From association to mechanism in microbiota–gut–brain research

The microbiota–gut–brain axis is now supported by a substantial body of experimental and human research.

Microbial colonisation can influence neurodevelopment, stress physiology, immune function, microglial maturation, intestinal and blood–brain barrier biology and behaviour in experimental models. Microbial metabolites can interact with neural, immune, endocrine and metabolic pathways. Human studies have identified associations between intestinal microbial features and psychiatric, neurological and behavioural phenotypes.^¹–⁴

The field has therefore moved beyond the question of whether communication between intestinal microorganisms and the nervous system is biologically possible.

The more difficult question is what any particular finding means.

A microbial association is not a mechanism. A plausible pathway is not a demonstrated causal sequence. A behavioural effect in an animal model is not evidence of the same clinical effect in humans. And the simultaneous presence of microbial, inflammatory, barrier and neural differences does not establish the order in which those differences arose.

As microbiota–gut–brain research matures, the central challenge is therefore one of inference:

How do we move from observing that two systems vary together to demonstrating how one contributes to change in the other?

Association is where many discoveries begin

Association is not weak science.

It is often the first indication that two biological processes may be connected.

Human case–control studies have identified microbial differences associated with depression, anxiety-related phenotypes, autism spectrum disorder, Parkinson’s disease and other neurological or psychiatric conditions.^⁵–⁸ Population studies have linked microbial features and functional potential with quality of life and depressive symptoms.^⁹

These observations create hypotheses.

They can identify organisms, pathways or metabolites worthy of further investigation. They can reveal reproducible patterns across populations. They can suggest that a clinical phenotype is accompanied by altered intestinal ecology.

What they cannot establish is causal direction.

If microbial composition differs between people with and without depression, the difference may contribute to depression. Depression may alter diet, sleep, medication use, intestinal transit or stress physiology and thereby change the microbiota. A third factor may influence both. Or reciprocal processes may operate simultaneously.^¹⁰–¹²

The association remains scientifically meaningful.

Its meaning is simply narrower than a causal claim.

Temporality is the first major step beyond association

A cause must precede its consequence.

This apparently simple principle becomes difficult in microbiome research because many human studies measure microbial composition only after disease is established.

Longitudinal studies improve inference by repeatedly measuring microbial and host variables over time. They can ask whether microbial changes precede, accompany or follow changes in symptoms, physiology or treatment.^¹³,¹⁴

This is one reason longitudinal design is central to our Meals, Microbiota and Mental Health in Children and Adolescents study. The MMM-Study integrates dietary intake, microbiota, intestinal-permeability-related measures and mental-health phenotypes over time rather than treating a single microbiome sample as a fixed exposure.^¹⁵

Temporal information does not prove causality.

A microbial change can precede an outcome because both are responding to an earlier unmeasured exposure.

But temporality eliminates some causal structures and strengthens others.

It changes the question from are these variables different together? to how do they change in relation to one another?

Confounding can create biologically convincing stories

Microbiota–gut–brain research is particularly vulnerable to confounding because many variables influence both intestinal microbial ecology and brain-related outcomes.

Diet is one.

Medication is another.

Age, sex, adiposity, geography, physical activity, sleep, alcohol, smoking, socioeconomic conditions, gastrointestinal symptoms and stress can all matter.^{¹⁰–¹²,¹⁶}

The problem is not merely statistical.

Confounding can create a biologically plausible narrative.

Suppose people with depression consume less dietary fibre. Their microbial communities and fermentation patterns differ from those of controls. The microbial difference correlates with symptom severity.

It is tempting to infer:

microbiome difference → depression

But the underlying structure may include:

depression → dietary change → microbiome difference

or:

socioeconomic or behavioural exposure → diet + depression + microbiome

or several reciprocal relationships.

Statistical adjustment can help, but only for variables that have been measured adequately and incorporated into an appropriate causal model.

No regression model can adjust for a causal structure that has not been considered.

Reverse causality is especially important in the gut–brain axis

The microbiota–gut–brain axis is explicitly bidirectional.

This makes reverse causality not an inconvenience but a central biological possibility.

