Answers are grounded in this paper and cite their sources.
Dietary intervention trials increasingly pair microbiome profiling with cardiometabolic phenotyping. This combination is promising, but it has created a recurring interpretive problem: the claim often travels farther than the measurement. A shift in relative taxon abundance is reported as a functional change; a predicted pathway is treated as active; a metabolite association is labelled microbial; or a randomized dietary effect is described as microbiome mediated. These statements are not interchangeable. Each requires evidence from a different biological layer.Our concern is not with any one assay. It is with treating their outputs as though they answer the same question. A taxonomic signal can be a useful discovery result without identifying a mechanism. Equally, the absence of a taxonomic shift does not rule out a functional response: different organisms may encode overlapping capabilities, and the same organism may alter its activity without changing in abundance [1,2]. When these layers are blurred, replication becomes harder to interpret and apparently inconsistent studies may not have tested the same claim at all. We therefore propose a five-level framework that links each measurement to the strongest claim it can directly support: composition, functional potential, realized function, host linkage, and causal attribution (Figure 1). The staircase is deliberate. It is not a checklist that every trial must complete; it is a reminder that the wording of the conclusion should stop where the evidence stops.The framework is intended to complement, not replace, existing guidance. Reporting checklists such as STORMS improve transparency and reproducibility, while formal causal-inference methods address specific estimands and assumptions. Our framework asks a preceding interpretive question: given the biological layer actually measured, what is the strongest claim that the data directly support? This claim-tomeasurement mapping is its main contribution and can be used alongside reporting standards, multi-omics integration, and causal analyses.The first level is composition: alpha diversity, community dissimilarity, taxa, strains, and, when measured, microbial load. Amplicon sequencing and taxonomic profiles from shotgun metagenomics can show that microbial ecology changed. They cannot, on their own, show that microbial function changed or that the microbiome produced a host effect. Relative-abundance data are compositional, so an apparent increase may reflect a decrease elsewhere; quantitative profiling can therefore change the interpretation [3]. The choice of differential-abundance method also matters [4]. Reporting standards such as STORMS are useful because they make it possible to judge whether a compositional finding is technically and analytically reproducible [5]. As a practical Level 1 example, the fermented-food arm of Wastyk et al. showed increased microbiota diversity; that result supports an ecological change, not by itself a mechanism for the accompanying immune changes [6].Shotgun metagenomics moves the question from who is present to what the community could do. Genes, pathways, and modules represent functional potential, not realized activity. Gene presence does not establish transcription, translation, enzyme flux, or substrate availability. This distinction is visible in both observational multi-omics work and diet trials: the gut metatranscriptome is related to, but not determined by, the metagenome [2], while Wastyk et al. found increased microbiome-encoded glycan-degrading CAZymes during a high-fiber intervention despite stable community diversity [6]. A Level 2 result therefore supports a change in encoded capacity, not proof that the corresponding biochemical function occurred.Activity is a different question. At Level 3, metatranscriptomics, metaproteomics, enzyme assays, and metabolite measurements can provide evidence of realized function, particularly when several layers agree [2,7,8]. Meslier et al., for example, combined shotgun metagenomics with fecal, plasma, and urinary metabolomics during a Mediterranean-diet intervention and observed changes in microbial functional modules and metabolites, including urolithins and bile-acid-related measures [7]. Metabolites still require careful source attribution: a molecule in stool or plasma may come from food, host metabolism, microbial metabolism, or all three. Diet can also alter transit time and the intestinal environment, changing both community structure and measured metabolite concentrations [9,10]. We would reserve the phrase "microbially produced" for findings supported by pathway concordance, paired microbial measurements, isotope tracing, or another source-resolving approach.To reach Level 4, the microbial signal must be linked to a host response in time. Two endpoints that both differ between baseline and study end are not enough. Repeated sampling should show that the microbial feature changes before, or alongside a plausible intermediate that precedes, the cardiometabolic outcome. Within-person coupling and dose-response patterns add credibility. Wang et al. analyzed longitudinal microbiome profiles alongside long-term dietary data and cardiometabolic biomarkers, showing that the association of a Mediterranean-style diet with cardiometabolic risk differed according to microbial composition [11]. This illustrates diet-microbiome-host linkage, but not temporal mediation; common causes, reverse direction, and unmeasured exposures can still explain the observed relationships. Deeply phenotyped cohort analyses provide complementary host-linked evidence [12,13].The final step is causal attribution: evidence that an effect is microbiome dependent or microbiome mediated. This usually requires deliberate perturbation, transfer, tracing, rescue, or a causal model whose assumptions are explicit. Kovatcheva-Datchary et al. combined a barley-kernel intervention with transfer of responder microbiota into germ-free mice, providing a stronger test of whether microbiota associated with response could reproduce a metabolic phenotype [14]. In humans, allogenic versus autologous fecal microbiota transfer has directly perturbed the microbiome while measuring insulin sensitivity [15]. Mendelian-randomization analyses provide a different causal contrast [16], and targeted fiber interventions can connect defined microbial guilds with metabolic outcomes [17]. Each design answers a narrower question than the umbrella phrase "the microbiome caused the effect."Importantly, the five levels classify claims, not entire studies. A multi-omics trial may generate Level 1 compositional evidence, Level 2 functional-potential evidence, and Level 3 metabolite evidence at the same time. Concordance across layers strengthens a proposed pathway, but adding