SUPPORTED. Cross-cohort leave-one-study-out ARI was 0.000 with HMP2 held out and 0.000 with Franzosa held out, mean 0.000, substantially lower than the taxonomic 0.113; the within-pooled bootstrap (80 % × 30) gave 0.937 against the taxonomic 0.16, but the report treats that 0.937 as misleading as a stability metric: it measures how reproducibly pooled K-means recovers the cohort-batch structure under subsampling, which is high because the batch effect is consistently visible in any subsample, and not whether the clustering is biologically informative. The methodological lesson drawn is that m/z-bridge metabolomics requires explicit batch correction (ComBat/SVA/RUV) for cross-cohort clustering, and that the taxonomic-feature ecotype framework at consensus K = 4 remains the project’s primary ecotype basis for the Pillar 4–5 cocktail design. ₁

Figure NB09d presents the PCA scatter coloured by metabolite K = 4 cluster, with PC1’s 79 % variance recorded as cohort batch, the per-cluster diagnosis composition stacked bar given in the caption as χ² = 88, p = 7e-17, and the stability-metrics summary of cross-cohort LOSO ARI 0.000 against the taxonomic 0.113 baseline. ₁

The H3d-clust verdict is NOT SUPPORTED: at cross-cohort LOSO ARI = 0.000 against the 0.113 taxonomic baseline, the metabolite-feature framework on the m/z-bridge panel does not achieve higher cross-cohort stability than the taxonomic framework. The verdict is scoped to raw m/z-bridge clustering without batch correction, since no batch correction was applied and the report expects ComBat, SVA or RUV would likely improve cross-cohort LOSO ARI substantially. ₁

The methodological lesson the report records is that m/z-bridge metabolomics carries unaddressed batch effects that dominate cluster structure: MetaPhlAn3 relative-abundance values are unitless and compositionally constrained, and that constraint supplies a natural normalization that handles cohort differences, whereas metabolite features inherit absolute-intensity scale differences between LC-MS runs (different instruments, ionization tuning, solvent batches) and would require explicit batch correction — ComBat, SVA, RUV or quantile normalization — before clustering. Within-cohort metabolite-feature ecotypes might still be useful for single-laboratory clinical translation; the failure here is specifically the cross-cohort-portability question. ₁

The taxonomic consensus K = 4 ecotype framework therefore remains the project’s primary ecotype basis for the Pillar 4–5 cocktail design, with metabolite-feature ecotypes set aside as a clinical-translation basis because they would require external batch correction first. ₁

A two-modality canonical correlation analysis (CCA) pilot on 106 paired HMP2 subjects produced four canonical correlations of 0.964, 0.928, 0.911, and 0.889. CC1 had r = 0.964 (quoted elsewhere in the report as r = 0.96), cliff CD-versus-nonIBD = +0.498 (quoted elsewhere as +0.50), and Mann-Whitney p = 4e-4; all six actionable Tier-A species loaded CD-positive, the ecotype-defining commensals R. bromii, A. putredinis, and L. eligens loaded CD-negative, urobilin and secondary bile acids loaded CD-negative, and polyamines, long-chain PUFAs, fatty-acid amides, sphingolipids, N-acetylputrescine and E. coli-specific cadaverine loaded CD-positive. The report interprets CC1 as a single dominant CD axis within which the iron-acquisition narrative centred on E. coli AIEC (adherent-invasive E. coli) and the bile-acid 7α-dehydroxylation narrative centred on F. plautii, E. lenta, and E. bolteae sit as orthogonal molecular sub-mechanisms, while noting that the analysis was a pilot and did not include the unavailable pathway modality. Its data scope was 106 paired HMP2 CSM* subjects × 130 species at ≥ 10 % prevalence (CLR-transformed) plus 582 named metabolites (log10 intensity, ≥ 30 % non-NaN coverage), with per-modality PCA to 30 components each feeding a CCA that returned 4 canonical pairs. ₁

Only CC1 separated the diagnoses. CC1 had mean subject scores of +0.235 in CD, +0.123 in UC and −0.593 in nonIBD, with cliff CD-versus-nonIBD = +0.498 at Mann-Whitney p = 4e-4; CC2, despite a canonical r of 0.928, reached only cliff +0.092 at p = 0.51; CC3 (r = 0.911) cliff −0.146 at p = 0.30; and CC4 (r = 0.889) cliff −0.274 at p = 0.05, borderline at best. ₁

The CC1 species loadings recapitulate the whole actionable Tier-A set on the CD-positive side — M. gnavus +0.195, E. coli +0.173, F. plautii +0.153, H. hathewayi +0.144, E. lenta +0.109, E. bolteae +0.103 — joined by the oral-gut Tier-B candidates V. parvula +0.194, A. intestini +0.161 and V. atypica +0.151, while the ecotype-defining commensals load negative: R. bromii −0.173, R. bicirculans −0.170, A. putredinis −0.169, plus L. eligens and B. intestinihominis. The report reads this as the pathobiont module structure independently rediscovered as a single principal direction in joint species–metabolite space, with CLR-Spearman-Louvain modules and CCA joint factors converging on the same biology. ₁

On the metabolite side, the CD-direction negative (depleted) loadings were urobilin in 3 instances at a maximum of −0.143, glycodeoxycholate −0.113, caproate −0.113 and lithocholate −0.080. ₁

The CD-direction positive (elevated) loadings group into the project’s mechanism axes: the fatty-acid amides linoleoyl ethanolamide twice (+0.110, +0.106) and palmitoylethanolamide +0.108, which the report ties to the CD-up ebf/ecf BGC family at p < 1e-31; sphingosine-isomer1/2 (+0.109, +0.105) and 7-methylguanine +0.121; the polyamines N-acetylputrescine +0.105, diacetylspermine +0.091, cadaverine +0.086 and putrescine +0.084, connecting to the polyamine OR = 14.6 enrichment and the cadaverine × E. coli ρ = +0.45; the long-chain PUFAs docosapentaenoate twice (+0.092, +0.091), adrenate +0.081 and arachidonate +0.048, connecting to the PUFA OR = 7.9 theme and its 75 % cross-cohort concordance; and the uremic-toxin marker ADMA/SDMA +0.099, which the report links to arginine catabolism alongside the TMA/choline axis. Together with the negative urobilin and bile-acid loadings on the same axis, the report reads these loadings as recapitulating all of its mechanism narratives — urobilin CD-down, polyamines, long-chain PUFAs, the ebf/ecf fatty-acid amides, cadaverine × E. coli, and bile-acid depletion — in a single joint axis. ₁

Figure NB07d presents the joint factor space as a CC1 × CC2 sample scatter coloured by diagnosis (CD/UC/nonIBD) across the 106 subjects, with the CC1 top 15 species loadings — Tier-A core CD-positive against commensal CD-negative — and the CC1 top 15 metabolite loadings, urobilin CD-DOWN with polyamines, fatty-acid amides, sphingosine and ADMA CD-UP. ₁

The pilot is two-modality rather than three because the pathway modality was dropped under the data-scope constraint — HMP2 pathway abundance is not in the mart, whose fact_pathway_abundance contains only CMD_IBD_PATHWAYS, so the planned 3-modality MOFA+ fell back to 2-modality joint factor analysis using sklearn CCA on per-modality PC scores, mofapy2 being unavailable; adding pathways by reprocessing cohort-aligned subjects would give a three-modality MOFA (multi-omics factor analysis) but requires sample-level pairing that the available pathway slice does not have with HMP2 metabolomics. ₁

CCA on PC scores captures the same canonical-correlation signal MOFA+ would in the two-modality case, but MOFA+ would additionally model modality-specific factors that load on only one modality; the four PCs used here are joint factors, and modality-specific structure remains in the per-modality PCA components. ₁

Ecotype is implicit in the species-loading structure — E1-Bacteroides2-transitional and E3-Bacteroides-expanded species both load positive on CC1 while E0-commensal species load negative — so the diagnosis-associated axis may equally be an ecotype axis; regressing factor scores on ecotype plus diagnosis to test whether residual factor variance encodes ecotype-specific biology remains future work. ₁

The analysis is explicitly pilot-scale and not a substitute for a full MOFA+ with three modalities and proper factor-relevance gating; that all four canonical correlations exceed 0.88 suggests the per-modality PCAs are over-aligned, where a proper MOFA+ would impose sparsity to gate factor count. ₁

At n = 106 subjects the decomposition is modest for four factors, and the signal beyond CC1 in CC2, CC3 and CC4 is suggestive rather than strongly disease-discriminative. ₁

Phage evidence and cocktail design

The Pillar 4 design framework starts from a per-pathobiont phage-availability profile — Tier-B in the project’s four-tier rubric — layered on top of the actionable Tier-A scoring and the Pillar-3 mechanism profile, with the foundational matrix built from ref_phage_biology, a 12-organism literature-curated synthesis drawn from the indicator-taxa literature review. ₁

The Tier-B phage-availability score is a 0–3 ordinal scale: 0 = no known phages, 1 = temperate or limited, 2 = lytic literature, 3 = clinical trial or commercial cocktail. ₁

The report flags as critical that the two highest-scoring species on the Tier-A scale — H. hathewayi at 4.0 and M. gnavus at 3.8 — have the weakest phage availability, and that F. plautii additionally carries the highest bile-acid coupling cost, so its phage targeting may be deprioritized in favour of bile-acid-binding co-therapy. On that scale the two coverage gaps and the temperate-only case are: H. hathewayi, the highest Tier-A score at 4.0, phage score 0 with no known phages and low bile-acid coupling cost, making external-database querying (INPHARED plus IMG/VR) the priority; M. gnavus at 3.8 with phage score 1, its 6 known phages all temperate and low coupling cost, leaving lytic-locked engineering or a biochemical glucorhamnan-synthesis target; and F. plautii at 3.3 with phage score 0 and phage status recorded as unknown because it is absent from ref_phage_biology, combined with the highest bile-acid coupling cost, so phage targeting is deprioritized in favour of bile-acid monitoring or biochemical alternatives. ₁

The three species with usable lytic options are: E. coli (AIEC) at Tier-A 3.6 with phage score 3, lytic plus clinical and low coupling cost — the EcoActive cocktail of 7 lytic phages in clinical trials, plus the FimH-targeting HER259, making it the most advanced case; E. lenta at 3.3 with phage score 2 from lytic literature and moderate coupling cost, PMBT5 a characterized siphovirus against a non-BGC drug-metabolism mechanism; and E. bolteae at 2.8 with phage score 2 and moderate coupling cost, PMBT24 a virulent 99,962 bp Kielviridae phage, described as the best-characterized lytic phage among the gut-anaerobe Tier-A species. ₁

The six actionable Tier-A species therefore stratify into four phage-availability classes: class 3 clinical (1 species, E. coli, with the EcoActive cocktail), class 2 lytic literature (2, E. lenta with PMBT5 and E. bolteae with PMBT24), class 1 temperate-limited (1, M. gnavus, with 6 temperate phages) and class 0 gap (2, H. hathewayi and F. plautii). ₁

