A custom AI reviewer for model-to-human relevance

PerfectPaper’s model-to-human relevance agent is a custom AI reviewer that checks whether a preclinical manuscript’s claims are supportable in the animal model actually used, flagging immune conclusions drawn in an immunodeficient host, microenvironment claims from a subcutaneous implant, and translational language that outruns the system it was demonstrated in. Copy the brief below into a custom agent slot to run it.

Model relevance is the objection that arrives late in review and is the hardest to answer, because the remedy is either a new experiment or a substantially rewritten discussion. A statistics comment can be answered in an afternoon with a re-analysis. A model comment asks whether the experiment you already finished was capable of producing the sentence you wrote about it, and no re-analysis reaches that.

The stakes are visible at the scale of the literature. Begley and Ellis, describing Amgen’s attempt to reproduce 53 landmark preclinical oncology papers in Nature (2012;483:531–533), reported that the scientific findings could be confirmed in 11% of them. Model choice is not the only reason a preclinical result fails to travel, but it is the reason that survives perfect data: an experiment can be flawlessly randomised, blinded and powered and still be run in a host that cannot contain the phenomenon its abstract describes. That is why this agent runs alongside, not instead of, in vivo experimental rigour.

When to use this agent

• Your conclusions concern immune infiltration, immune evasion, or immunotherapy response
• Tumours were implanted subcutaneously and the claim concerns the tumour microenvironment
• The work uses nude, SCID or NSG mice and the discussion talks about immunity
• Human cells were grown in a mouse and the conclusion is about human disease
• Your abstract uses the word translational
• A reviewer has already written that the model is not appropriate
• The manuscript reports a patient-derived xenograft and the discussion describes stromal or immune biology
• A mouse dose appears in the abstract described as clinically relevant or clinically achievable
• Cell line identity, authentication method or passage number is absent from the methods

What model-to-human relevance means

Model-to-human relevance is construct validity: whether the experimental system used could in principle generate the evidence the manuscript claims from it, before any question about how well that evidence was collected. Construct validity, internal validity and external validity fail in different places and are repaired differently, and conflating them is why model comments are so often answered with the wrong response.

Internal validity is whether the difference between groups was caused by the intervention rather than by allocation, measurement or analysis; it is repaired with randomisation, blinding and a pre-specified analysis. External validity is whether a result obtained in one valid system generalises to another population. Construct validity sits before both: it asks whether the system instantiates the biology named in the claim at all. A nude mouse experiment on T cell exhaustion has an internal validity problem only in the sense that a thermometer in an empty room has a precision problem — the measurement may be excellent and the referent is absent.

The practical consequence is that adding animals does not help, repeating the experiment does not help, and a tighter confidence interval makes the situation worse by lending precision to a claim the host cannot support. This is the same asymmetry that makes a mechanism-not-shown objection so hard to answer: both are objections to what the design can demonstrate, not to how well it was executed.

The host strain decides which claims are available

The host strain, not the tumour, sets the ceiling on immune claims, and the correct objection differs strain by strain. Nude mice are athymic and lack mature T cells while retaining B cells, natural killer cells, myeloid cells and complement. NSG mice, by contrast, lack mature T cells, B cells and NK cells, carry impaired macrophage and dendritic cell function, and sit on a NOD background that is deficient in haemolytic complement. A reviewer who writes “the mice are immunodeficient” is being imprecise; a reviewer who writes “NK-mediated antibody-dependent cytotoxicity cannot occur in an NSG host” has ended the argument.

Two mechanisms explain most of the table. The interleukin-2 receptor gamma chain deleted in NSG mice is shared by the receptors for IL-2, IL-4, IL-7, IL-9, IL-15 and IL-21, so removing it removes IL-15 signalling and with it natural killer cell development — this is why NSG hosts, unlike nude and SCID hosts, cannot support an NK-dependent readout. And the NOD background carries a signal regulatory protein alpha variant that binds human CD47, delivering the “do not eat me” signal to mouse macrophages and permitting human cell engraftment that other backgrounds reject. Engraftment efficiency and immune competence trade against each other by construction: the strain that accepts the human tumour best is the strain least able to say anything about immunity.