more assays does not automatically establish Level 4 linkage or Level 5 causality. Each conclusion should therefore be matched to the specific evidence that supports it.Causal wording should also distinguish necessity, sufficiency, and mediation. Blocking a microbial pathway can test whether it is necessary in a defined model; adding a community, strain, or metabolite can test sufficiency under the receiving conditions; mediation asks how much of the dietary effect operates through a prespecified post-exposure variable [18]. These questions require different contrasts and need not produce the same answer. A microbial feature that predicts response may be useful for stratification, but prediction does not show that changing the feature will change the outcome.Randomization settles one important question, not every question. It strengthens the inference that assignment to the dietary intervention caused the measured outcome, assuming missing data and contamination are handled appropriately. It does not establish that the microbiome mediated that outcome. The microbiome is measured after randomization and can be influenced by adherence, medications, illness, transit, and other time-varying factors. Trials of microbiota-targeted diets and Mediterranean dietary patterns show how an intervention can produce coordinated immunologic, metabolic, microbiome, and metabolome changes [6,7]. That coordination is biologically informative, but mediation still requires a prespecified mediator, appropriate temporal sampling, attention to mediator-outcome confounding, and an analysis that matches the causal question [18].A related problem is the phrase "the microbiome improved." Improvement has no clear meaning unless a functional or clinical criterion was defined in advance. There is no universal taxonomic signature of a healthy microbiome, and a taxon associated with benefit in one dietary or ecological setting may not behave similarly in another. It is more informative to report the feature measured, the direction and magnitude of change, its uncertainty, and the host outcome to which it was linked. Interindividual variation should not automatically be dismissed as noise. Baseline community structure, habitual diet, metabolic phenotype, and ecological constraints may all influence response to the same intervention. Responder analyses are most credible when prespecified, expressed on a continuous scale, and accompanied by uncertainty. If a baseline microbiome feature is proposed as an effect modifier, the interaction should be tested directly and then validated, rather than inferred from separate significance tests within subgroups.Four principles cut across the staircase. Exposure fidelity should be documented with delivery records, adherence measures, and, when feasible, dietary biomarkers. Temporality should be built into the sampling schedule rather than inferred after the fact. Quantitative rigor requires prespecified primary features, appropriate compositional methods, consideration of microbial load, multiplicity control, and power calculations tied to the intended claim [3,4,19]. Reproducibility depends on complete reporting of sample collection, storage, extraction, sequencing, preprocessing, statistical decisions, and null findings [5].At the protocol stage, five questions make this operational: (1) What exact biological claim is the study intended to support? (2) Which measurement level directly tests that claim? (3) Does the sampling schedule establish the required temporal order? (4) What alternative explanations -including diet, host metabolism, medications, transit, and source attribution -must be addressed? (5) If a microbiomedependent or mediated effect is claimed, is the causal contrast, mediator, confounding structure, and estimand specified in advance? Writing these answers before data collection makes it less likely that a sparse or exploratory design will later be described as mechanistic.Confirmatory and exploratory analyses should remain visibly distinct. A small number of primary microbial and host endpoints can anchor the main claim, while high-dimensional findings are reported as discovery results. Effect sizes and compatible intervals should accompany P values. Sensitivity analyses can show whether conclusions survive alternative preprocessing choices, missing-data assumptions, medication changes, and plausible mediator-outcome confounding. Sharing code, feature definitions, and metadata dictionaries makes replication possible even when individual-level data require controlled access.Figure 1 should not be read as a ranking of assay sophistication. It is a restraint on wording. Composition can identify ecological responders and generate hypotheses; functional potential can prioritize pathways; realized function can establish activity; host linkage can define a candidate diet-microbiome-host pathway; and causal designs can test whether that pathway is necessary, sufficient, or mediating under stated conditions. Each level is useful, but each licenses a different conclusion.The framework is also deliberately simplified. The levels are not always discrete, some assays bridge more than one biological layer, and interpretation depends on biological context, study design, timing, and the quality of data integration. Causal evidence may arise through triangulation across complementary designs rather than a single experiment. The staircase should therefore be used as a reasoning aid, not as a rigid scoring system or a requirement that every study progress sequentially from Level 1 to Level 5.The framework may make evidence synthesis more honest. Trials should be grouped by the biological layer measured and the claim tested, not only by dietary label or sequencing platform. An apparent inconsistency may disappear when one study measured community composition, another inferred encoded capacity, and a third quantified metabolic output. Conversely, repeated taxonomic associations should not be counted as repeated confirmation of a biochemical mechanism.The practical message is deliberately simple: do not move the claim higher than the measurement. When the intended claim guides the sampling schedule, assays, and analysis from the outset, positive findings become easier to interpret, null findings become more informative, and mechanistic conclusions are more likely to hold across populations and dietary contexts. Levels classify claims rather than entire studies; a single multi-omics study may contribute evidence at several levels. Exposure fidelity, temporality, quantitative rigor, and reproducibility apply across all levels. The framework does not require every study to reach Level 5; it requires that claims do not exceed the evidence generated.
No takes yet. Share an insight, caveat, or question.
Salamat et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: