Reviewer says the mechanism is not shown

A reviewer comment that the mechanism is not shown usually means the manuscript’s title or abstract claimed a causal step — via, through, -dependent — that no figure tests, not that a full mechanistic study is required. The fix is often three rewritten sentences; a genuine gap instead needs an epistasis series, a rescue, or a reconstitution.

Reframing the comment this way changes the arithmetic of a revision. A mismatch between claim and evidence is closed by editing three sentences. A genuine mechanistic gap is closed by an epistasis series, a rescue, or a biochemical reconstitution — weeks to months of work. The sections below separate the two, name what each version of the objection is asking for, and set out what to do when the experiment the reviewer wants is not available in your system.

Read your own title first

If the title says “X promotes Y via Z”, you have made a mechanistic claim and the reviewer is entitled to ask for evidence of Z specifically. If your data show that X promotes Y and that Z is involved somehow, the honest fix may be the title rather than an experiment.

The same applies to the abstract’s final sentence and the discussion’s opening, which is where mechanistic claims tend to appear without support elsewhere in the paper. A phenotype attributed to a pathway is a mechanistic claim even when the word mechanism never appears.

Four constructions carry a mechanistic commitment and are worth grepping for in your own manuscript: via, through, by, and -dependent. “IL-6–dependent proliferation” asserts that removing IL-6 signalling removes the proliferation, which is an epistasis result; if you only showed that IL-6 is elevated, the hyphen has made a claim your figures do not carry. The same applies to mediates, drives, acts on and targets. Attach a figure number to every one of them. A mechanistic word with no figure number beside it is the exact string the reviewer will quote back, and the fastest way to find them all before submission is the overclaim check.

Check the graphical abstract and the cover letter too. Both are read before the figures, both routinely carry an arrow the data do not support, and neither is usually revised when the manuscript is; guidance on the letter is in how to write a cover letter.

What counts as mechanism

Reviewers generally mean one of three things, and they are not equally demanding.

The molecular intermediate. What is between X and Y. Requires identifying the intermediate and showing that perturbing it breaks the link.

Necessity and sufficiency of the proposed step. If you propose that X acts through Z, then removing Z should block the effect of X, and activating Z should reproduce it. Epistasis experiments of this shape are what most reviewers mean.

Direct versus indirect. Whether X acts on Y itself or through something else. Binding, structural, or reconstitution evidence.

Which one is being asked for is usually inferable from the reviewer’s other comments and from the journal’s expectations.

The fourth row is the common case, and it is the one authors skip past because it does not feel like an answer.

A sufficiency test bought by overexpression carries its own artifact risk, and specialist reviewers know it. Driving Z far above its physiological range with a strong constitutive promoter can produce a phenotype that a normal, endogenous level of Z never would — the rescue “succeeds” but says nothing about the pathway at physiological concentrations. A reviewer who has seen this before asks whether the rescue used an inducible or dose-titratable system tunable toward the endogenous level, not only whether a rescue exists at all.

Common questions about what a mechanism comment means

Reviewer 2 says the mechanism isn’t shown — do I actually need to run new experiments?

Not automatically. Check whether your title or abstract used a connective — via, through, -dependent, mediates — that no figure tests; if so, the fastest fix is narrowing that sentence, not a new experiment. New work is only required when your claims already match your data and the reviewer still wants the intermediate identified or perturbed.

How do I tell if the reviewer wants an experiment or just doesn’t like my title?

Compare your title and abstract’s final sentence against your figures, line by line. If a causal verb — drives, mediates, via — has no figure number attached to it anywhere in the manuscript, that is a claim–evidence mismatch a rewrite fixes. If every claim already has a figure and the reviewer still asks for mechanism, they want new data, not new wording.

Is showing that a pathway is activated alongside my phenotype enough to claim mechanism?

No — that is a correlation within your own dataset, not a mechanistic result. Showing phospho-SMAD3 rises when KIN1 is lost tells you TGF-β signalling moves with the phenotype; it does not show the phenotype requires that signalling. Only removing the pathway in the same background and watching the phenotype persist or disappear answers that question.

Mechanism is not the same objection as causation

Mechanism and causation are different objections with different remedies, and conflating them is why revisions overshoot. Correlative, not causal says you have not shown that X does anything to Y. “Mechanism not shown” concedes that X does something to Y and asks how — what sits between them.