Humanised mice narrow the gap without closing it. Human T cells developing in an HSC-engrafted NSG mouse are selected on mouse thymic epithelium, so their HLA restriction does not match the donor’s, and human myeloid reconstitution is poor unless human cytokines are supplied transgenically. NSG-SGM3 — NOD.Cg-Prkdc^scid Il2rg^tm1Wjl Tg(CMV-IL3,CSF2,KITLG)1Eav/MloySzJ — exists precisely because human IL-3, GM-CSF and stem cell factor had to be added: the mouse orthologues do not do the job. The agent treats the existence of those transgenes as the evidence that species mismatch is real rather than theoretical.

Moving to an immunocompetent syngeneic host solves the compartment problem but creates a different one. CT26, MC38 and B16-F10 were established decades ago by chemical carcinogenesis or spontaneous transformation, and they carry mutation loads far above the average human tumour. Castle and colleagues sequenced CT26 and reported 1,688 non-synonymous point variations, a coding rate of 53 non-synonymous mutations per megabase against the roughly 4 per megabase they cite as the average for spontaneous human tumours — while noting in the same paper that this still sits inside the range for primary human colorectal tumours, which runs from under 1 to over 100 per megabase (BMC Genomics 2014;15:190). The same group found 962 non-synonymous somatic point mutations in B16F10, 563 of them in expressed genes (Cancer Research 2012;72:1081–1091). So the objection to state is not that these lines are more mutated than a patient’s tumour could be; it is that each line is one fixed point in a distribution patients span, selected over decades for growing reliably in its host.

Mutation load is also not the variable that decides the answer, which is why a syngeneic result cuts both ways. Lechner and colleagues profiled six murine models, CT26 and B16 among them, and found tumour MHC class I expression to be “a singular marker to predict immunotherapy response” with the regimens they tested (Journal of Immunotherapy 2013;36:477–489). A brisk response to checkpoint blockade in an antigenic, well-presenting line is partly a response to that line, so it is not by itself evidence that a low-mutational-burden human cancer — most prostate and pancreatic tumours, for instance — would respond the same way. A null result in a poorly presenting line is the mirror image: evidence that the therapy failed in a host with little to present, not that the mechanism is wrong. The agent reports the direction of the inference either way.

Common questions about immunodeficient mouse strains

Can I make immune claims using nude mice?

Innate ones, within limits. Nude mice lack mature T cells but retain B cells, natural killer cells and innate immunity, so claims about innate mechanisms may hold while claims about adaptive immunity or T cell mediated therapy do not. The agent names the specific compartment involved.

What’s the real difference between nude, SCID and NSG mice?

Nude mice lack mature T cells but keep B cells, NK cells and complement. SCID mice lack functional T and B cells, keep NK cells, and often turn “leaky” with age as endogenous lymphocytes appear — a real confound in longer studies. NSG mice lack T, B and NK cells and have impaired macrophage function; they do not leak.

I used NSG mice for my checkpoint inhibitor study — can reviewers say the results don’t hold?

Yes, and they usually will. A checkpoint inhibitor works by releasing an existing but exhausted T cell response, and NSG mice have no mature T cell compartment to release. A reviewer is more likely to write that the mechanism cannot be demonstrated in that host than that the data are wrong — the fix is a syngeneic or genetically engineered model, or a narrower mechanistic claim.

Implantation site: what a flank tumour cannot show

A subcutaneous flank implant reproduces subcutaneous tissue, and the environment a tumour microenvironment claim refers to is the tissue of origin. Stromal composition, resident immune populations, vascular architecture, interstitial pressure, oxygen tension, microbiome exposure and drug penetration all differ between the flank and the pancreas, brain, lung or bone marrow. A subcutaneous experiment is a legitimate and often correct choice for pharmacokinetics, target engagement, tolerability and relative growth inhibition. It is the wrong instrument for a sentence containing the phrase tumour microenvironment.

Metastasis is the sharpest case. Spontaneous metastatic spread from a flank implant is rare, so a subcutaneous model reports on primary growth and effectively nothing about dissemination. The 4T1 mammary carcinoma line is the standard demonstration: implanted in the mammary fat pad of a BALB/c mouse it metastasises spontaneously to lung, liver and bone, and the same cells in the flank do not reproduce that behaviour. A tail vein injection is an experimental metastasis assay that begins at the arrival step; it reports on extravasation and colonisation and says nothing about local invasion or intravasation. A manuscript that uses tail vein injection and writes about metastasis without qualification has described a different biological process from the one its readers will assume.