Stress can alter autonomic activity, gastrointestinal motility, secretion, immune responses and intestinal permeability. These changes can reshape the microbial environment.^{¹⁷–¹⁹}

Neurological disease can alter mobility, diet, bowel function and medication exposure.

Psychiatric symptoms can alter sleep and eating behaviour.

The brain can therefore generate conditions that produce the microbial signature later associated with the brain phenotype.

A cross-sectional microbiome difference cannot distinguish these routes.

The more bidirectional the biological system, the more cautious unidirectional causal language should become.

Biological plausibility strengthens a hypothesis but does not complete it

The microbiota–gut–brain axis contains many plausible mechanisms.

The vagus nerve can transmit visceral information.

Microbial metabolites can influence epithelial, immune and endocrine signalling.

Cytokines can affect neural function.

Tryptophan metabolism can intersect with serotonergic and kynurenine pathways.

Short-chain fatty acids can influence immune and cellular physiology.

Barrier dysregulation may alter host exposure to luminal material.^{¹–⁴,²⁰}–^²³

These pathways make a microbial effect on brain physiology plausible.

But plausibility is not evidence that a particular pathway mediated a particular human association.

The existence of a vagal route does not demonstrate that an observed depression-associated microbiome signal was transmitted through the vagus.

The ability of butyrate to influence immune cells does not demonstrate that a clinical phenotype resulted from altered butyrate exposure.

The existence of intestinal permeability dysregulation does not establish that it connected a microbiome difference to a neural outcome.

Mechanistic knowledge narrows what is possible.

It does not tell us what actually happened in a specific study.

Mediation requires measuring the middle of the pathway

Many microbiota–gut–brain claims contain an implicit chain.

For example:

microbial change → altered metabolite → immune response → brain effect → behavioural phenotype

If only the microbial composition and behavioural phenotype are measured, everything in the middle remains inferred.

A stronger study measures intermediate steps.

Did the metabolite actually change?

Was the immune pathway activated?

Did the neural outcome follow?

Did variation in the intermediate signal statistically or experimentally account for part of the relationship?

Even then, statistical mediation is not equivalent to biological proof.

But measuring the proposed intermediates transforms a narrative into a testable pathway.

This is one of the major advantages of integrating microbiome data with metabolomics, immunology, barrier measurements, endocrine physiology and neural phenotyping.^{¹³,²⁴,²⁵}

Multi-omics becomes most useful when it tests a predefined biological sequence rather than simply producing more correlations.

Experimental manipulation strengthens causal inference

Experimental models can move beyond observation by deliberately altering the proposed causal variable.

Germ-free models test dependence on microbial exposure.

Defined colonisation tests selected organisms.

Antibiotic perturbation alters established communities.

Microbiota transfer tests whether aspects of a phenotype are transmissible.

Vagotomy can test neural-pathway dependence.

Receptor knockout can test host signalling.

Microbial gene manipulation can test whether a specific microbial function is necessary.^{²⁶–³⁰}

The logic is powerful.

If altering a proposed cause changes the outcome, causal inference strengthens.

If blocking the proposed intermediate pathway prevents the effect, it strengthens further.

But experimental evidence must remain bounded by the model.

A causal mechanism in a mouse is evidence that the mechanism can operate in that biological system.

Human relevance remains a separate question.

Microbiota transfer provides causality — within limits

Human-to-animal microbiota transfer has become particularly influential because it begins with a human microbial community and tests its effects experimentally.

Studies have reported behavioural or physiological differences after transfer of microbiota from individuals with depression, autism spectrum disorder, Parkinson’s disease and other phenotypes into recipient animals.^³¹–³⁴

These experiments provide evidence that characteristics of the transferred microbial community can contribute to recipient phenotypes.

That is substantially stronger than a human association alone.

But the conclusion should remain exact.

The human disease itself has not been transferred.

The recipient is a different species.

Engraftment is incomplete.

Diet, immune biology and intestinal ecology differ.

And behavioural assays measure selected dimensions rather than human diagnostic states.

The experiment demonstrates microbiota-dependent transmissibility of a measured phenotype under controlled conditions.

That is a meaningful causal statement without extending beyond the evidence.

Intervention in humans adds another causal layer

Randomised interventions can test whether deliberately changing microbial exposure alters a human outcome.