The practical consequence is that a rescue experiment answers the causal objection and does not, on its own, answer the mechanistic one. Removing X, losing Y, and restoring X with a reagent-resistant construct establishes that X is required. It says nothing about the intermediate. If the reviewer has already accepted your perturbation of X and still writes about mechanism, adding a second perturbation of X will not move them; the missing experiment is about Z.

The inverse also holds and is worth knowing before you plan a revision: a demonstrated mechanism does not establish causation on its own. Showing that X binds Z, and that Z regulates Y, does not show that the X–Y relationship in your system runs that way. Reviewers who have seen a plausible biochemical story attached to an unperturbed phenotype write both comments in the same review.

Reading which of the three is being asked

The rest of the review tells you which reading you are facing, and it is worth spending twenty minutes on this before ordering reagents.

Vocabulary about identity — “the responsible factor”, “which effector”, “what lies downstream” — means the intermediate is missing and the reviewer expects a screen, a proteomic or transcriptomic comparison, or a candidate test.

Vocabulary about consequence — “necessary”, “sufficient”, “epistatic”, “dose-dependent” — means the intermediate is named but untested, and the reviewer wants perturbation of Z.

Vocabulary about physical relationship — “direct”, “binds”, “in vitro”, “purified”, “structural basis” — means the reviewer doubts a step you have drawn as an arrow, and no amount of cell biology will settle it.

A fourth signal is silence about mechanism in the numbered comments. If the reviewer’s specific points are all about controls and reagent specificity rather than pathway logic, the mechanism sentence is probably a summary judgement about framing rather than a request for an experiment.

The rewriting route

If your data establish a relationship without establishing how, describing it accurately is legitimate and frequently the right revision. “X is required for Y” is a real finding. “X is required for Y, and our data suggest this occurs through Z, though we have not tested this directly” is honest and reads as competence.

What triggers a second round is a mechanism asserted in the abstract, hedged in the discussion, and untested anywhere. Reviewers notice the inconsistency and read it as an attempt to have the claim without the evidence.

The rewrite has a shape. Move the mechanistic sentence out of the results and into a clearly marked paragraph in the discussion. Replace the causal verb with the one your evidence supports — is required for rather than drives, is associated with rather than mediates. Name the experiment that would test the speculation, in one sentence, so the reader can see you know what is missing. Then make the title match. A title narrowed from “X promotes Y via Z” to “X is required for Y” has lost no finding, and it removes the sentence the reviewer was objecting to.

Authors resist this because a narrower title feels like a smaller paper. In review it reads as the opposite: a claim that exactly fits its evidence signals that the rest of the manuscript can be trusted, and it removes the specific inconsistency that generates a second round.

Worked example: a title that outran its figures

Consider an ordinary submission: “Loss of the kinase KIN1 drives fibrosis via TGF-β signalling”, supported by four figures. Figure 1 shows KIN1 protein reduced in fibrotic tissue. Figure 2 shows conditional deletion of KIN1 in mice increases collagen deposition. Figure 3 shows phospho-SMAD3 elevated in the KIN1-null tissue. Figure 4 shows a TGF-β receptor inhibitor reduces collagen in wild-type animals.

The reviewer writes that the mechanism is not shown, and they are right in a precise way. Figures 1 to 3 establish that KIN1 loss produces fibrosis and that TGF-β signalling is elevated when it does. Figure 4 establishes that TGF-β signalling contributes to fibrosis in animals that still have KIN1. Nothing in the set tests whether the fibrosis caused by KIN1 loss requires TGF-β, because the inhibitor was never given to the KIN1-null animals. The word via is carried entirely by two experiments that never met.

The missing experiment is one arm: the inhibitor in the KIN1-null background, with the wild-type and untreated controls run alongside. That is a four-group design, and in most laboratories it is a single animal study rather than a research programme. The rewriting alternative is a title of “Loss of KIN1 promotes fibrosis and elevated TGF-β signalling”, which is what the four figures show.

Note which mistake this is. Nothing is wrong with any experiment; each figure is well controlled and would survive a methods critique. The defect is that the connective in the title was never tested, and it is invisible while you read the figures one at a time.