A per-pathobiont design matrix crosses Tier-A score, iron specialization, bile-acid-network role and mechanism class into a design recommendation, and places three species in a Tier-1 phage-target class. H. hathewayi, the highest score at 4.0, is listed with no iron specialization and as not in the bile-acid network, its mechanism read as species abundance plus a within-carrier metabolic shift (biosynthesis-up, sugar-down) from the within-carrier CD-versus-nonIBD meta-analysis, which allowed a single-substudy fallback where cells were too small — the same analysis that put E. coli’s per-pathway abundance CD-down within carriers — so its low bile-acid cost makes broad targeting acceptable. M. gnavus at 3.8 is likewise listed with no iron specialization and no bile-acid-network role, with a species-abundance mechanism whose mucin-glucorhamnan production the report attributes to Henke 2019; the matrix annotates it as showing no Kumbhari strain-adaptation signal, which is inconsistent with the report’s own strain-adaptation section recording M. gnavus as absent from the 59-species panel and therefore never tested — an unreconciled internal inconsistency, not a measured null. E. coli at 3.6 is a Tier-1 target with a strain-resolution requirement: its Yersiniabactin/Enterobactin/Colibactin content is the matrix’s dominant iron specialization, within-carrier CD-down plus 54 iron MIBiG matches make the mechanism strain-content-mediated and specific to the AIEC subset, and the proposed targeting criterion is to select pks/Yersiniabactin/Enterobactin-positive strains specifically; it sits outside the bile-acid network at low bile-acid cost. ₁

The three Tier-2 targets all carry bile-acid coupling costs. F. plautii at 3.3, listed with weak iron specialization, is an active 7α-dehydroxylator (ρ × cholate = −0.26, × lithocholate = +0.15), and the matrix reads its CD association as species-abundance-mediated on the strength of the strain-adaptation informative null — an interpretation of that null rather than a demonstrated exclusion of strain-content effects; it carries the highest bile-acid coupling cost on the prediction, untested here, that depletion would shift the bile-acid pool toward inflammatory primary tauro-conjugated forms, so the report suggests co-administering UDCA or a bile-acid-binding agent. E. bolteae at 2.8, with no iron specialization and a mixed mechanism, is an active 7α-dehydroxylator (× deoxycholate = +0.17, × lithocholate = +0.18) at moderate bile-acid cost. E. lenta at 3.3, with no iron specialization, is a partial 7α-dehydroxylator (× ketodeoxycholate = −0.14) at moderate cost whose primary CD association is a non-BGC drug-metabolism mechanism (Koppel 2018 Cgr2; CB-ORF CD-down 54 %), so it may not require strain resolution. ₁

Figure NB12, the Pillar-4 opener, presents the pathobiont × phage targetability matrix in two panels: a Tier-A × Tier-B scatter and a Pillar-5 priority bar. ₁

The strategy tiers that follow are: direct phage targeting (Tier-1) for the E. coli AIEC subset with the EcoActive clinical-trial cocktail or the project’s own NB13-derived five-phage cocktail of DIJ07_P2, LF73_P1, AL505_Ev3, 55989_P2 and LF110_P2, requiring a strain-resolution AIEC diagnostic by pks-island plus Yersiniabactin plus Enterobactin gene-presence detection; phage targeting with monitoring (Tier-2) for E. lenta (PMBT5) and E. bolteae (PMBT24) with bile-acid pool monitoring, and for M. gnavus if lytic-locked phage engineering succeeds; highest-priority external-database querying of INPHARED plus IMG/VR for H. hathewayi, falling back to GAG-degrading enzyme inhibitors if no phages are found; and lowest Pillar-5 priority for F. plautii on its highest bile-acid cost, with deprioritization or co-administration of UDCA or a bile-acid-binding agent considered instead. ₁

The three-layer phage evidence stack classified the six actionable targets into clinical-trial-stage E. coli (the EcoActive precedent), lytic-literature E. lenta (PMBT5) and E. bolteae (PMBT24), temperate-only M. gnavus (6 phages), and coverage gaps for H. hathewayi and F. plautii. PhageFoundry’s phagefoundry_strain_modelling dataset, which the report calls the definitive E. coli AIEC phage-cocktail design dataset for this project, contained 96 phages, 188 E. coli strains, and 17,672 experimentally tested susceptibility pairs from the Gaborieau 2025-10-02 phage-prediction experiment reported at AUC = 0.88, with 3,929 susceptible and 13,743 resistant pairs and a 22 % susceptibility rate. Greedy minimum-set-cover design produced a five-phage cocktail—DIJ07_P2, LF73_P1, AL505_Ev3, 55989_P2, and LF110_P2—covering 94.7 % of the 188 strains; the report also describes this as approximately 95 % coverage, and pairs that greedy minimum-set-cover result with an AIEC phylogroup B2/D coverage of 65/94 phages. ₁

The greedy ladder reads: DIJ07_P2 (Phapecoctavirus) alone covers 63.8 % of the 188 strains; adding LF73_P1 (Tequatrovirus) gives 81.4 %; four phages, adding AL505_Ev3 and 55989_P2, give 92.6 %; the five-phage row labelled as the ≥ 95 % coverage target reaches only 94.7 %, so that target is not met; and the eight-phage row labelled as the ≥ 99 % target, adding NIC06_P2, LF73_P4 and BCH953_P4, reaches only 98.4 %, so that target is likewise not met. ₁

The extended eight-phage design reached 98.4 % coverage. Of 94 phages with host-phylogroup information, 65 (69 %) were isolated against B2/D hosts, the phylogroups emphasized as AIEC-relevant, and the broadest-host-range B2/D phages include ones isolated against the canonical AIEC and UPEC reference strains LF82, LF73 and 536 — LF82_P8 (60.1 %, Mosigvirus), LF73_P1 (62.8 %), and 536_P7 / 536_P9. Twenty-six strains (14 %) were phage-resistant at ≤5 % susceptibility and are flagged as escape candidates whose clinical significance the report says needs cross-reference to AIEC pks-island annotation. HMP2 endogenous phageome analysis covered 630 samples and found Gokushovirus WZ-2015a CD-down in E1 with cliff δ = −0.358, FDR = 5e-7, n_CD = 231, and n_HC = 125; E. coli correlated with Podoviridae at ρ = +0.183 and Myoviridae at ρ = +0.125. ₁

Figure NB13 presents the PhageFoundry quantitative E. coli phage-cocktail design in three panels: phage host range, strain susceptibility, and the cocktail-coverage curve. ₁

The PhageFoundry collection is E. coli-rich but sparse for everything else: its 96 phages × 188 E. coli strains × 17,672 susceptibility pairs are what make a quantitative five-phage cocktail design possible for E. coli AIEC at all, a design the report labels as 95 % strain coverage on the tested panel — a rounded label distinct from the 94.7 % its set-cover table actually reaches — while the other 5 actionable Tier-A species are not present in phagefoundry_strain_modelling, so their phage availability rests on the ref_phage_biology literature curation alone. ₁

The PhageFoundry and HMP2 viromics evidence layers do not overlap at all: the unique E. coli phages observed in HMP2 fact_viromics — D108, EC6, ECML-117, Murica, slur16, vB_EcoM-VpaE1 and vB_EcoM_AYO145A, given here as seven, a count the report states elsewhere as 16 unique E. coli phages across the same sample-rows without reconciling the two — share no names with any PhageFoundry phage. The report treats the two datasets as complementary rather than overlapping, PhageFoundry being research and clinical isolates with experimental susceptibility while HMP2 viromics records natural phages observed in patient stool, and draws from the zero overlap that in-vivo gut delivery of the cocktail phages is not validated by the project’s data and would require in-vivo testing. ₁

The project’s own NB13 synthesis row records the PhageFoundry inputs as 96 phages × 188 strains × 17,672 susceptibility pairs, with a five-phage cocktail covering 95 % of E. coli strains and 69 % of phages in an AIEC-relevant phylogroup. The 96 phages are the dataset total, while the host-phylogroup analysis is computed over the 94 phages that carry host-phylogroup information — two distinct denominators the report keeps separate. The rounded 95 %, by contrast, differs from the 94.7 % the set-cover table actually reaches, so that percentage should be read as a hand-off label or coverage target rather than the measured result, and the report does not reconcile it. ₁

The report attributes to Dogan 2014 and Dubinsky 2022 an approximate AIEC phylogroup distribution of B2 ~80 % and D ~20 %, and groups the PhageFoundry phages by isolation-host phylogroup: 50 phages isolated against B2 hosts, the primary AIEC phylogroup, at median 21.3 % and maximum 63.3 % host range; 15 against D hosts, the secondary AIEC phylogroup, at median 13.8 % and maximum 63.8 %; and 10 against predominantly commensal A hosts at median 16.5 % and maximum 39.4 % — those host ranges being measured across the tested 188-strain E. coli panel rather than within the isolation phylogroup. ₁

Direct phage observations corroborate the gut-anaerobe coverage gap from the observational rather than the literature-curation side: across 3,039 HMP2 fact_viromics sample-rows of patient stool, 16 unique E. coli phages and 13 unique Klebsiella phages were observed, with no observations at all of phages targeting H. hathewayi, M. gnavus, F. plautii, E. lenta or E. bolteae. ₁

A separate precomputed 22-virus CD-versus-nonIBD differential-abundance table (ref_viromics_cd_vs_nonibd) had as its top hit Gokushovirus WZ-2015a CD-down at log2FC = −2.7, FDR = 1e-11, a microvirus depleted in CD. ₁

The per-ecotype endogenous phageome analysis cross-referenced HMP2 fact_viromics — 3,039 sample-rows × 273 viruses on a VirMAP taxonomic profile — with the ecotype projections, yielding 630 of 648 viromics samples (97 %) with ecotype calls spanning CD, UC and healthy; the report’s own per-ecotype breakdown of 27 (E0), 484 (E1), 55 (E2) and 50 (E3) does not add up to that stated 630 total, a discrepancy it does not address. ₁

Across 85 per-virus CD-versus-nonIBD Mann-Whitney differential-abundance (DA) tests spanning the four ecotypes, three passed false-discovery rate (FDR) < 0.10: Gokushovirus WZ-2015a in E1 at cliff δ = −0.358 on 231 versus 125 samples, FDR = 5e-7; Gokushovirus WZ-2015a in E2 at −0.471 on 21 versus 20 samples, FDR = 0.056; and human feces pecovirus in E2 at −0.300 on the same 21 versus 20 samples, FDR = 0.056. Gokushovirus WZ-2015a is therefore the virus the report calls robustly CD-down across multiple ecotypes, strongest in E1; across the 3,039 HMP2 fact_viromics rows and the 630 ecotype-called samples (97 % ecotype-call coverage) that Microviridae depletion — E1 cliff = −0.358, FDR = 5e-7, n_CD = 231 versus n_HC = 125 — is recorded as consistent with Norman 2015 and Clooney 2019, and it independently rediscovers the precomputed ref_viromics_cd_vs_nonibd top hit (log2fc = −2.7, FDR = 1e-11). ₁

Species × phage-family Spearman correlations on the 630 paired viromics and MetaPhlAn3 samples gave E. coli × Podoviridae ρ = +0.183 (p = 4e-6), H. hathewayi × Unknown ρ = −0.150 (p = 2e-4), M. gnavus × Unknown ρ = −0.134 (p = 7e-4), E. coli × Myoviridae ρ = +0.125 (p = 0.002) and E. bolteae × Myoviridae ρ = +0.119 (p = 0.003); the report gives the two E. coli values elsewhere as Podoviridae +0.18 and Myoviridae +0.13 and reads them as confirming in-vivo carriage of E. coli phages. ₁

No correlation reached the report’s strong-candidate threshold of |ρ| > 0.30, so it records no strong endogenous phage candidates targeting Tier-A pathobionts; its accompanying statement that all correlations are |ρ| ≤ 0.18 conflicts with its own tabulated +0.183 and is left unreconciled. The negative H. hathewayi and M. gnavus associations are with the “Unknown” family that captures 80 % of HMP2 viromics observations VirMAP could not classify, which limits per-family DA and correlation power and hides species-level signal behind family aggregation — the report judges per-virus DA more informative than per-family for IBD-relevant phages — so its reading of those negatives as pathobiont blooms displacing unclassified phages is an interpretation rather than a measurement. The report also contrasts this channel with species-level taxonomy, where within-substudy effect sizes routinely reach |CLR-Δ| > 1.0 while the phageome × ecotype × diagnosis effects are modest at its stated maximum |ρ| ≤ 0.18, and states that the three-layer phage-evidence convergence therefore rests on directional concordance rather than on effect-size matching across layers. ₁