The agent therefore reports the site as a separate finding from the strain, because a paper can be immunologically clean and site-wrong. A syngeneic MC38 tumour in a fully immunocompetent C57BL/6 mouse still sits under the skin, and the tumour-infiltrating lymphocyte populations recovered from a flank tumour are not the populations that would be recovered from a colon tumour. The honest form of the sentence names the site: “in subcutaneous MC38 tumours” takes four words and removes the objection.

Moving the implant to the correct organ narrows the site objection without closing it completely. An orthotopically implanted cell line still arrives as a single-step bolus of already-transformed cells; it skips the field cancerization, sequential mutation and stepwise immune editing that a spontaneously arising tumour undergoes in its tissue of origin, which is why a genetically engineered model — where the tumour develops in place — is the stronger design for questions about how the microenvironment forms over time, not only what it looks like once the tumour is established.

Common questions about implantation site and the tumor microenvironment

Is a subcutaneous model acceptable for microenvironment work?

Flank implants are acceptable for many questions and poor for those that depend on organ context — stromal composition, tissue-specific immune populations, and metastatic tropism all differ from the orthotopic setting. The agent reports the mismatch where the claim depends on it.

Reviewer says my flank tumor doesn’t reflect the real tumor microenvironment — is that a fair comment?

Usually, yes. A subcutaneous flank implant sits in skin and fat, not in the stroma, vasculature or resident immune populations of the tumour’s actual tissue of origin, so a microenvironment claim drawn from it is a claim the site cannot support. The fix is rarely to dispute the comment — naming the site in the sentence (“in subcutaneous tumours”) usually resolves it.

If my tumor spread after a tail vein injection, can I call that spontaneous metastasis?

No. A tail vein injection puts tumour cells directly into circulation, so it models only the later steps of metastasis — extravasation and colonisation at a distant site — not the local invasion and intravasation “spontaneous” implies. Report it as an experimental metastasis assay, and reserve “spontaneous metastasis” for a model like orthotopic 4T1, where cells leave the primary site unaided.

Species mismatch: which signals do not cross

A human tumour in a mouse host is a two-species system in which tumour-derived signals are human and every stromal, vascular and immune signal reaching them is murine. Many ligand-receptor pairs are not cross-reactive, and where cross-reactivity exists it is frequently one-directional, so a xenograft is not simply a quieter version of a patient tumour — it is a system with specific circuits disconnected.

Named examples the agent looks for. Mice have no direct orthologue of human CXCL8, so a claim about IL-8 driven neutrophil recruitment in a xenograft is describing a chemokine the host cannot make; the mouse functional analogues, CXCL1 and CXCL2, are different proteins with different receptors. Bevacizumab neutralises human VEGF-A and not mouse VEGF-A, so in a xenograft it can only act on tumour-derived VEGF while the host-derived angiogenic supply continues — a result that will not scale to a patient in whom both compartments are human. Antibody effector claims fail at the host: antibody-dependent cellular cytotoxicity requires NK cells and Fc gamma receptor engagement, both absent or impaired in the strains used for engraftment, so a therapeutic antibody tested in an NSG host has been tested for its direct effects only.

The corresponding check on the reagent side is quick and frequently productive. Where a manuscript reports immune phenotyping in a xenograft, the agent asks which species each antibody clone recognises. Human CD8 staining in a xenograft counts cells that arrived with the graft or with a humanising engraftment; mouse CD8 staining in an athymic host should count almost nothing. A figure reporting large CD8 populations in nude mice usually indicates either a mislabelled antibody, an unrecognised leaky phenotype, or a legend that does not match the experiment, and it is worth resolving before submission rather than in a response to reviewers.

The mismatch also cuts both ways within a single experiment, which is what makes a xenograft non-response hard to interpret. A therapeutic antibody built for the human system engages mouse Fc gamma receptors with a different affinity and a different receptor hierarchy than it would engage human ones, even where its target-binding arm works, so a xenograft can simultaneously under-deliver the drug’s immune-effector function and under-represent the tumour’s human signalling repertoire. When a xenograft study reports no effect, the agent treats that as ambiguous between the drug failing and the model failing until the manuscript’s own data — target engagement, exposure, effector-cell presence — rule one of the two out.