Epistasis, and how it is misread

Epistasis experiments place two perturbations in the same animal, cell or extract and ask whether their effects combine or collapse. If X acts through Z, then perturbing both should look like perturbing Z alone; if they act in parallel, the double perturbation should exceed either single one. That comparison, not the two single perturbations, is the mechanistic result.

The word carries two distinct meanings, and reviewers use both. William Bateson’s original genetic sense — one locus masking the phenotypic effect of another — is the pathway-ordering usage above. R. A. Fisher’s statistical sense, introduced in “The correlation between relatives on the supposition of Mendelian inheritance” (Transactions of the Royal Society of Edinburgh 1918;52:399–433), is a departure from additivity on a chosen scale. A significant interaction term is Fisher’s epistasis and is not by itself a pathway order.

Three design errors account for most epistasis experiments that fail review.

Saturation. If either single perturbation already drives the readout to its floor or ceiling, the double perturbation cannot exceed it and non-additivity is uninformative. Run the perturbations at a sub-maximal level — a partial knockdown, a lower inhibitor concentration, a hypomorphic allele — where there is room in the assay for the double to be worse.

Missing arms. The design has four groups: neither perturbation, X alone, Z alone, both. Manuscripts routinely report three and leave the reader to infer the fourth. Reviewers count the arms.

Scale. Additive and multiplicative expectations differ, and “more than additive” on a log scale can be “less than additive” on a linear one. State the null model you are testing against before the result, and report the interaction estimate with its confidence interval rather than three sets of asterisks.

The sufficiency arm is the second half and is skipped more often than the necessity arm. Activating Z in a background lacking X — a constitutively active allele, a ligand, a degron released — is what converts “Z is required downstream of X” into “Z is the step”.

Epistasis logic further assumes the relationship between Z and the readout is monotonic — that more Z reliably means more of the phenotype, or less. Where Z has a biphasic or U-shaped effect, which is common for signalling nodes once expression moves well past the physiological range, a double perturbation can look additive or even suppressive for reasons that have nothing to do with pathway order. The four-arm design alone will not catch this; it needs an independent dose-response curve for Z by itself before the double-perturbation result can be trusted as a pathway statement rather than an artifact of where on that curve the experiment landed.

Questions about epistasis and knockout experiments

Do I need to identify the actual intermediate, or does showing necessity count as mechanism?

Which reading the reviewer means decides the answer. Necessity and sufficiency of a named step is the reading usually meant, and it is satisfied without identifying every downstream node: remove Z and lose the effect of X, restore Z and restore it. Identity of the intermediate is a separate, harder ask, and reviewers signal it with vocabulary like “the responsible factor” or “which effector.”

My knockout and knockdown of the same gene are giving me different phenotypes — is that a mechanism problem?

A knockout-versus-knockdown mismatch is a specific, diagnosable result rather than noise to explain away. A germline knockout can trigger transcriptional adaptation that a knockdown does not, so the knockout phenotype can look milder, or different in kind, than the knockdown — meaning a mechanism inferred from the knockout allele may describe the compensation rather than the gene itself.

Is n=3 mice per group enough for an epistasis experiment?

Sample size does not rescue a design with the wrong arms. An epistasis experiment needs all four groups — neither perturbation, X alone, Z alone, both together — run at a sub-saturating dose of each. Three well-powered animals per group in a correctly designed four-arm study is more defensible to a reviewer than five animals spread across a three-arm design that is missing the double perturbation.

Direct versus indirect, and what evidence separates them

A claim that X acts directly on Y is a biochemistry claim, and it is answered with purified components rather than in cells. Co-immunoprecipitation from a lysate does not distinguish a direct interaction from a bridged one, because everything in the lysate is available to bridge it. Reviewers who write “direct” are asking for recombinant protein, a crosslink at a defined residue, a structure, or a separation-of-function mutant that abolishes the interaction and the phenotype together.

A separation-of-function mutant is only as good as its folding control. A point mutation that abolishes the X–Z interaction but also destabilises the protein will read as loss of direct binding when it is really loss of protein, so a reviewer asking for this evidence increasingly also expects a check — a thermal-shift assay, a size-exclusion profile, or simply comparable expression on a blot — that the mutant still folds and accumulates before its binding is scored as lost.

Where the perturbation is a small molecule, the standard the reviewer is applying is usually the kind of chemical probe criterion set out by Arrowsmith and colleagues in “The promise and peril of chemical probes” (Nature Chemical Biology 2015;11:536–541), who give as a working example the thresholds the Structural Genomics Consortium requires of probes for epigenetics targets: “in vitro potency at the target protein of <100 nM”, “>30-fold selectivity relative to other sequence-related proteins of the same target family”, and “demonstrated on-target effects in cells at <1 μM”. The same paper reports that the consortium “has strongly encouraged and now requires that the chemical probe be accompanied by an inactive close analog of the compound to serve as a negative control”, and that “each protein should be targeted by another equally well-characterized ‘orthogonal’ chemical probe having a completely different chemical structure”. A compound used at 10 μM with no inactive analogue is not evidence about its nominal target, and that is a mechanistic objection even when the reviewer writes it as a concentration complaint. Choosing a second reagent that fails differently is the substance of orthogonal validation.

Timing is the other discriminator between direct and downstream. Acute depletion systems remove a protein fast enough that the first changes to appear are close to it: the auxin-inducible degron (Nature Methods 2009;6:917–922), its low-leak successor AID2 (Nature Communications 2020;11:5701), and the dTAG system pairing a degrader of FKBP12^F36V with an in-frame knock-in at the endogenous locus (Nature Chemical Biology 2018;14:431–441) all act on a timescale of minutes to hours. A chronic knockout measured at steady state cannot separate a primary consequence from an adaptation, and where a knockout and a knockdown disagree, that disagreement is its own diagnosable result rather than a nuisance. El-Brolosy and colleagues reported in Nature (2019;568:193–197) that “alleles that fail to transcribe the mutated gene do not exhibit transcriptional adaptation, and these alleles give rise to more severe phenotypes than alleles displaying mutant mRNA decay” — which means a mechanism inferred from a premature-termination-codon allele may be a mechanism of the compensation rather than of the gene.

Mechanism in observational and clinical work

In observational and clinical work, “mechanism not shown” almost always means the manuscript has proposed a pathway variable and analysed it informally. The remedy is a named mediation analysis rather than a covariate.

Reporting that an association attenuates after adjusting for a candidate intermediate is not a mechanistic result; it is a coefficient comparison whose behaviour depends on the model, and the intermediate–outcome relationship is confounded in ways adjustment cannot see. The distinction between a variable on the path and a variable beside it is the substance of mediator versus confounder, and it is settled by a graph rather than by the data — write the directed acyclic graph that produced your adjustment set so the reviewer can dispute the graph instead of guessing at it.

The method to describe is no longer the causal-steps procedure of Baron and Kenny (Journal of Personality and Social Psychology 1986;51:1173–1182), which remains in wide use despite having no formal identification conditions and no way to accommodate an exposure–mediator interaction. VanderWeele’s four-way decomposition (Epidemiology 2014;25:749–761) separates a controlled direct effect, a reference interaction, a mediated interaction and a pure indirect effect, and its identification rests on four no-unmeasured-confounding conditions that a methods section should name rather than assume. Implementations exist in the R packages mediation (Tingley and colleagues, Journal of Statistical Software 2014;59(5)) and CMAverse (Shi, Choirat, Coull, VanderWeele and Valeri, Epidemiology 2021;32:e20–e22).

Two reporting details draw the comment on their own. Report the proportion mediated only when the total effect’s confidence interval excludes the null, because it is a ratio with the total effect in its denominator and becomes unstable, exceeding 1 or turning negative, as that denominator approaches zero. And bound what you cannot exclude: mediational E-values (Smith and VanderWeele, Epidemiology 2019;30:835–837) give the minimum strength of unmeasured mediator–outcome confounding required to explain away a direct or indirect effect. A separate objection arises when the intermediate is also the outcome being reported, which is the surrogate endpoint problem rather than a mediation one.