Probiotics, prebiotics, synbiotics, dietary interventions and, in selected contexts, microbiota-transfer approaches have been investigated in relation to stress, mood, cognition and neurological or psychiatric symptoms.^{³⁵–³⁹}

Randomisation reduces several forms of confounding.

If an intervention changes an outcome compared with an appropriate control, the intervention contributed causally to that difference.

But even here, the mechanism may remain uncertain.

A probiotic may influence gastrointestinal symptoms, immune signalling, microbial metabolism or host physiology without producing a large measurable change in community composition.

A dietary intervention may act directly on host metabolism while simultaneously changing the microbiota.

A clinical effect therefore establishes intervention efficacy under the trial conditions.

It does not automatically establish microbiome mediation.

To make that claim, the microbial pathway itself must be measured and linked to the outcome.

Strain specificity limits generalisation

This is particularly important in probiotic research.

Microbial effects can be strain-specific.

Evidence obtained with one Lactobacillus or Bifidobacterium strain cannot automatically be attributed to the species, genus or probiotics as a class.^⁴⁰,⁴¹

The classic vagotomy experiment using Lactobacillus rhamnosus JB-1 illustrates the point. The behavioural and neurochemical findings belong to that experimental strain and model.^²⁸

They do not establish that every strain carrying the label L. rhamnosus acts through the vagus.

Clinical translation therefore requires precise microbial identity.

The same principle applies to microbial consortia and faecal microbiota transfer.

The biological exposure must be defined before its effect can be generalised.

Dose and exposure matter

Mechanistic studies often demonstrate that a microbial metabolite or molecule can produce a biological effect.

The next question is whether the host encounters that compound at a sufficient concentration under physiological conditions.

This is a classic problem of exposure.

A metabolite may alter neuronal or immune function in vitro at a concentration much higher than that found in human circulation.

An orally administered compound may produce tissue exposure unlike endogenous microbial production.

A faecal concentration may not reflect absorption.

A plasma concentration may be influenced by hepatic or renal physiology.^¹³,²²–^²⁵

Demonstrating biological activity is therefore not enough.

Mechanism requires plausible exposure.

Barrier mechanisms require evidence on both sides of the barrier

Intestinal permeability is frequently positioned between microbial dysbiosis and systemic or neural effects.

A common proposed sequence is:

dysbiosis → increased permeability → microbial translocation → systemic inflammation → neuroinflammation → behavioural change

Every element is biologically plausible.

But each is also a distinct measurement problem.^{⁴²–⁴⁴}

Microbial composition does not measure permeability.

A zonulin-related assay does not automatically measure functional permeability.

A circulating microbial-associated marker does not establish the anatomical route by which it entered the circulation.

A systemic cytokine does not demonstrate neuroinflammation.

And a behavioural phenotype does not identify the neural mechanism responsible.

This is precisely where attractive narratives can outrun evidence.

The chain is only as strong as the links actually measured.

My own research illustrates the value of keeping levels separate

This distinction between levels of evidence has shaped my own work.

Our systematic review and meta-analysis of zonulin-dependent intestinal permeability in children diagnosed with mental disorders examined whether the available human literature supported altered zonulin-related measurements in these populations.^⁴⁵

The evidence was suggestive in some contexts but heterogeneous.

It did not establish a causal microbiota–permeability–brain pathway.

In experimental work using wild-type and zonulin-transgenic mice, we examined intestinal microbiota and behavioural outcomes following bovine-colostrum intervention.^⁴⁶ Subsequent transcriptomic analysis identified intestinal responses involving tight-junction and inflammatory pathways, including sex-specific patterns.^⁴⁷

These studies provide evidence at different levels.

The systematic review synthesises human association.

The mouse model allows experimental manipulation.

The transcriptomic study adds molecular resolution.

None alone completes the causal chain.

Together, they show why convergence matters.

Triangulation is stronger than one decisive experiment

Complex biological systems rarely yield to a single definitive study.

Causal confidence grows when different methods with different limitations converge.

A human association identifies relevance.

A longitudinal study establishes temporal ordering.

A microbial-transfer experiment tests transmissibility.

A defined microbial model isolates the organism.

Metabolomics identifies the candidate mediator.