Per-ecotype phage-family abundance showed modest ecotype-specific variation: Anelloviridae E1-specific, Parvoviridae E2-elevated, and the Unknown family E2-dominant. ₁

Figure NB14 presents the HMP2 endogenous phageome by ecotype and diagnosis in three panels: the ecotype × family abundance heatmap, the top 12 DA viruses, and the species × family ρ heatmap. ₁

Per-ecotype DA power is asymmetric — E1 contributes 231 CD samples while E0, E2 and E3 contribute 12-30 each — which limits within-ecotype detection in the smaller ecotypes; the report flags an E0 Gokushovirus cliff of −0.55 as biologically interesting but failing strict FDR < 0.10 on power. HMP2 viromics also carries no CRISPR-spacer-derived phage-host predictions, so discovering phages against Tier-A hosts would require an IMG/VR cross-reference. ₁

The combined Pillar-4 verdict is that phage-therapy feasibility is high for E. coli, with clinical cocktail, experimental susceptibility and in-vivo phage correlation all aligned across the three evidence layers (curated literature from ref_phage_biology; experimental susceptibility from PhageFoundry strain_modelling, 96 phages, 188 E. coli strains and 17,672 pairs; and the in-vivo phageome from HMP2 fact_viromics, 630 samples), while for M. gnavus, H. hathewayi and F. plautii the gap is confirmed across all three layers — no clinical phages, no PhageFoundry coverage, and no HMP2 in-vivo Podoviridae/Myoviridae signal. ₁

Read target by target, the three layers agree: E. coli AIEC is the clinical-trial case (EcoActive) with a five-phage cocktail the matrix labels at 95 % strain coverage — the rounded label recorded above, against the 94.7 % the set-cover table reaches — plus 65/94 phages in an AIEC-relevant phylogroup and +0.18 Podoviridae / +0.13 Myoviridae endogenous correlation, which the report classes Tier-1 phage-feasible at high confidence with all three layers aligned; E. lenta has PMBT5 siphovirus literature but is not in the PhageFoundry collection and shows no detectable in-vivo correlation, giving Tier-2 phage-feasible with monitoring. The remaining rows of the same per-target table read: E. bolteae carries the virulent PMBT24 Kielviridae literature precedent, is not in the PhageFoundry collection, and shows a +0.12 Myoviridae correlation, classed Tier-2 phage-feasible with monitoring; M. gnavus has 6 temperate siphoviruses, is not in the PhageFoundry collection, and shows a weak in-vivo signal, classed limited and temperate-only with alternatives needed; H. hathewayi has no known phages, is not in the PhageFoundry collection, and has no detectable in-vivo phage correlation, classed GAP with external-database querying the priority; and F. plautii is absent from the literature curation and from the PhageFoundry collection with no detectable in-vivo phage correlation, classed GAP and carrying the highest bile-acid coupling cost, so alternatives are preferred. The report treats that per-target agreement as its Pillar-4 rigor signal, explicitly analogous to the five-line iron-acquisition narrative (Novel Contribution #14) and the two cross-corroborated narratives (#17): where one layer shows phage availability the other two tend to, and where one layer shows GAP the others confirm GAP. The inference is bounded, though — absence from an isolate collection and non-detection of a family-level correlation do not establish that no phage exists for a target. ₁

The phage evidence supports direct targeting for E. coli, monitored lytic-phage targeting for E. lenta and E. bolteae, limited or engineered targeting for M. gnavus, and non-phage alternatives or external database searches for H. hathewayi and F. plautii. The two phage-gap species carry Tier-A scores of 4.0 and 3.3, so the recommended INPHARED and IMG/VR queries for the three gut-anaerobe gaps remain the highest-priority follow-up rather than completed evidence, and the report emphasizes that the current phage evidence does not establish in-vivo efficacy of the PhageFoundry cocktail. ₁

Read back onto the ecotype strata, the report’s per-ecotype cocktail-appropriateness note is that E0, 27 % of the UC Davis cohort, has minimal pathobiont burden and possibly does not need a cocktail at all; E1, 42 %, is the dominant CD ecotype for which a full Tier-1 cocktail — E. coli AIEC plus Tier-2 E. bolteae/E. lenta plus bile-acid monitoring — is appropriate; and E3, 31 %, is the severe Bacteroides-expanded stratum in which E. coli and M. gnavus are prominent and the cocktail plus its ecological-cost annotation is more critical because of the higher pathobiont burden. These are ecotype-level design readings rather than tested assignments, and the E1 reading sits beside the report’s own translation projection that a pooled-cohort cocktail could instead be over-aggressive for E1 patients. ₁

Per-patient framework and longitudinal validation

The proposed workflow begins with patient ecotype assignment by projection of stool MetaPhlAn3 profiles onto E0, E1 or E3 — E2 being absent in active CD — and then applies the per-ecotype Tier-A target list: E1_CD module 0 contains H. hathewayi, F. plautii, E. bolteae, E. lenta and M. gnavus, while E3_CD module 1 contains E. lenta, H. hathewayi, E. coli and M. gnavus, so F. plautii is E1-specific and E. coli is E3-specific. ₁

Among 23 UC Davis CD patients, with per-patient inputs comprising the NB02 ecotype call, demographics (Montreal, calprotectin, medication), Kuehl_WGS Kaiju Tier-A pathobiont presence at a ≥ 0.001 relative-abundance threshold, the Tier-A score, the mechanism profile and phage availability, M. gnavus was present in 21 (91 %), H. hathewayi in 19 (83 %), E. bolteae in 19 (83 %), F. plautii in 18 (78 %), E. lenta in 16 (70 %), and E. coli in 8 (35 %) — much lower carriage than expected for an AIEC-canonical species, so the five-phage cocktail is directly applicable to only about 8 of 23 patients, which the report extrapolates to roughly 35 % of the cohort even though E. coli detection alone does not establish AIEC strain status, while the near-universal carriage of H. hathewayi and M. gnavus makes their phage gap the dominant unmet need. The cohort contained 7 E0, 9 E1, 6 E3, and 1 mixed-ecotype patient. Fourteen of 23 patients (61 %) received concrete phage cocktail drafts — a pure phage cocktail being infeasible for E1, which requires a three-strategy hybrid and carries the F. plautii bile-acid-cost constraint as its dominant design limit — and all 9 E1 patients carried the full Tier-A pathobiont module, although the report also records 22 patients with a target-level cocktail profile before distinguishing concrete phage availability. ₁

The per-ecotype summary gives, by ecotype, patient count, mean Tier-A targets per patient, patients with a cocktail profile and patients with concrete phage components: E0 7 patients, mean 1.71 targets, 7 cocktail profiles and 0 concrete phages, because its priority targets H. hathewayi and M. gnavus are a phage gap or temperate-only; E1 9, 4.89, 9 and 9, with the full Tier-A module and concrete cocktails; E3 6, 2.33, 5 and 4; and the single mixed-ecotype patient 6967 at 3.0, 1 and 1. The total row reads 23 patients, mean 4.39 targets, 22 with cocktails and 14 with concrete phage components; that reported total mean does not align with the per-ecotype means and counts above, and the report does not reconcile it. Fourteen of 23 patients therefore have concrete phage cocktail drafts — all 9 E1 patients receiving PMBT24 plus PMBT5 plus the AIEC five-phage cocktail if E. coli is present, 4 of 6 E3 patients, and the one mixed patient — while E0 patients lack concrete components entirely because their priority targets sit in the Pillar-4 gap or temperate-only classes. ₁

Patients were sorted into priority classes. Class A, active disease with multiple targets — 3701 (calprotectin 8000), 1835 (3340), 5843 (2150) and 6434 (251, with 6 targets) — is the highest Pillar-5 priority and is assigned a full hybrid cocktail of phages plus alternatives for the gap species. Class C, quiescent disease — 1112 (41), 2708 (3), 1492 (9) and 1460 (7) — reserves the cocktail for flares with calprotectin monitoring. Class D, mixed ecotype with longitudinal drift, holds only patient 6967 (E1 ↔ E3), assigned state-dependent dosing and treated as the central per-patient stability test for Pillar 5. The active-versus-quiescent split uses a fecal-calprotectin threshold of 250 µg/g, which the report calls a clinical convention not validated against per-patient response data, noting that some patients at calprotectin 41–248 µg/g currently placed in the quiescent class may benefit from an active cocktail and that the threshold should be re-evaluated against clinical response in a pilot. ₁

The per-class strategy draft assigns approximately 3 E1 patients with active calprotectin and E. coli present the full Tier-1 regimen (the NB13 five-phage E. coli cocktail plus PMBT24 plus PMBT5 plus alternatives for F. plautii, H. hathewayi and M. gnavus); approximately 5 active-calprotectin E1 patients without E. coli a Tier-2 regimen of PMBT24 plus PMBT5 plus alternatives; 1-2 quiescent E1 patients a cocktail reserved for flares with calprotectin monitoring; all 7 E0 patients a limited strategy with deprioritization considered, since their priority targets are a gap or temperate-only; approximately 3 E3 patients with E. coli the five-phage cocktail plus PMBT5 plus alternatives; and approximately 3 E3 patients without E. coli a limited E. lenta plus H. hathewayi strategy. These are conditional design recommendations, not patient-level validation, and none rests on a patient-specific AIEC strain diagnostic. ₁

The final NB17 strategy distribution assigns 12 patients to reserve-for-flare with no active cocktail but calprotectin monitoring plus M. gnavus qPCR; 4 to an E1 hybrid three-strategy regimen without E. coli (PMBT24 plus PMBT5 plus alternatives for H. hathewayi, F. plautii and M. gnavus plus UDCA or a bile-acid-binding agent); 4 to an E0 limited cocktail whose priority targets are GAP, with non-phage strategies considered; 1 to the full E1 hybrid (the NB13 five-phage E. coli cocktail plus PMBT24 plus PMBT5 plus alternatives plus UDCA); 1 to an E3-focused regimen of the five-phage E. coli cocktail plus PMBT5 plus alternatives; and 1 to state-dependent dosing for patient 6967, whose cocktail rebalances on E1↔E3 ecotype shifts. The 12 quiescent patients — calprotectin below 250 or unmeasured — receive the same Tier-A panel reserved for flare rather than as an active cocktail. The E1-full row includes E. coli AIEC, which sits against the report’s own state-dependent rule calling E. coli inclusion E3-specific, and the report does not reconcile the two. ₁

Figure NB15 presents the UC Davis per-patient cocktail draft as the Pillar-5 opener, in three panels: a 23-patient × Tier-A presence heatmap sorted by ecotype and calprotectin, a per-ecotype prescribing-rate bar recording M. gnavus 91 % and E. coli 35 %, and a calprotectin × number-of-targets scatter marking the four patient-design categories. ₁

All 9 E1 patients carried the full five-species E1 pathobiont module, but H. hathewayi and F. plautii were phage gaps and M. gnavus was temperate-only. The report states that only 3 of those 5 species have direct lytic phage options at present, while the examples it gives are E. lenta (PMBT5), E. bolteae (PMBT24) and E. coli AIEC if present — E. coli not being one of the five E1 module species, an internal inconsistency it does not reconcile. The report therefore concludes that a pure phage cocktail is structurally infeasible for E1 and requires a three-strategy hybrid: direct phage targeting where lytic options exist (E. coli AIEC, E. bolteae PMBT24, E. lenta PMBT5), alternatives such as H. hathewayi GAG-degrading enzyme inhibitors or F. plautii bile-acid co-therapy, and limited or engineered approaches for M. gnavus such as lytic-locked phage engineering or a biochemical glucorhamnan-synthesis target; the INPHARED and IMG/VR external-database queries that could close the two gaps are flagged as the highest-priority external-data extension before clinical translation but are out of scope under plan v1.9, which restricted the project to platform-internal data, so the hybrid structure reflects gut-anaerobe phage availability rather than a project limitation per se. F. plautii — present in 18 of 23 patients (78 %), including all nine E1 patients — is deprioritized despite its Tier-A score because it combines the highest bile-acid coupling cost with a phage gap, a double penalty the report calls the dominant E1 design constraint, so that the three-strategy hybrid requirement and the F. plautii bile-acid cost together sit inside a scheme of four patient-design categories — A active with many targets, B active with few targets, C quiescent, and D mixed-ecotype; because phage targeting is predicted to shift the bile-acid pool toward inflammatory primary tauro-conjugated forms, the proposed remedy is to deprioritize F. plautii from the cocktail or co-administer UDCA (ursodeoxycholic acid) or a bile-acid-binding agent — a design proposal, not a validated or patient-level tested regimen. ₁

The report states that deprioritization two ways without reconciling them: the E1 design passage calls F. plautii’s combination of highest bile-acid coupling cost and phage gap a double penalty, while the NB15 design principle calls it a triple penalty and treats F. plautii as a “leave alone” target despite its NB05 Tier-A score of 3.3 — present in 78 % of UC Davis patients, highest bile-acid coupling cost as an active 7α-dehydroxylator, and Pillar-4 phage GAP through absence from ref_phage_biology. The remedy proposed under the triple-penalty framing is to avoid F. plautii in the cocktail and instead use bile-acid-binding co-therapy (UDCA or cholestyramine) to maintain the secondary bile-acid pool that F. plautii’s depletion would otherwise erode — a design proposal rather than a measured outcome. ₁

A second NB15 design principle is that E. coli AIEC is over-emphasized in the Crohn’s-disease literature relative to this cohort’s prevalence: only 8 of 23 UC Davis patients carry detectable E. coli (35 %), against the canonical framing of AIEC as the prototypical CD pathobiont (Darfeuille-Michaud 2004), so the report estimates the NB13 five-phage cocktail to be directly applicable to about 35 % of patients while H. hathewayi at 83 % and M. gnavus at 91 % carry the unmet need. That applicability figure is the report’s own conditional estimate rather than a measured coverage: detection of E. coli does not establish AIEC strain status, no patient-specific AIEC strain-resolution diagnostic was run, and no per-patient susceptibility or in-vivo delivery validation exists in this project’s data. That same passage also calls both unmet-need species Pillar-4 GAP, which conflicts with the report’s own availability table classifying M. gnavus as temperate-only and limited rather than a gap; the report does not reconcile the two labels. ₁

Patient 6967, the only multi-timepoint UC Davis CD patient with biological-replicate samples and the only one with documented E1↔E3 ecotype drift across two visits, shifted from E1 (projection confidence 0.64) to E3 (confidence 0.41). M. gnavus increased from 0.53 to 7.45 reads, a 14.0× expansion; H. hathewayi went from 0.27 to 0.36, a 1.3× change; E. lenta increased 3.1× (0.40 to 1.24), F. plautii 1.9× (0.84 to 1.62) and E. bolteae 2.1× (0.25 to 0.54), while E. coli remained 0.00 at both visits. The report calls the M. gnavus 14× expansion the dominant signature of the E1→E3 transition, with all other Tier-A pathobionts also expanding (1.3–3.1×) as general dysbiosis worsening, and the visit-to-visit cocktail Jaccard of 0.60 (3 shared components, 2 visit-1-only) as its design consequence — a single-patient observation rather than a cohort pattern. ₁

At visit 1 the E1_CD module-0 list gave 5 priority targets, all 5 present, so the drafted cocktail was H. hathewayi, M. gnavus, E. bolteae, E. lenta and F. plautii; at visit 2 the E3_CD module-1 list gave 4 priority targets of which 3 were present, E. coli being absent — which the report reads as patient 6967 not being an AIEC carrier — so the draft became E. lenta, H. hathewayi and M. gnavus. H. hathewayi, M. gnavus and E. lenta were shared across both visits — the report’s universal Tier-1 trio — while E. bolteae and F. plautii were visit-1-only, E1-specific targets that would be deprioritized on an E3 transition. The visit-level cocktail Jaccard was 0.60, which the report reads as moderate overlap implying non-trivial cocktail re-design rather than a complete rewrite. ₁

Patient 6967 is the only multi-timepoint UC Davis CD patient with biological-replicate samples, while patient 1112 contributed two resequencing replicates of the same biological sample and was used as a Kaiju reliability validation rather than as a longitudinal comparison: H. hathewayi 0.445 versus 0.436 (2 %), M. gnavus 7.916 versus 7.792 (1.6 %), E. coli 0.000 versus 0.000, E. lenta 0.712 versus 0.642 (10 %), F. plautii 2.019 versus 1.994 (1.2 %) and E. bolteae 0.864 versus 0.862 (0.2 %), with Tier-A Spearman ρ = 1.000 at p < 0.001 — perfect rank concordance across resequencing replicates, which the report reads as Kaiju being reliable across such replicates and as establishing the technical-noise floor. ₁

Figure NB16 presents patient 6967’s longitudinal trajectory together with the technical-replicate concordance and the cocktail-composition shift, in three panels: patient 6967’s per-visit Tier-A abundance with the E1 → E3 transition and the M. gnavus 14× expansion highlighted, patient 1112’s technical-replicate scatter at Spearman ρ = 1.000, and the cocktail-composition shift Venn whose visit-1 ∩ visit-2 intersection is the 3 universal Tier-1 species at Jaccard 0.60. ₁

The per-patient cocktail draft therefore covers the 23 UC Davis patients crossed with ecotype and Tier-A profile plus bile-acid coupling-cost annotations and medication harmonization, with patient 6967’s longitudinal E1 ↔ E3 shift as the central per-patient stability test; the report presents the H3 framework plus the CC1 integrative pilot as giving this design layer a mechanistically grounded multi-attribute decision framework rather than a single-axis Tier-A score. The Pillar-5 hand-off is recorded as a complete clinical workflow of an initial visit, follow-up calprotectin and qPCR monitoring, and an ecotype-shift decision tree. ₁

The proposed state-dependent workflow recommends ecotype reassessment every 3–6 months, retaining the universal Tier-1 trio of M. gnavus, H. hathewayi, and E. lenta — which the report says spans both E1 and E3 and therefore does not need re-evaluation on an ecotype shift — as the cocktail backbone, dropping F. plautii after E1→E3 transition, and considering E. coli targeting in E3 when an AIEC strain diagnostic is positive. A five-fold change in M. gnavus abundance measured by quantitative PCR (qPCR), motivated by the 14× M. gnavus expansion at patient 6967’s E3 transition, is proposed as a cheap clinical proxy triggering full ecotype retesting in place of routine metagenomics, but the report explicitly labels this qPCR proxy and the dosing rules as hypotheses requiring prospective validation, with validation of the qPCR trigger requiring a prospective cohort carrying paired qPCR and metagenomics across timepoints. The Pillar-5 longitudinal verdict pairs the visit-to-visit cocktail Jaccard of 0.60 and the patient-1112 technical-replicate Spearman ρ = 1.000 with these five state-dependent dosing rules, which the report numbers as Novel Contribution #24. The re-test interval is specified for active CD patients on phage cocktail therapy; F. plautii inclusion is E1-specific, so an E1→E3 transition deprioritizes it — which the report says also reduces the bile-acid-coupling-cost concern, F. plautii being an E1_CD module-0 and not an E3_CD module-1 species — while E. coli inclusion is symmetrically E3-specific, so an E3→E1 transition deprioritizes E. coli subject to AIEC strain detection, E. coli sitting on the E3_CD module-1 priority list rather than the E1_CD one. ₁

Project-level synthesis and translation implications

The report’s closing narrative records 23 notebooks across Pillars 1–5 and two adversarial reviews that caught 9 critical plus 16 important methodological issues. Separately, it records that two runs of the project’s routine review command on the pre-rigor-repair NB04 state concluded “no critical issues” while an adversarial reviewer — a general-purpose agent given explicit “find flaws” framing — caught 5 critical plus 6 important issues, all empirically confirmed by NB04b and NB04c; that count is distinct from the project-level tally of 9 critical plus 16 important issues, and the report does not combine them. ₁

The committed artifacts carry the same record in machine-readable form: data/ecotype_assignments.tsv holds 8,489 rows, the K = 4 consensus ecotype per curatedMetagenomicData sample with the separate LDA and GMM calls and a methods_agree flag; data/nb04_tier_a_candidates.tsv retains the 33-row original within-ecotype Tier-A list as retracted and kept for audit; data/nb04c_lme.tsv is empty at 0 rows, which is itself the documentation that the pooled CD-versus-HC linear mixed-effects model is structurally unidentifiable; and the formal verdict files record the leave-one-substudy-out ecotype test as “study-dependent” at mean ARI 0.113, the pathway-ecotype-versus-taxon-ecotype comparison as PARTIAL at ARI 0.113, and the HMP2 external replication as PASS. ₁

The first two claimed deliverables are a reproducible four-ecotype framework at consensus K = 4 — stratifying 8,489 curatedMetagenomicData samples plus 1,627 HMP2 samples plus 23 UC Davis patients into E0 healthy-commensal, E1 Bact2-transitional, E2 Prevotella and E3 severe-Bacteroides, and externally replicated on HMP2 at 80 % projection confidence with disease-stratification χ²(2) = 15.6, p = 0.016, a summary form of the replication section’s 80.4 % of samples at projection confidence > 0.70 and χ² = 15.61, p = 0.016 that the report does not reconcile — and a confound-free within-IBD-substudy CD-versus-nonIBD meta-analysis design credited with resolving both the feature-leakage and substudy-confounding pitfalls of pooled-cohort differential abundance, from which the six actionable Tier-A targets emerge: H. hathewayi (4.0), M. gnavus (3.8), E. coli (3.6), E. lenta (3.3), F. plautii (3.3) and E. bolteae (2.8). The report also claims the framework reproduces published enterotype structure on 8.5 K samples with disease-stratifying signal and improved disease resolution, separating E1 transitional from E3 severe within the Bacteroides-dominant space. The same closing list pairs the six actionable Tier-A species — H. hathewayi, M. gnavus, E. coli, E. lenta, F. plautii and E. bolteae — with HMP2 external replication of the E1 Tier-A list at 88.2 %. ₁

Two cross-corroborated six-line mechanism narratives are claimed: iron acquisition in the E. coli AIEC subset (OR = 8.1 pathway, OR = 44 BGC) and bile-acid 7α-dehydroxylation by F. plautii, E. lenta and E. bolteae through a paired-sample substrate-product signature, each described in this closing section as resting on six independent evidence streams — where the mechanism section above enumerates five streams for iron, a count the report leaves unreconciled. The report further states that all Pillar 3 narratives collapse into a single multi-omics joint factor (NB07d CC1, canonical r = 0.96, cliff CD-versus-nonIBD = +0.50), reading the canonical correlation result as an independent rediscovery confirming that the per-modality analyses converge on the same biology rather than measuring orthogonal phenomena. ₁

The design-side deliverable is a per-target cocktail-design profile distinguishing iron/AIEC-mediated from other CD specialization mechanisms, bile-acid coupling cost, and species-abundance-mediated from strain-content-mediated mechanisms; the shared-environment reframing of H3a is credited with sharpening that profile’s “metabolic-coupling cost” annotation into the more rigorous “bile-acid coupling cost” annotation. ₁

Also claimed are three cross-cohort-replicated metabolite signatures at ≥ 75 % HMP2-to-Franzosa sign concordance — urobilin CD-down, read as loss of bilirubin-reducer commensals; acyl-carnitines CD-up; and long-chain PUFAs CD-up as inflammatory eicosanoid precursors — and five generalizable methodology contributions beyond the project’s specific findings, listed as cMD substudy-nesting unidentifiability, feature leakage in cluster-stratified differential abundance, LOSO ARI as an honest stability metric, pool ≠ flux for pathway-versus-metabolite differential abundance, multi-line cross-corroboration as a portable rigor pattern, and cohort-batch-dominates-clustering for absolute-intensity multi-cohort modalities. ₁

The same closing section is internally inconsistent about completion: it describes the H3 falsifiability framework as fully closed and the five pillars as delivered, while an interpretation paragraph immediately after lists the phage targetability matrix per Tier-A core (PhageFoundry plus external phage databases) combined with the per-target cocktail-design profile, and the per-ecotype and per-patient cocktail drafts for the 23 UC Davis patients, as what remains for Pillars 4–5; the report does not reconcile the two statements. ₁

Two translation consequences are drawn, both as design projections rather than tested outcomes. A cocktail designed against the pooled-cohort top-pathobiont list (R. gnavus, E. bolteae, E. coli, E. lenta, C. difficile, K. oxytoca, from the preliminary report’s donor-2708 engraftment analysis) is projected to mismatch the ~27 % of UC Davis patients in E0, who may have near-healthy microbiomes and not need the cocktail at all, and to be possibly over-aggressive for the 42 % in E1, whose pathobiont burden differs from the severe E3 cluster. And because the clinical-covariate classifier cannot separate E1 from E3 without metagenomics — its nominal AUC passing the bar while the UC Davis translation fails — trial design needs stool metagenomics as a screening step unless a rapid qPCR panel on ecotype-defining species can be developed or the within-IBD signal can be boosted with better training data (HMP2 plus Franzosa medication metadata, pending HMP2 MetaPhlAn3 ingestion), neither of which is demonstrated here. ₁

NB17 deliverables and the out-of-scope clinical-translation roadmap

NB17 records three deliverables: a per-patient master table for the 23 UC Davis CD patients combining all per-pillar attributes plus design category plus cocktail strategy plus longitudinal status; a target decision matrix of 6 actionable Tier-A species × 5 attributes resolving to a final priority class; and a 4-phase clinical-translation roadmap spanning immediate, near, mid and long term. ₁

The decision matrix resolves each actionable species to a final priority class across score, ecotype membership, bile-acid coupling cost, phage tier and UC Davis prevalence: H. hathewayi at 4.0, E1+E3, no bile-acid coupling cost, phage GAP, 83 % — a Tier-1 phage GAP making INPHARED/IMG-VR querying the priority; M. gnavus at 3.8, E1+E3, no bile-acid cost, temperate-only, 91 % — Tier-1 limited, needing lytic-locked engineering or a biochemical glucorhamnan target; E. coli at 3.6, E3-specific, no bile-acid cost, clinical-trial, 35 % — Tier-1 phage with strain resolution via the NB13 five-phage cocktail and a required AIEC diagnostic; E. lenta at 3.3, universal Tier-1 across ecotypes, moderate bile-acid cost, lytic (PMBT5), 70 % — Tier-2 phage with bile-acid monitoring; F. plautii at 3.3, E1-specific, highest bile-acid cost, phage GAP, 78 % — Tier-2 deprioritize on its triple penalty, with UDCA or bile-acid-binding co-therapy; and E. bolteae at 2.8, E1+E3, moderate bile-acid cost, lytic (PMBT24), 83 % — Tier-2 phage with bile-acid monitoring. ₁

The matrix exposes the structural shape of the clinical-translation problem rather than resolving it: the two species the report’s own wording calls highest-scoring (H. hathewayi and F. plautii, a ranking it contradicts elsewhere) are both Pillar-4 GAP, the species with the most complete phage precedent (E. coli) is the rarest in UC Davis at 35 % carriage, F. plautii should be deprioritized despite actionable status, and M. gnavus is near-universal but temperate-only. ₁

The project-outline summary compresses the phage deliverable to a concrete 5-phage E. coli AIEC cocktail at 95 % strain coverage with the gut-anaerobe Tier-A coverage gap confirmed across all 3 layers — the 95 % being the rounded label recorded above, against the 94.7 % the set-cover table reaches. The final per-pillar verdict reads all 5 pillars closed, 31 notebooks and 24 Novel Contributions, a notebook count that sits against the 23 notebooks the closing narrative records, and the report does not reconcile the two. ₁

The claim of completion is explicitly scoped: the report states that the project’s central science question is fully answered within plan v1.9 scope (no raw reads), and flags the items in the NB17 clinical-translation roadmap as out-of-scope follow-ups required to bridge from “computational deliverables” to a “clinical pilot” rather than as work performed here. ₁

The roadmap’s phage-discovery item is driven by the coverage gap: its own wording calls H. hathewayi at 4.0 and F. plautii at 3.3 “the 2 highest-NB05-scored species” that are Pillar-4 GAP, plus M. gnavus temperate-only. That ranking is a disputed source claim rather than an established ordering, because the Tier-A scores the report retains elsewhere place M. gnavus at 3.8 and E. coli at 3.6 above F. plautii’s 3.3, and the report does not reconcile the two; the gap itself stands, and the report names INPHARED (~25K phages) plus IMG/VR (~3M UViGs) as covering the gut-anaerobe phage-discovery space — external datasets proposed for follow-up, not queried in this project. ₁

A second roadmap item is an AIEC strain-resolution diagnostic for the 8 of 23 E. coli-positive UC Davis patients, on the report’s interpretation that the E. coli CD-association is strain-content-mediated (the AIEC subset). The NB13 five-phage cocktail’s clinical applicability requires per-patient strain genotyping — pks-island plus Yersiniabactin plus Enterobactin gene-presence detection via PCR or amplicon panel — so applicability is conditional on a test not run here. ₁

A third item is a cheaper follow-up assay: the 14× M. gnavus expansion in patient 6967’s E3 transition suggests qPCR could substitute for full metagenomics in clinical follow-up, with validation requiring a prospective cohort carrying paired qPCR and metagenomics across timepoints. The proposed targeted ecotype panel is 4–6 species (F. prausnitzii, P. copri, P. vulgatus, B. fragilis, M. gnavus), proposed because H1c demonstrated that clinical-covariate-only assignment is not viable. ₁

A fourth item is per-patient bile-acid measurement: the bile-acid coupling cost is currently ecotype-level (NB09c §13) rather than per-patient, so clinical translation requires per-patient bile-acid measurements. ₁

A fifth item is multi-cohort serology meta-analysis, which the report says would resolve what that future-directions line calls the cohort-aggregate versus site-stratified r = 0.46 versus +0.31 gap; those labels are inverted relative to the report’s own NB11 results, where +0.46 is the Harvard site-specific partial r and +0.31 the site-adjusted cohort-level value, and the report does not reconcile the two labelings. ₁

A sixth item is multi-patient longitudinal cohorts: the state-dependent dosing rule is derived from an n = 1 trajectory, and such cohorts would test whether ecotype drift is common, whether the M. gnavus 5× expansion is the right qPCR threshold, and whether the cocktail Jaccard of 0.60 generalizes. ₁

A seventh item is additional cMD sub-studies with CD plus nonIBD samples that populate E3, of which only HallAB_2017 currently qualifies, so independent E3 Tier-A replication remains blocked on new qualifying data. ₁

The roadmap stages these follow-ups by horizon rather than reporting them as done: near-term (6–12 months) feasible external-data extensions — INPHARED plus IMG/VR queries for the 3 GAP species, an AIEC strain-resolution diagnostic, and M. gnavus qPCR validation; mid-term (12–24 months) additional cohort and assay generation — a targeted qPCR ecotype panel, a per-patient bile-acid panel, multi-cohort serology meta-analysis, and longitudinal sampling expanded beyond patient 6967; and long-term (24+ months) clinical-pilot territory — a hybrid three-strategy cocktail clinical pilot, lytic-locked phage engineering for M. gnavus, and GAG-degrading enzyme inhibitor screening for H. hathewayi. ₁

On the multi-omics side, NB07d ran 2-modality CCA on HMP2 because pathway data is not paired with HMP2 metabolomics, whereas the cMD_IBD pathway slice in the mart is paired with cMD_IBD species at sample level. The report proposes on that basis a “3-modality MOFA” on cMD_IBD subjects giving “species + pathway + (cMD_IBD-only) joint factors”, but it names only two modalities — species and pathway — and leaves the putative third unspecified, so this is recorded as the report’s own proposal rather than evidence that three paired modalities are available. ₁

On the metabolomics side, NB09d’s finding that cohort batch dominates m/z-bridge clustering is recorded as Novel Contribution #19, and ComBat/SVA/RUV harmonization prior to PCA plus K-means is proposed to reveal whether metabolite-feature ecotypes are more cross-cohort-portable after batch correction — a test not run here. The same bridge fails on polyamines through the low-mass HILIC-positive gap, so the HMP2 polyamine OR = 14.6 stands as a single-cohort finding. ₁

Literature grounding and external context

The clustering choices are grounded in external method literature: LDA on pseudo-counts is reported to yield equivalent assignments at this sample scale (Ding & Schloss 2014), which the project says it confirmed by cross-method ARI against an independent CLR-based GMM, and the 48.9 % between-method agreement of the K = 4 assignments is read as consistent with a real microbiome continuum that the framework discretizes for operational purposes rather than as a defect. ₁

For differential abundance the report cites Gloor et al. (2017), Lin & Peddada (2020) and Tsilimigras & Fodor (2016) as establishing that raw Mann-Whitney on relative abundance is systematically biased by the sum-to-constant constraint, and Zhou et al. (2022) for LinDA — linear CLR regression with bias correction — as the simple pure-Python-implementable second-method concordance check it uses; NB00 reproduces the compositional-bias result directly on a curated battery. ₁

The AIEC compound glosses are literature-anchored rather than project-measured: Yersiniabactin is the iron-capture siderophore AIEC LF82 uses to survive in macrophage phagolysosomes (Prudent 2021, Dalmasso 2021, Dogan 2014), Colibactin is the pks+ genotoxin implicated in AIEC-associated colorectal carcinogenesis (Veziant 2016), and Enterobactin is a further AIEC siderophore active in mucosal-associated IBD settings; what the project itself contributes alongside them is the E. coli MIBiG matches to Yersiniabactin and Enterobactin as siderophores. ₁

The iron axis is given external causal context rather than project-internal causal evidence: Ellermann 2020 (PMID 31179826) demonstrated that dietary iron variably modulates intestinal microbiota assembly in colitis-resistant and colitis-susceptible mice — mouse-model evidence, not a uniform human effect — and the report notes that oral iron supplementation can exacerbate disease activity in a subset of IBD patients, not in all of them. Its therapeutic reading is that per-patient iron-supplementation decisions should account for ecotype-specific pathobiont iron-system expression and that phage targeting could specifically deplete the iron-system-enriched AIEC subset without affecting iron-pathway-deficient commensals — proposals, not demonstrated interventions. ₁