A mediator claim also carries a temporal-order requirement that a purely cross-sectional dataset cannot satisfy: the candidate mediator has to be measured after the exposure and before the outcome, not at the same wave as the outcome. A mediator measured contemporaneously with the outcome cannot be ordered on the causal path, and adjusting for it risks conditioning on a collider rather than isolating an indirect effect — a distinction the DAG should encode explicitly, before the model is specified, rather than being discovered in the covariate list afterward.

Questions about direct evidence and mediation analysis

The reviewer wants direct binding evidence — is co-immunoprecipitation from a cell lysate enough?

No. Co-IP from a lysate cannot distinguish a direct interaction from one bridged by a third protein, because everything else in the lysate is available to do the bridging. Reviewers asking for “direct” generally want recombinant protein, a defined crosslink, a solved structure, or a separation-of-function mutant that loses the interaction without losing its fold.

My compound only works at 10 micromolar — is that a mechanism problem or a dose problem?

Both, and reviewers read them as the same objection. A compound used far above its reported in vitro potency, with no inactive structural analogue run alongside it, is not evidence about its nominal target — off-target effects at that concentration are at least as plausible as the intended one. The fix is a lower concentration plus a negative-control analogue, not a larger sample size.

Is a mediation analysis the same thing as showing mechanism in a clinical paper?

A mediation analysis is the clinical-research equivalent of mechanism, with its own failure mode: reporting that an association attenuates after adjusting for a candidate mediator is a coefficient comparison, not an identified indirect effect. A named method — VanderWeele’s four-way decomposition, for instance — with its no-unmeasured-confounding assumptions stated explicitly is what a reviewer asking for mechanism in observational work actually wants to see.

When mechanism is genuinely required

Some venues expect a mechanistic advance as a condition of publication, and no rewriting satisfies that — it is a scope judgement rather than a rigour one. If the objection is really “this is descriptive and we publish mechanism”, the choice is to do the work or to move the paper to a venue whose scope fits. That is a legitimate outcome and often the faster route to publication.

Three signals distinguish a scope judgement from a rigour objection. The comment appears in the reviewer’s summary paragraph rather than among the numbered points. The editor’s letter repeats it in the editor’s own words. And the reviewer does not name an experiment — a rigour objection almost always specifies what would answer it, while a scope judgement says the paper is “descriptive” and stops. Where all three hold, a revision aimed at the mechanism comment is unlikely to change the decision, and journal selection is the more productive move. Reviewer says this is incremental covers the scope version of this conversation.

What counts as sufficient mechanistic evidence is itself field-specific, and calibrating to the wrong standard wastes a revision cycle. A structural biology or biochemistry journal generally holds authors to the direct-versus-indirect standard described above; a genetics or cell-biology journal usually accepts a clean epistasis series as the ceiling of what is asked; a clinical or epidemiological journal accepts a pre-registered mediation analysis with its identification assumptions stated. Reading two or three recent mechanism-focused papers the target journal has actually published is a faster way to calibrate the expected standard than inferring it from one reviewer’s comment.

It is worth stating the boundary plainly, because authors internalise the opposite. Absence of a demonstrated mechanism does not refute a causal claim. Austin Bradford Hill listed plausibility among the nine viewpoints in “The environment and disease: association or causation?” (Proceedings of the Royal Society of Medicine 1965;58:295–300) and presented them as considerations to weigh rather than a checklist to satisfy. Descriptive and phenomenological work has priority claims of its own, and a well-characterised phenomenon reported without a mechanism is a contribution. What is not defensible is a mechanism asserted in the title of such a paper.

When the obvious remedy is unavailable

The epistasis experiment is unavailable more often than reviewers assume: the intermediate is essential and its deletion kills the cell, the organism has no conditional genetics, the phenotype exists only in human tissue, or the proposed step has no selective inhibitor. Say so specifically, then substitute.

Four substitutions carry weight. A hypomorphic or partial perturbation with the residual level stated — “70% protein reduction, quantified in Supplementary Figure 3” — establishes dose dependence where a null cannot. A conditional or degron allele converts an essential gene into a tractable one and an acute one at the same time. A naturally occurring loss-of-function variant, or a patient-derived line carrying the deletion, is a perturbation nobody had to perform. And a negative epistasis result is publishable: reporting that the double perturbation exceeded either single one, and concluding that the two act in parallel, is a mechanistic finding that removes the via from your title on the strength of evidence rather than caution.