A receptor knockout tests pathway necessity.

A human intervention tests modifiability.

Replication tests robustness.

When these independent lines point towards the same mechanism, the conclusion becomes stronger because the weaknesses of one method are partly compensated by the strengths of another.^⁴⁸

This is triangulation.

It is different from accumulating many studies that all use the same design and therefore share the same bias.

Null findings are part of mechanism

Mechanistic narratives are often built from positive findings.

But null and discordant results can be equally informative.

If a probiotic changes microbial composition but not behaviour, the microbial change may not be sufficient for that outcome.

If a metabolite changes but the proposed immune pathway does not, the pathway may be wrong or incomplete.

If an effect observed in germ-free animals disappears in conventionally colonised animals, the extreme microbial context may be necessary.

If an association fails to replicate across human populations, it may depend on diet, geography, medication or another environmental factor.

A mechanism is defined partly by the conditions under which it fails.

Negative evidence creates boundaries.

Sex, age and developmental timing can modify causality

A causal pathway may not operate equally in every host.

Sex hormones, immune development, neural maturation and microbial ecology change across the life course. Experimental studies have demonstrated sex-dependent microbiota–brain effects, while early-life microbial exposure can influence developmental processes differently from microbial manipulation in adulthood.^{⁴⁹–⁵¹}

This means that heterogeneity is not always noise.

A microbial pathway may be relevant only during a developmental window or in one physiological context.

Our own transcriptomic findings in zonulin-transgenic mice showed sex-specific intestinal responses to intervention.^⁴⁷

Such interactions should be investigated rather than averaged away.

Causal inference becomes more precise when we ask not only does this pathway operate? but for whom, when and under which biological conditions?

Replication is a causal tool

Replication is often discussed as a quality-control issue.

It is also part of causal inference.

If a microbial association appears only in one cohort, one laboratory or one mouse colony, its general biological relevance remains uncertain.

Replication across populations with different diets and environments can reveal which findings are robust and which depend on context.^¹⁰,⁵²

Mechanistic replication is even stronger.

If a microbial metabolite produces the predicted host response in cell systems, animal models and human observational data, confidence increases.

If an intervention alters the metabolite and the downstream host response in the predicted direction, it increases further.

Replication is therefore not simply repeating an experiment.

It tests whether the proposed mechanism survives changes in context.

Causal language should match causal evidence

Scientific language matters because it determines how evidence is communicated.

Associated with is not weaker wording when the study is observational.

It is accurate wording.

Contributed to may be appropriate when an experimental manipulation changes an outcome but does not fully explain it.

Mediated by requires evidence that the intermediate pathway accounts for the relationship.

Caused requires stronger evidence still.

Using causal language conservatively does not diminish the importance of microbiota–gut–brain research.

It protects it.

The field becomes more credible when the strength of the verb matches the strength of the design.

The next phase of the field is not more association

The microbiota–gut–brain axis has generated thousands of associations.

More will continue to be useful, particularly in large longitudinal cohorts and understudied populations.

But the major scientific opportunity now lies in connecting the levels.

Which microbial gene produces the relevant molecule?

Does that molecule reach the host?

Which host receptor detects it?

Which cell responds?

Which neural or endocrine pathway changes?

Does blocking that pathway remove the phenotype?

Does the mechanism operate in humans?

Can it be modified safely?

Does modification improve a meaningful clinical outcome?

These questions move the field from description towards mechanism.

From a compelling story to a demonstrated pathway

The microbiota–gut–brain axis is compelling partly because the pieces fit together so naturally.

Microbes produce metabolites.

The intestine contains nerves and immune cells.

The barrier regulates exposure.

The circulation connects organs.

The brain responds to immune, endocrine and metabolic signals.

Stress changes gastrointestinal physiology.

Every component of the story is biologically plausible.

That is precisely why methodological discipline matters.

When a narrative is compelling, it becomes easier to overlook which links were measured and which were inferred.

The goal is not to dismantle the microbiota–gut–brain concept.

It is to make it more exact.

A mature field should be able to say where evidence is strong, where it is suggestive, where models disagree and where a causal bridge has not yet been built.

The most important transition in microbiota–gut–brain research is therefore not from gut to brain. It is from association to mechanism.