Cell line provenance changes what the model is

Cell line identity is a model relevance question, not a housekeeping question, because a misidentified line changes the disease the experiment is about. MDA-MB-435 is the standard cautionary case: distributed for years as a breast carcinoma line, it is a derivative of the M14 melanoma line, it carries the melanoma-typifying BRAF p.Val600Glu mutation, and it appears in the International Cell Line Authentication Committee Register of Misidentified Cell Lines. Every conclusion about breast cancer drawn from it was a conclusion about melanoma.

Three items settle provenance in three sentences of methods, and the agent reports each absence briefly rather than lecturing: short tandem repeat authentication with the date it was performed and the reference profile it matched; passage number at the time of the reported experiments; and mycoplasma testing with its method and date. Mycoplasma belongs on that list for a biological reason, not a procedural one — contamination alters proliferation, metabolism and immune-relevant gene expression, so an untested line is an uncharacterised variable in every figure. A line received from a collaborator or a repository is not exempt: identity and mycoplasma status can both change during ongoing culture, so the authentication that matters is the one performed on the cells actually used for the reported experiments, not the one performed when the line first arrived in the lab. Nor is the problem confined to human lines. Schrörs and colleagues compared MC38 stocks held at two widely used sources and found only 35.3% of non-synonymous variants shared between them, along with different mutational signatures and different neoantigen repertoires (Frontiers in Immunology 2023;14:1102282) — so “MC38” with no source named is under-specified in any manuscript whose conclusion depends on how antigenic the tumour was.

Patient-derived xenografts carry a provenance problem of their own that is easy to miss. Ben-David and colleagues monitored copy number alterations in 1,110 PDX samples across 24 cancer types (Nature Genetics 2017;49:1567–1575) and found copy number alterations accumulating rapidly during passaging, often through selection of pre-existing minor clones, with several alterations recurrent in primary tumours gradually disappearing in the mouse. A late-passage PDX is therefore not automatically a closer model of the patient than a cell line; it is a model whose divergence is undocumented unless the manuscript reports the passage number and any genomic comparison to the originating tumour. Human stroma is also replaced by murine stroma within early passages, which returns the whole section to the species mismatch above.

Common questions about species mismatch and cell line identity

Why does species mismatch matter in a xenograft?

Because tumour and stroma communicate through ligand-receptor pairs, and many of those pairs do not function across species — the CXCL8/CXCL1 and human/mouse VEGF-A examples above are the concrete version of this. A human tumour in a mouse may not receive signals it would receive in a patient, which limits conclusions about tumour-stroma crosstalk specifically, not about the tumour’s own cell-intrinsic biology.

Can a patient-derived xenograft support a claim about the tumor microenvironment?

Not a human microenvironment claim. Human stroma is replaced by murine stroma within early passages, and the host strains used for engraftment lack the immune compartments a microenvironment argument usually invokes. Report the passage number and confine the claim to tumour-intrinsic biology.

Do I need STR authentication if I just got the cell line from a collaborator?

Yes — provenance does not transfer with the tube. A line can drift, be mislabelled, or be cross-contaminated between institutions, so a collaborator’s assurance is not a substitute for your own short tandem repeat profile matched against a reference database. Report the authentication date and the reference profile it matched, the same as for a line ordered directly from a repository.

Dose and schedule: the conversion a reviewer will do

A dose used in mice becomes a claim about clinical promise only when the manuscript states its relationship to a human dose, and the conversion is arithmetic a reviewer can do in ten seconds. The FDA guidance Estimating the Maximum Safe Starting Dose in Initial Clinical Trials for Therapeutics in Adult Healthy Volunteers (July 2005) gives body surface area normalisation factors in Table 1: the mouse km is 3 and the human km is 37, so a mouse mg/kg dose is divided by 12.3 to give the human equivalent dose in mg/kg for a 60 kg adult.

Body-surface-area conversion is a regulatory safety convention, not a claim of pharmacological equivalence, and the guidance is explicit about which of the two it is doing. Its introduction states that “the purpose of this process is to ensure the safety of the human volunteers”, and its Step 5, on the pharmacologically active dose — the dose at which the compound does something rather than the dose that is tolerated — says that “selection of a PAD is beyond the scope of this guidance”. The efficacy half of the question is not answered by the conversion; it is explicitly left outside it. Where pharmacokinetic data exist for the compound, an exposure-matched comparison — a human equivalent dose estimated from AUC or Cmax rather than from body surface area alone — is the more defensible basis for a clinical-promise claim, because two compounds with the same km-based conversion can differ severalfold in clearance, protein binding and half-life between mouse and human.