Several CD-associated organisms carry literature glosses with no project measurement behind them: E. innocuum is documented as a vancomycin-resistant pathobiome in IBD with clinically consequential phenotypes — creeping-fat formation and intestinal strictures in CD, and reduced UC remission rates — with faecal microbiota transplantation under evaluation as an intervention (Le 2025, PMID 40074633); the NB05 S. salivarius Salivaricin MIBiG hits are lantibiotic bacteriocins whose proimmune activity regulates the oral microbiome (Barbour 2023); and S. salivarius is separately reported as a potential rather than established salivary biomarker for orofacial granulomatosis co-occurring with CD (Goel 2019). The project’s own observation alongside these is that the top E1 Tier-A candidates S. salivarius, S. thermophilus and S. parasanguinis are all oral streptococci whose within-substudy CD-enrichment is consistent with that framework. ₁

Two further readings of the E1 Tier-A list are interpretive rather than measured. The report groups ranks 6–10 — Enterocloster asparagiformis, Intestinibacter bartlettii, Hungatella symbiosa, Erysipelatoclostridium ramosum, together with the related clostridial reclassifications E. bolteae, E. clostridioformis and E. citroniae in the lower-ranked list — as Lachnospiraceae / Erysipelotrichaceae expansion species characteristic of the Bacteroides-2 dysbiosis subtype (Vandeputte 2017) that defines E1. And it flags Gordonibacter pamelaeae’s CD-up signal in E1 as difficult to interpret as pathobiont activity, because the cited literature has it producing urolithins from dietary ellagitannins (Selma 2014, PMID 24744017) and increasing during microbiome recovery from a dysbiosis-inducing insult (Tierney 2023, PMID 36840551). ₁

Phage precedent and intervention context likewise come from outside the project: Galtier 2017 (PMID 28130329) demonstrated that AIEC-specific bacteriophages targeting CEACAM6-bound E. coli can reduce AIEC-associated intestinal mucosal signal in CD models, and Sheikh 2024 (PMID 38532703) showed that CD-patient microbiome transplanted into germ-free mice produces colitis with a discontinued-pattern, proximal colonic localization — background mechanistic context that carries no phage-efficacy bearing for this project’s targets. Combining the AIEC-phage result with the NB06 finding that 5 of 6 actionable Tier-A candidates co-occur in a single module, the report argues for multi-target phage cocktails rather than monovalent AIEC-only treatment on the reasoning that the pathobiont module would likely re-equilibrate around a remaining hub if only one species were targeted — an inference, not a tested depletion. ₁

The clustering and enterotype references are Holmes, Harris and Quince (2012), “Dirichlet multinomial mixtures: generative models for microbial metagenomics”, PLoS One 7(2):e30126, PMID 22319561, and Costea et al. (2018), “Enterotypes in the landscape of gut microbial community composition”, Nat Microbiol 3(1):8–16, PMID 29255284. ₁

The compositional-data references are Gloor et al. (2017), “Microbiome datasets are compositional: and this is not optional”, Front Microbiol 8:2224, PMID 29187837, which identifies microbiome datasets as compositional; Lin and Peddada (2020), “Analysis of compositions of microbiomes with bias correction”, Nat Commun 11(1):3514, PMID 32665548; and Tsilimigras and Fodor (2016), a review of the fundamentals, tools and challenges of compositional microbiome analysis, Ann Epidemiol 26(5):330–335, PMID 27255738. ₁

Oral-microbiome context is carried by four citations with no project measurement behind them: Xiang et al. (2024) on the involvement of oral bacteria in inflammatory bowel disease (Gastroenterol Rep (Oxf) 12:goae076, PMID 39188957); Guo et al. (2024), reporting that oral pathobiont-derived metabolites promote IBD (Gut Microbes 16(1):2333463, PMID 38545880); Tanwar et al. (2023) on the periodontitis–inflammatory-bowel-disease link, challenges and outlook, which specifies no causal result (Cells 12, PMID 37645044); and Goel et al. (2019), proposing Streptococcus salivarius as a potential — not established — salivary biomarker for orofacial granulomatosis and Crohn’s disease (Inflamm Bowel Dis, PMID 30796823). ₁

Two intervention-side citations are bibliographic rather than load-bearing: Le et al. (2025) is a pilot study of fecal microbiota transplantation for vancomycin-resistant Clostridium innocuum infection in inflammatory bowel disease evaluating safety and clinical and microbiota outcome, and the citation alone does not establish safety or efficacy (J Microbiol Immunol Infect, PMID 40074633); and Tierney et al. (2023) examines the capacity of a microbial synbiotic to rescue the in vitro metabolic activity of the gut microbiome following perturbation with alcohol or antibiotics, without a quantified effect in the citation (Appl Environ Microbiol 89(3):e01880-22, PMID 36840551). ₁

The module-anchor commensals carry their own external literature: Geirnaert et al. (2015) report interindividual differences in response to treatment with butyrate-producing Butyricicoccus pullicaecorum 25-3T in an in vitro gut model (FEMS Microbiol Ecol 91(6):fiv054, PMID 25999470); Steppe et al. (2014) is a safety assessment of the same strain as a potential probiotic for patients with inflammatory bowel disease, with no safety outcome stated in the citation (Food Chem Toxicol 72:129–137, PMID 25007784); Jeraldo et al. (2016) reconstructed the genome of a novel butyrate-producing clostridial scavenger from metagenomic sequence data (Front Microbiol 7:783, PMID 27303377); and Selma et al. (2014) describe Gordonibacter urolithinfaciens as a urolithin-producing bacterium isolated from the human gut (Int J Syst Evol Microbiol 64:2346–2352, PMID 24744017). ₁

The AIEC and iron citations carry PMIDs and, in two cases, scopes the records themselves flag: Dalmasso et al. (2021) on the Yersiniabactin siderophore of Crohn’s disease-associated adherent-invasive Escherichia coli (Int J Mol Sci 22(7):3512, PMID 33805299); Prudent et al. (2021), reporting that the Crohn’s disease-related bacterial strain LF82 assembles biofilm-like communities to protect itself from phagolysosomal attack (Commun Biol 4(1), PMID 34035436) — a cited single-strain phenotype used as background; Dogan et al. (2014) on enrichment of propanediol-use, iron-use and M-cell-translocation pathways in inflammation-associated adherent-invasive E. coli (Inflamm Bowel Dis 21(1):92–111, PMID 25230163); Dubinsky et al. (2022) on disease-specific genomic adaptations in E. coli strains from patients with inflammatory bowel diseases (J Crohns Colitis 16(9):1484–1497, PMID 35560165); Veziant et al. (2016) on the association of colorectal cancer with pathogenic E. coli, a disease context distinct from Crohn’s disease (World J Clin Oncol 7(3):293–302, PMID 27298769); and Buret et al. (2019), whose title raises rather than settles a possible role for iron in pathobiont release from dysbiotic gut microbiota biofilms in intestinal inflammatory diseases (J Biomed Sci 26(1):1, PMID 30602371). ₁

The quaternary-amine and polyamine citations are recorded with their own limits: Timsina, Gora and Ferguson (2025), “Proteomic and metabolomic analysis reveals new insights into quaternary amine metabolism” (mSphere 10(2):e00421-25, PMID 40853000), an entry that gives no specific quantitative result; Cheng, Hung and Lin (2025) on the association of trimethylamine N-oxide and metabolites with kidney function decline in patients with chronic kidney disease, a context distinct from inflammatory bowel disease (Clin Nutr 44:18–25, PMID 39709651); Jiang et al. (2024), a mouse study reporting that polyphenols from hickory nut reduce the occurrence of atherosclerosis while improving intestinal microbiota and inhibiting trimethylamine N-oxide production (Phytomedicine 130:155349, PMID 38522315); and Pegg (2014), “Toxicity of polyamines and their metabolic products” (Pharmacol Res 84:80-89, PMID 24747196), cited as a review of polyamine biology plus IBD. ₁

Two butyrate-producer characterizations stand behind the module-anchor reading of A. caccae, both external rather than project measurements: Schwiertz et al. (2002) describe Anaerostipes caccae as a new saccharolytic, acetate-utilising, butyrate-producing bacterium from human faeces (Syst Appl Microbiol 25(1):46–51, PMID 12086188), and Duncan et al. (2004) report lactate-utilizing bacteria, isolated from human feces, that produce butyrate as a major fermentation product (Appl Environ Microbiol 70(10):5810–5817, PMID 15466518). ₁

The reference list also identifies Elmassry et al. (2025), “Diet, lifestyle, and the microbiome interaction reveals fatty acid amide functional signatures of inflammatory bowel disease” (Cell Host & Microbe 33(2):243–256, PMID 39837311), as the BGC catalog source and the origin of the ebf/ecf RPKM CD-versus-HC finding, which the report says NB08a §4 replicated independently; the citation entry itself carries no effect size. ₁

Three further anchors are cited rather than measured here: Henke et al. (2019) report that Ruminococcus gnavus, a member of the human gut microbiome associated with Crohn’s disease, produces an inflammatory polysaccharide (Proc Natl Acad Sci USA 116(26):12672–12677, PMID 31182571); Darfeuille-Michaud et al. (2004) report a high prevalence of adherent-invasive Escherichia coli associated with ileal mucosa in Crohn’s disease (Gastroenterology 127(2):412–421, PMID 15300573); and Vujkovic-Cvijin et al. (2020) warn that host variables confound gut-microbiota studies of human disease (Nature 587(7834):448–454, PMID 33149306). ₁

Two specific reference entries are pointers rather than full citations: reference 11, the Elmassry et al. (2025) gut-microbiome BGC catalog in Cell Host & Microbe, is referenced through ref_bgc_catalog metadata with its full citation stated to live in dim_studies — the report does give a full Elmassry 2025 citation separately, as recorded just above — and reference 12, the Kumbhari et al. (2024) strain-frequency / IBD-adaptation source, comes from the ref_kumbhari_s7_* supplementary with its full citation likewise stated to live in dim_studies. ₁

The second compositional differential-abundance method is cited in full as Zhou, He, Chen and Zhang (2022), “LinDA: linear models for differential abundance analysis of microbiome compositional data”, Genome Biology 23(1):95, PMID 35421994, used in NB04c §4 alongside CLR plus Mann-Whitney and implemented in pure NumPy for this project. ₁

The metabolite-side narratives carry their own external anchors: Wang et al. (2018) on plasma polyamine signatures in patients with inflammatory bowel disease (Inflamm Bowel Dis 24(2):324–334, PMID 29361096), which the report describes as polyamine CD biomarker confirmation; Franzosa et al. (2019) on gut-microbiome structure and metabolic activity in inflammatory bowel disease (Nat Microbiol 4(2):293–305, PMID 30531976), which the report links through HMP2 metabolomics plus microbiome integration to a canonical bile-acid 7α-dehydroxylation deficit and polyamine and lipid signatures; Hall et al. (2024) identifying bilirubin reductase as a gut-bacterial enzyme that produces urobilinogen and connects intestinal microbiota to host bilirubin metabolism (Nature); Vital et al. (2018) on the diversity of bacteria exhibiting bilirubin-reducing activity in the human intestinal microbiota, the urobilinoid-producing commensals (Front Microbiol 9:486, PMID 29615976); and Devlin and Fischbach (2015) on a biosynthetic pathway for a prominent class of microbiota-derived bile acids (Nat Chem Biol 11(9):685–690, PMID 26412091), whose bai-operon 7α-dehydroxylation mechanism in Clostridium scindens-like bacteria the report takes as the anchor for its NB09c bile-acid network finding. ₁

The report’s author list includes Adam Arkin (ORCID 0000-0002-4999-2931), U.C. Berkeley / Lawrence Berkeley National Laboratory. ₁