Where none of these is available, narrow the claim and say what would test it. A limitations paragraph that names the experiment, names the reason it could not be done, and names what the result would distinguish is an analysis; see how to write a limitations section. “The mechanism warrants further investigation” is the sentence that draws the comment again in the next round.

What reviewers actually write

Reviewers rarely write “the mechanism is not shown” and stop. The specific versions below are what arrives, and each maps to one of the three readings.

“The authors conclude that X acts through Z, but Z is never perturbed.” “The inhibitor experiment in Figure 4 was performed in wild-type animals only; the epistatic relationship claimed in the title requires the inhibitor in the mutant background.” “The title states ‘via’, which the data do not support.” “The interaction is shown by co-immunoprecipitation from cell lysate and is described as direct throughout; please demonstrate a direct interaction or amend the wording.” “Elevated phospho-SMAD3 in the mutant is a correlation within the mutant, not evidence that the phenotype requires SMAD3.” “The compound was used at 10 μM, well above its reported potency, without an inactive analogue.” “The mediation analysis reports an attenuated coefficient; please report natural direct and indirect effects with their identification assumptions.” “The work is descriptive and does not advance mechanistic understanding.”

The last of those is the scope judgement, and it is the only one on the list that a revision cannot answer.

Wording the response

Answer a mechanism comment in one of exactly three shapes, and say at the top of the reply which shape you are using.

You did the experiment. Give the new figure number in the first sentence, state the result including the interaction estimate, and say which of the three readings it addresses. Do not bury it under a paragraph of agreement.

You narrowed the claim. Quote the old title and the new one side by side, list the sentences you changed with their line numbers, and state plainly that the original wording asserted more than the data support. Reviewers accept this far more readily than authors expect, because it concedes the exact point they made.

You could not do it. Name the obstacle concretely — the gene is essential in this line, the inhibitor has no selective analogue, the tissue is not obtainable — then give the substitute you ran and the claim you narrowed to. An unexplained refusal reads as evasion; a named obstacle with a substitution reads as judgement. The general structure is in how to write a response to reviewers.

Where two reviewers disagree about whether mechanism is required, answer the more demanding one in the manuscript and address the disagreement directly in the letter to the editor, who is the person adjudicating it.

Frequently asked questions about responding to the reviewer

I can’t do the epistasis experiment my reviewer wants because the gene is essential — what should I write in my rebuttal?

Name the obstacle concretely — essential gene, no conditional allele, no selective inhibitor — then give the substitute you ran: a hypomorph, a degron system, or a patient-derived line carrying a natural loss-of-function variant. An unexplained refusal reads as evasion; a named obstacle paired with a substitution and a narrowed claim reads as judgement.

Reviewer says my paper is descriptive, not mechanistic — should I argue the point or just submit elsewhere?

Check where the comment sits first. If it is in the summary paragraph, repeated in the editor’s own words, and names no specific experiment, it is a scope judgement about the journal’s remit rather than a rigour objection — and a well-matched journal usually publishes the finding faster than a mechanistic study aimed at satisfying that one venue’s scope.

Can I put the mechanism claim in the discussion as speculation instead of removing it?

Yes, if it is clearly labelled as speculation and kept separate from your results. “Our data suggest this may occur through Z, though we have not tested this directly” is a normal, honest discussion sentence; the same idea stated as a result, without the hedge, in the abstract, is what draws the objection back in the next round.

Before submission

The overclaim check traces every claim in the title, abstract and discussion back to the specific result meant to support it, and mechanism asserted without evidence is one of the categories it reports explicitly — attributing a phenotype to a target requires evidence of engagement with that target. Running it before submission is how you find out that your title makes a claim your figures do not.

PerfectPaper reads the title, the abstract’s final sentence and the figure legends together, and reports every mechanistic connective — via, through, -dependent, mediates — that no figure in the manuscript tests. Each finding names the sentence, the reading of “mechanism” it commits you to, and the arm of the design that is missing, so the choice between an experiment and a rewrite is made before a reviewer makes it for you.

More in this family: reviewer comments on methods and design.

Related

Correlative, not causal · Orthogonal validation · Knockdown and knockout disagree · The experiment was not replicated · Preclinical review

Review my manuscript