species-cooccurrence-modules-as-intervention-units

A species-level co-occurrence module is the natural candidate unit for a multi-target microbiome intervention: if the species one wants to remove sit together in one correlated block, one cocktail addresses them as an ecological group rather than one by one. The IBD phage-targeting project tested this by building four per-subnet correlation networks via CLR transform (centered log-ratio, a transform applied to compositional abundance vectors before correlation) plus rank-based Pearson correlation (equivalent to Spearman rho), with per-edge BH-FDR (Benjamini-Hochberg false discovery rate control), thresholded at |rho| > 0.3 AND FDR < 0.05, then ran Louvain community detection (networkx.community.louvain_communities, edge-weighted by |rho|); Networkx 3.5’s built-in Louvain was sufficient and FastSpar / SpiecEasi installation was held back as unnecessary for the question given clear module structure at CLR-Spearman ₁.

Module structure across ecotype subnets

The four subnets differ in sample count, and the resulting networks differ in both density and module count ₁:

The headline test of “modules as intervention units” failed on its own stated metric but passed on inspection: the raw mean-actionable-per-module is 1.38 on E1_all + E3_all, below the >= 2 bar stated in the plan, yet the signal is not uniformly distributed across modules, because in every subnet a single module contains 4-5 of the 6 actionable Tier-A candidates ₁.

That concentration is what gives the module its role as a candidate unit, and its composition and size vary by subnet ₁:

This refines rather than rescues the mean-based criterion: the remaining modules per subnet are commensal / Prevotella / diverse-healthy communities that naturally contain 0 Tier-A hits by construction, so the mean-per-module statistic is diluted by these biologically-irrelevant-to-the-question modules ₁. The practical lesson is that a module-level enrichment claim should be evaluated on the module where the targets are expected, not as an average over modules that cannot contain them.

Membership is not stable across community states

The unit is not transferable between community states unchanged. F. plautii is in the main pathobiont module in E1 but in the generalist module in E3, which the report reads as meaning that for E1 patients F. plautii plus main-pathobiont co-targeting is ecologically coherent, while for E3 patients F. plautii is less linked and may need a separate phage; that cocktail implication is a prospective design inference, not a tested result ₁. See gut-microbiome-ecotypes-as-patient-strata for the subnet definitions this depends on.

E. coli is in the pathobiont module in E3 only, not E1, which the report interprets as consistent with adherent-invasive E. coli (AIEC) being more characteristic of severe-Bacteroides-expanded E3 than transitional E1 — an interpretation layered on a single cross-sectional co-occurrence analysis rather than an independently measured strain-level finding ₁.

Module anchors and the collateral-damage question

The non-target members of a pathobiont module are worth enumerating because of a risk hypothesis, not a demonstrated effect: co-occurrence and hub membership do not show that a phage cocktail aimed at module pathobionts would perturb their module neighbours, and nothing in this project tested that. Top-degree non-Tier-A hubs in the pathobiont modules differ between subnets: E1_all module 1 is anchored by Firmicutes bacterium CAG 110, Collinsella massiliensis and Phascolarctobacterium sp CAG 266, while E3_all module 1 is anchored by Butyricicoccus pullicaecorum, Anaerostipes caccae and Lactococcus lactis ₁.

The Crohn’s-disease-restricted subnets show the same pattern at smaller scale: E1_CD module 0 (75 nodes) has anchor commensals Clostridiales bacterium 1_7_47FAA, Anaerostipes caccae (the only genuine butyrate-producer among the module-anchor commensals) and Bacteroides nordii, together with the 5 actionable Tier-A pathobionts of that subnet (H. hathewayi, F. plautii, E. bolteae, E. lenta, M. gnavus); E3_CD module 1 (57 nodes) has anchor commensals Actinomyces sp. oral-taxon-181 and Actinomyces sp. HMSC035G02 (both oral cavity ectopic colonizers) plus Lactonifactor longoviformis (lactate utilizer), with 4 module pathobionts (E. lenta, H. hathewayi, E. coli, M. gnavus) ₁. The presence of a butyrate producer inside the E1_CD target module is the kind of adjacency that feeds ecological-cost-of-microbiome-target-depletion.

Evidence strength

All of the above rests on one project’s cross-sectional CLR-Spearman co-occurrence networks over metagenomic abundance profiles; correlation modules are not measured interaction modules, and no perturbation, isolate co-culture or longitudinal series tested whether removing one module member moves the others ₁. The reusable claim is therefore stated as a hypothesis: co-occurrence modules concentrate candidate targets strongly enough to be useful design scaffolds, but because membership switches with community state (F. plautii, E. coli), a module defined in one patient stratum should not be assumed to be the same intervention unit in another ₁. Translational use of these units is further constrained by the standards discussed in phage-therapy-evidence-translation.

Tensions

The criterion stated in the project’s plan (mean >= 2 actionable candidates per module) and the project’s post-hoc inspection (one module per subnet holding 4-5 of the 6 actionable Tier-A candidates) point in opposite directions on whether modules qualify as intervention units; the report keeps both rather than choosing, attributing the gap to modules that cannot contain targets by construction ₁. The available evidence does not record what threshold, if any, was set for the expected module considered alone, so which reading should govern cannot be settled from this corpus.

Open Directions

• Rebuild the four subnet networks under alternative edge estimators — FastSpar and SpiecEasi, which the project deliberately did not install, the latter targeting conditional independence rather than marginal correlation — on the same CLR matrices, and report whether the single-pathobiont-module concentration of 4-5 of the 6 Tier-A candidates is stable across estimators; this measures how far the module definition depends on the CLR-Spearman edge definition, without assuming any one estimator removes indirect association ₁.
• Define the module-enrichment criterion over target-containing modules only, with a permutation null that reassigns Tier-A labels across nodes, so the reported 1.38 mean and the per-subnet concentration can be judged against the same prespecified test rather than one before and one after inspection ₁.
• Test module stability directly by bootstrapping samples within each subnet and recording how often F. plautii and E. coli land in the pathobiont module versus the generalist module, which would say whether the E1/E3 membership switch is a community-state effect or a sampling artifact of the differing subnet sizes (n samples = 2,601 for E1_all, 1,364 for E3_all) ₁.
• Score each pathobiont module for collateral risk by annotating its anchor commensals (e.g. the butyrate producer Anaerostipes caccae in E1_CD module 0) for functions that would be lost if those taxa declined, giving a per-subnet risk list to pair with the per-subnet target list ₁.
• Repeat the module construction in a longitudinal cohort so that module co-membership can be checked against within-patient co-movement over time. This is a stability analysis only: temporal co-movement would strengthen the module-as-unit hypothesis but could not by itself establish that acting on the module produces the intended community change, which needs perturbation data ₁.