Committed figure set

The committed figure set records the project’s data basis: NB00_cohort_summary.png gives sample counts per study × data type and participants by diagnosis; NB00_missingness_heatmap.png the per-column null fraction per table; NB00_ucdavis_cohort.png UC Davis calprotectin by Montreal location and medication class; and NB00_protective_species_paradox.png raw log₂FC versus CLR Δ on 14 curated protective plus pathobiont species. ₁

NB01b_K_selection.png plots held-out perplexity, cross-method ARI and GMM BIC across K = 2..8; NB01b_ecotype_species_heatmap.png the top species × ecotype mean relative abundance; NB01b_ecotype_by_diagnosis.png the stacked ecotype distribution per diagnosis class; and NB02_ucdavis_ecotype_assignment.png the per-UC-Davis-patient ecotype call ordered by calprotectin, carried inline under the caption of UC Davis per-patient ecotype calls and reference comparison. ₁

NB02_ucdavis_ecotype_x_clinical.png crosses UC Davis ecotype with Montreal location and with medication class, while the two proxy-classifier figures are NB03_feature_importance.png, giving classifier feature importance for the minimal versus extended covariate sets, and NB03_h1c_auc.png, plotting the one-vs-rest AUCs per ecotype against a 0.70 threshold. ₁

Two superseded NB04 figures are kept for audit: NB04_h2c_paradox_resolution.png, the 14-curated-species × pooled raw / pooled CLR / E1 CLR / E3 CLR heatmap, and NB04_top_tier_a_per_ecotype.png, the union of the top-15 within-ecotype Tier-A candidates; NB04b_jaccard_null.png shows the permutation null and observed Jaccard for H2b (200 perms; empirical p = 0.000). ₁

NB04f_loso_stability.png shows per-substudy leave-one-substudy-out ARI with per-sample agreement bars as honest cross-study variation; NB04g_pathway_vs_taxon_ecotype.png the pathway-ecotype × taxon-ecotype cross-table heatmap with per-ecotype agreement bars; and NB04h_hmp2_external_replication.png the HMP2 external replication as an ecotype × diagnosis stacked bar, a projection-confidence histogram and an E1 Tier-A effect scatter. ₁

NB05_tier_a_scored.png presents the top-30 A3–A6 scoring matrix heatmap with a total-score bar chart marking the 2.5 actionable threshold, and NB06_cooccurrence_networks.png the per-subnet spring-layout networks for E1_all, E1_CD, E3_all and E3_CD with the actionable Tier-A species highlighted. ₁

Four figures carry the pathway-mechanism layer: NB07a_pathway_DA_h3a.png, the three-clause falsifiability dashboard of per-pathway CLR-Δ histogram, permutation null and top pathway-pathobiont attribution scatter; NB07b_stratified_pathway_DA.png, the per-Tier-A-core species stratified pathway effect heatmap with per-species CD-up and CD-down counts; NB07_H3a_v18_class_enrichment.png, the v1.8 cohort-level Fisher enrichment bar chart with iron/heme dominant plus the per-species per-theme heatmap; and NB07c_anchor_pathobiont_coupling.png, the anchor × pathobiont species ρ heatmap with E1_CD in the top half and E3_CD in the bottom half beside the mean iron-pathway co-variation per pathobiont. ₁

The figure index carries its own captions for the Pillar-4 and Pillar-5 figures, which stand alongside rather than replace the panel descriptions given in the prose above: NB12_phage_targetability_matrix.png as a per-Tier-A-core × phage-availability-score heatmap plus the 4-class stratification of clinical-trial, lytic-literature, temperate-only and GAP, with a therapeutic-targets cross-reference for alternative strategies; NB13_phage_cocktail.png — indexed separately from the NB13_phagefoundry_cocktail.png figure described above — as a 3-panel cumulative strain-coverage curve with the 5-phage cocktail at 94.7 %, the per-phage isolate-host phylogroup distribution with AIEC B2/D dominant, and the cocktail-composition strain-coverage heatmap; and NB14_phageome_ecotype.png as a 3-panel per-phage-family CD-vs-HC by ecotype with the E1 Gokushovirus result (cliff = −0.358, FDR = 5e-7) highlighted, the Tier-A pathobiont × phage-family endogenous correlation heatmap (E. coli × Podo/Myo +0.18/+0.13), and the family-classification gap of 80 % “Unknown”. ₁

NB17_synthesis.png closes the set in three panels: a target decision-matrix heatmap whose index caption reads 6 actionable Tier-A × 4 attributes while the accompanying NB17 prose reads 6 actionable Tier-A × 5 attributes (NB05 score, ecotype membership, BA-coupling cost, mediation mode, phage tier), a disagreement in attribute count the report does not reconcile; a UC Davis per-patient cocktail-design map of 4 design categories × 7 cocktail strategies with the patient count in each cell; and a 4-phase clinical-translation roadmap timeline spanning immediate, near, mid and long term. ₁

The figure index also lists additional supporting files — NB00_cohort_summary.png, NB00_missingness_heatmap.png and NB01_*.png — with the legacy K = 8 fit preserved for methodology audit. ₁

Caveats

The ecotype framework is operationally useful but not uniformly reproducible: LOSO ARI averaged 0.113, the E3 Tier-A list came from a single eligible cMD study, E0 and E2 lacked viable within-ecotype CD-versus-nonIBD contrasts, and hard ecotype assignments discretize an underlying continuum (Costea 2018), with soft probabilities preserved in data/ecotype_assignments.tsv so the hard calls are to be treated as operational labels rather than biology. The report’s own summary of this is that Pillar 2 cross-cohort replication is weak: single-study E3 evidence combined with marginal ecotype stability means the Tier-A list, particularly for E3, should be treated as hypotheses for further experimental validation rather than established targets, and the Pillar 5 UC Davis cocktail drafts inherit that limitation, with per-candidate cross-cohort support to be annotated explicitly. The UC Davis Kuehl projection onto the K = 4 framework accordingly used LDA as primary and GMM as advisory: the Kaiju-to-MetaPhlAn3 projection showed method asymmetry, with LDA more robust, whereas GMM assigned all 26 UC Davis samples to E3 with confidence >0.97 under sparse feature overlap, an interpretation judged artifactual. ₁

The project used a partial multi-method differential-abundance consensus because LinDA was implemented in pure Python and ANCOM-BC/MaAsLin2 were not completed; the plan had called for ≥ 2 / 3 of {ANCOM-BC, MaAsLin2, LinDA} to agree, and the report calls the bootstrap-CI and within-substudy-meta outputs additional independent evidence streams that are not drop-in ANCOM-BC / MaAsLin2 substitutes, leaving a full three-method R-native consensus as a publication-grade follow-up. That independence framing sits against the report’s own statement, recorded above, that bootstrap-stable within-ecotype differential abundance and LinDA run on the same ecotype-defined subsamples and so share the selection-on-outcome bias, leaving only the within-substudy CD-versus-nonIBD contrast as an independent evidence source; the report does not reconcile the two framings. The NB08a species × BGC interaction test was not performed — the analysis instead used the precomputed ref_cborf_enrichment table, which carries a CD-versus-HC main effect rather than a species × BGC interaction, so the per-species CB-ORF enrichment is consistent with H3c but cannot formally distinguish a species × BGC interaction from a species main effect plus a BGC main effect, and the interaction test was dropped under plan v1.9 as a structural project-scope limitation rather than a deferred follow-up. Its background comparator is the full BGC catalog of 10,060 BGCs across all species rather than a matched comparator, so the iron_siderophore Fisher OR of 44× is partly inflated by the catalog-wide rarity of iron MIBiG matches (51 / 9,774 = 0.5 %) and should not be treated as a precise effect size — a more conservative comparator would be a matched-niche pathobiont set of gut Proteobacteria plus CD-associated Firmicutes at similar genome-assembly depth — while the qualitative finding that E. coli alone among the actionable Tier-A core carries the iron-siderophore signature is described as robust across choice of comparator; raw-read strain-resolution analyses were dropped under the no-raw-reads scope; AIEC-specific targeting therefore depends on a proposed kbase_genomes pks plus iron-BGC genome-content query to identify Yersiniabactin/Enterobactin/Colibactin-positive strains rather than on strain resolution from curatedMetagenomicData raw reads; plan v1.9 dropped the raw-reads-based cMD strain-resolution test as not feasible, leaving a genome-content survey — a kbase_genomes pks plus Yersiniabactin plus Enterobactin co-occurrence query cross-referenced with the Kumbhari fact_strain_competition strain-frequency data — as the no-raw-reads alternative. ₁

The strain-adaptation result rests on one cohort (Kumbhari/LSS-PRISM), whose cross-cohort sensitivity test was acknowledged-weak by design, and about 95 % of FDR-passing genes are unclassified by the gene-symbol regex of the curated 13-category classifier, so the functional-enrichment signal belongs to the category-classified subset rather than to the full gene set. The planned cross-cohort species-level sensitivity test on curatedMetagenomicData was specified as sensitivity-only and then dropped unexecuted because raw reads were unavailable. ₁

The serology analysis is structurally low-powered rather than conclusive: n = 67 across 3 unevenly distributed sites is, in the report’s terms, structurally low-power for |r| > 0.40 detection at α = 0.05 with correction for 48 tests, and the plan acknowledged this test as weak by design. Subject-level aggregation discards the multiple serology and metagenomics visits each subject contributed, and while mixed-effects regression would have more power, unbalanced visit counts complicate interpretation; OmpC and ANCA also cross-react with related antigens beyond their named targets, so per-species attributions are correlative rather than exclusive. The single-cohort caveat is structural — there is no project-internal replication path, and multi-cohort meta-analysis would be needed to firm the signals up. ₁

The metabolomics ecotype clustering result is specifically limited by uncorrected batch effects: pooled HMP2 and FRANZOSA m/z-bridge clustering separated cohorts, with PC1 explaining 79 % of variance and cross-cohort LOSO ARI = 0.000, below the taxonomic baseline of 0.113. The 3-modality MOFA+ plan was reduced to a 2-modality CCA pilot because HMP2 pathway abundance was unavailable in the mart, and the polyamine signature remains single-cohort because no m/z bridge was found. The m/z-bridge feature space is also small at 111 unique features; the report notes that a larger feature space of 500-1000 metabolites with name annotations across cohorts might give different results, but that this requires external Franzosa supplementary annotations placed out of scope by the plan. ₁

PhageFoundry coverage applies to its 188 tested E. coli strains, not UC Davis patient isolates — the report states that coverage as 95 % of the 188 PhageFoundry-tested strains, a rounded label distinct from the 94.7 % its set-cover table reaches, and per-patient applicability requires an AIEC strain-resolution diagnostic on UC Davis patient stool isolates, queued in the near-term roadmap — and the dataset lacks explicit AIEC-versus-commensal labels, burst-size measurements, titer data, and in-vivo delivery validation. Those 188 strains do include canonical AIEC strains named as phage isolation hosts — LF82, LF73, 536, NIC06 and BCH953 — but accurate AIEC coverage estimation would require cross-reference to literature isolate annotations, and the cocktail covers broad E. coli strain diversity without by itself distinguishing AIEC from commensal E. coli, so the report makes a strain-level diagnostic (per-patient pks-island / Yersiniabactin / Enterobactin gene-presence detection) a requirement before cocktail selection. The susceptibility matrix is also binary, with no titer or burst size, while real-world cocktail dosing depends on phage burst size plus receptor-binding kinetics plus gastric-passage survival — all out of project scope. HMP2 viromics had an 80 % family-classification Unknown fraction, so the Tier-A × phage-family correlations — given in that caveat as E. coli × Podoviridae +0.18 and × Myoviridae +0.13 — are restricted to the classifiable 20 %, and improving family resolution would require re-running VirMAP/MARVEL against newer phage reference databases, which the report places out of scope; endogenous phage correlations were modest: the report states a maximum absolute ρ ≤ 0.18, a bound inconsistent with its own tabulated E. coli × Podoviridae ρ = +0.183 and left unreconciled, so it should be read as the report’s statement rather than an exact ceiling, with the substantive point being that no correlation reached its |ρ| > 0.30 strong-candidate threshold. The three gut-anaerobe coverage gaps — H. hathewayi, M. gnavus lytic alternatives and F. plautii — remain unresolved pending external-database queries of INPHARED (~25K phages) and IMG/VR (~3M UViGs), and those queries, multi-cohort prospective validation of the state-dependent dosing rule derived from the n = 1 patient 6967 trajectory, and a per-patient AIEC strain-resolution diagnostic are all flagged as required clinical-translation prerequisites in the NB17 roadmap’s near-term tier, out of scope here rather than performed. ₁

The UC Davis per-patient framework is based on only 23 patients, so its outputs are exemplars and templates rather than statistically robust per-patient validation and generalization requires multi-center validation on external CD cohorts; it has no patient-specific bile-acid measurements — the bile-acid coupling cost is an ecotype-level, not a per-patient, quantity — no AIEC strain-resolution diagnostics, and it uses Kaiju-derived target presence, which the report flags as lower-confidence than the MetaPhlAn3-based curatedMetagenomicData analyses given the NB02 classifier-mismatch asymmetry. Patient 6967 is the only biological-replicate longitudinal trajectory, so the E1→E3 drift, cocktail Jaccard = 0.60, 3–6-month reassessment interval, five-fold qPCR trigger, and state-dependent dosing rules are hypotheses rather than clinically validated rules. There is also no patient-specific AIEC strain-resolution diagnostic for the 8 E. coli-positive patients, so the cocktail recommendation assumes an AIEC subset prevalence taken from Dogan 2014 and Dubinsky 2022, with per-patient AIEC genotyping (pks-island plus Yersiniabactin plus Enterobactin) recorded as a near-term clinical-translation prerequisite; and the F. plautii deprioritization is precautionary, based on the NB09c mechanism, with clinical validation requiring per-patient bile-acid panels. The cocktail-strategy distribution on this cohort — 12 reserve, 4 E1 hybrid, 4 E0 limited, 1 E1 full, 1 E3 and 1 state-dependent — is the operational stratification the report says should be revisited on a larger CD cohort. ₁

The longitudinal and dosing conclusions rest on a single trajectory: patient 6967 is the only multi-timepoint patient with biological-replicate samples in the UC Davis cohort, so the ecotype-drift conclusions are based on an n = 1 longitudinal trajectory. Both ecotype calls are moderate-confidence (visit 1 = 0.64, visit 2 = 0.41), so the label shift could partly reflect classifier uncertainty, although the 14× M. gnavus and 3× E. lenta expansions are quantitative biological observations independent of the ecotype label. Patient 1112’s technical-replicate Spearman ρ = 1.000 validates the technical-noise floor for this longitudinal contrast but is not itself a biological-replicate longitudinal trajectory, so patient 6967 remains the only such trajectory in the cohort. There is no timing information between the two visits, so the duration of the E1→E3 drift is unknown, which limits clinical-workflow timing recommendations, the 3-6 month re-test interval being a clinical convention default rather than a value derived from these data; the state-dependent dosing rule is theoretical and not yet validated in clinical practice; and the cocktail Jaccard of 0.60 is moderate, with whether a 3-of-5 component overlap implies clinically meaningful cocktail re-design depending on patient-response heterogeneity, which is out of project scope. ₁

Slots Into

• condition-specific-fitness — Ecotype-specific Tier-A targets, condition-dependent pathobiont modules, and state-dependent cocktail dosing connect microbial fitness to CD ecological states. ₁
• multi-omics-integration — The CCA pilot integrates taxonomy and metabolomics into CC1, with r = 0.964 and cliff CD-versus-nonIBD = +0.498, while exposing pathway–metabolite pool-versus-flux differences. ₁
• gene-function-acquisition-depth — Iron/heme pathway enrichment, E. coli BGC content, strain-adaptation genes, and bile-acid metabolic signatures distinguish species-abundance-mediated from strain-content-mediated mechanisms. ₁
• pangenome-integration — PhageFoundry strain susceptibility, AIEC phylogroup coverage, and the proposed pks/Yersiniabactin/Enterobactin diagnostic connect strain-level genome content to cocktail selection. ₁
• cross-tenant-data-bridging — The report demonstrates cross-cohort and cross-dataset integration across curatedMetagenomicData, HMP2, FRANZOSA, PhageFoundry, viromics, and UC Davis profiles, while documenting taxonomy synonymy and batch limitations. ₁
• phage-therapy-evidence-translation — The three-layer evidence stack, the five-phage cocktail at 94.7 % coverage (short of the ≥ 95 % target it was selected for, though the report also calls it approximately 95 %), the coverage gaps at Tier-A scores 4.0 and 3.3, and the out-of-scope prospective-validation prerequisites set the distance between database coverage and clinical claim. ₁
• ecotype-clustering-validity — K = 4 chosen by cross-method ARI, LOSO ARI = 0.113, the E1→E3 patient drift, and the 26-sample/23-patient projection basis test how far discrete ecotypes survive as strata. ₁
• cross-cohort-microbiome-portability — HMP_2019_ibdmdb replication (1,627 samples, χ² = 15.61, 45/51 sign concordance) alongside the single-study E3 list and the internally contradictory S. thermophilus replication statements. ₁
• competitive-exclusion-consortium-design — Six actionable targets from 71 candidates, the donor-2708 engraftment pathobionts, and the 57–84-node pathobiont module define multi-target intervention units. ₁
• species-cooccurrence-modules-as-intervention-units — Every subnet places 4–5 of 6 actionable targets in one module, with ecotype-specific membership for F. plautii and E. coli. ₁
• community-metabolic-interdependence — The SCFA null, the 7 strict cross-feeding triangles, and the opposite lactate associations withdraw support from the A. caccae cross-feeding interpretation. ₁
• metabolite-class-signatures-of-gut-inflammation — Polyamine (OR = 14.6) and long-chain PUFA (OR = 7.9) enrichment against unsupported themes including acyl-carnitines, bile acids, SCFAs, tma_choline and the zero-member tryptophan/indole, fatty-acid-amide, aromatic-amino-acid and oxidative themes define the CD metabolite-class signature here. ₁
• occurrence-versus-catabolic-activity — The polyamine metabolite pool inverts its pathway-level theme (OR 0.42 vs 14.6), and urobilin and tauro-form shifts index activity rather than presence. ₁
• callability-limited-comparative-inference — K. oxytoca below the prevalence filter, the untested species × BGC interaction, and the unmeasured iron/heme metabolite axis each bound what could be compared. ₁
• taxonomic-nomenclature-reconciliation — 2,417 aliases mapped to 1,848 canonical species via NCBI taxid and GTDB r214+ renames made pooling and projection possible. ₁
• compositional-robustness-of-differential-abundance-calls — Raw Mann-Whitney mis-directed 4+ of 8 protective species that CLR correction recovered. ₁
• adversarial-research-quality-assurance — Three self-retractions (C. scindens, the Jaccard framing, the 33-species list) and an unreconciled internal contradiction about S. thermophilus replication. ₁
• selection-on-outcome-leakage — Clustering on taxon abundances and then testing the same taxa within cluster, compounded by substudy–diagnosis nesting, produced the spurious C. scindens null. ₁
• classifier-database-compatibility-in-taxonomic-quantification — Kaiju-to-MetaPhlAn3 projection asymmetry and the ρ = 1.000 resequencing replicate bound classifier-driven target calls. ₁
• cross-condition-metabolic-comparability — 579 testable of 592 named metabolites, unmeasured siderophores, and LC-MS SCFA under-sampling limit which conditions can be compared. ₁
• ontology-and-category-schema-sensitivity — Acyl-carnitine and lipid-class themes stayed non-significant (OR 1.06, 1.64) because the chemical-class backgrounds diluted real member-level shifts. ₁
• taxonomic-resolution-dependent-functional-inference — ebf/ecf RPKM without per-species attribution can monitor response but cannot select targets. ₁
• pathobiont-biosynthetic-gene-cluster-repertoires — E. coli’s 146 BGCs including 54 iron and 25 genotoxin against zeros in five co-targets anchor the BGC-repertoire contrast. ₁
• iron acquisition and pathobiont expansion — Five converging evidence streams place iron acquisition at the centre of the E. coli AIEC mechanism narrative. ₁
• gut-microbiome-ecotypes-as-patient-strata — Four ecotypes trained on 8,489 samples and projected onto 26 samples from 23 patients function as the stratification layer for cocktail design. ₁
• gut virome disease association — Gokushovirus WZ-2015a CD-down in E1 and modest E. coli–Podoviridae/Myoviridae correlations describe the endogenous phageome signal. ₁
• ecological-cost-of-microbiome-target-depletion — The bile-acid coupling cost of removing F. plautii, E. lenta, and E. bolteae is the explicit cost term in target selection. ₁
• proxy-assays-for-community-state-assignment — The clinical-covariate classifier’s 41 % agreement and the proposed M. gnavus qPCR trigger are both proxies for metagenomic state calls. ₁
• adaptive-versus-housekeeping-functional-differentiation — IBD-biased strain genes were enriched for adaptation (OR = 1.38) and depleted for housekeeping (OR = 0.62) across 59 commensal species. ₁
• carriage-prevalence-versus-within-carrier-function — E. lenta’s 54 % CD-down CB-ORFs are read as carriage-prevalence-driven rather than within-carrier abundance-shifted. ₁
• genomic-under-representation — Five Tier-A core species sit in MIBiG dark matter, carrying substantial BGC content with no annotated iron or genotoxin clusters. ₁
• confirmatory-exploratory-ecological-association-discordance — The serology H3e test fails its strict |r| > 0.40 plus FDR < 0.10 threshold while single sites reach |r| ≈ 0.40–0.46. ₁
• study-batch-confounding-of-environmental-associations — Serology associations concentrate in 1 of 2 productive sites and dilute on cohort aggregation. ₁
• functional-marker-validation — Serology serves as a CD-versus-UC stratifier but not a per-target abundance predictor, with OmpC/ANCA cross-reactivity blocking exclusive attribution. ₁
• pooled-run-pseudoreplication-and-metadata-label-noise — Subject-level aggregation of multi-visit serology and metagenomics discards longitudinal structure under unbalanced visit counts. ₁
• genome-wide-versus-locus-specific-ecological-adaptation — F. plautii’s 0 FDR-passing strain-adaptation genes among 3,245 tested are read by the report as placing its CD association at species abundance rather than strain gene content, an interpretation of an informative null. ₁
• cultivation-collection-bias-in-ecological-genomics — 65 of 94 PhageFoundry phages were isolated against B2/D hosts, so the tested susceptibility matrix inherits the isolation collection’s phylogroup skew. ₁
• evidence-triangulation-for-functional-annotation — The iron narrowing to E. coli moves from species × pathway correlation, where 8 of 15 iron pathways are non-specific menaquinol biosynthesis, to direct genomic BGC content, with the raw-read species-stratified attribution path dropped. ₁
• fitness-matched-null-models — The iron-siderophore OR of 44× rests on a full-catalog rather than matched-niche comparator, and per-species CB-ORF CD-up rates are read against a 2.5 % catalog background. ₁