Run the calculation on your own abstract before someone else does. A tumour growth experiment at 100 mg/kg daily in mice implies a human equivalent dose near 8 mg/kg per day, about 490 mg daily for a 60 kg adult, before any safety factor is applied. Whether that is achievable depends on the compound’s exposure, tolerability and half-life, and the manuscript rarely says. The agent’s finding is not that the dose is too high; it is that a claim of clinical promise has been made with no stated relationship between the schedule tested and any schedule a person could receive. Route matters on the same terms: intraperitoneal dosing has no clinical counterpart for most small molecules, and a paper that reports intraperitoneal delivery and then discusses oral therapy has skipped a step. Depth on this specific objection sits in translational dose relevance, and the escalation-safety version of the argument in dose escalation and safety reporting.

The agent brief

The brief below is ready to paste into a custom agent. Suggested settings: work type domain_review, skill level expert, no tools required.

Name: Model-to-human relevance

You are reviewing whether the conclusions of a preclinical study are
supportable in the model system actually used. Read the methods to establish
the system, then read the abstract, results and discussion for claims that
exceed it.

Determine and report:

1. The system. State precisely what was used: cell line xenograft, patient
   derived xenograft, syngeneic allograft, genetically engineered model,
   humanised mouse, organoid, or co-culture. Identify the host strain and its
   immune status. Identify the implantation site where relevant.

2. Immune claims. Where the host is immunodeficient, identify which immune
   compartments are absent or impaired. Report every claim about immune
   infiltration, immune evasion, adaptive response or immunotherapy mechanism
   that cannot hold in that host. Be specific about which compartment is
   missing and which sentence depends on it. Note that nude mice retain B cells
   and innate immunity while lacking mature T cells, and that NSG mice lack
   substantially more, since the correct objection depends on the strain.

3. Site. Where tumours were implanted subcutaneously and the conclusion
   concerns the tumour microenvironment, stromal interaction, metastatic
   behaviour or organ-specific biology, report that the site does not
   reproduce the relevant environment.

4. Species mismatch. Where human tumour cells grow in a mouse host, identify
   claims that depend on interaction between tumour and stroma, since many
   ligand-receptor pairs are not cross-reactive between species. Report claims
   about cytokine or chemokine signalling that assume cross-species function.

5. Cell line provenance. Determine whether cell lines are authenticated,
   whether passage number is given, and whether mycoplasma testing is stated.
   Report each absence briefly rather than at length.

6. Dosing and schedule. Determine whether the treatment schedule bears a
   stated relationship to clinical use. Report claims of clinical promise
   where the schedule has no such relationship.

7. Language. Identify sentences whose strength exceeds the evidence,
   particularly in the abstract and the final discussion paragraph. For each,
   quote the sentence and propose a specific rewording that the model does
   support.

Where a claim is unsupportable, say what would be required to support it: a
different host, a different site, or a narrower sentence. Prefer the narrower
sentence where it preserves the finding.

Editing the brief is expected, and adding a numbered item is the usual way to extend it. Reviewers of vaccine and infection work commonly add an item on the microbiome and housing facility; reviewers of neuroscience work add one on strain background and substrain drift. How to write a custom reviewer agent covers the brief structure this one follows.

What this agent catches

How to find this in your own manuscript before a reviewer does

Model relevance defects surface through three passes over your own text, run in this order, usually within a single sitting. They are found by cross-referencing sections, which is exactly the reading that authors of a manuscript they have lived with for a year stop doing.

Pass one: write the system down in one sentence. From the methods alone, state the host strain with its nomenclature, its immune status, the implantation site, the cell or tissue source, and the species of every biological reagent. If that sentence takes more than one attempt to write, the methods do not specify the system, and the ARRIVE reporting items for experimental animals and procedures are the place to start fixing it.

Pass two: search the abstract and discussion for the claim vocabulary. The retrieval handles are short and specific: immune, T cell, CD8, exhaustion, evasion, checkpoint, microenvironment, stroma, metastasis, translational, clinically, patients, therapeutic window. Every hit is a sentence to check against the one-sentence system description from pass one.

Pass three: check the figure legends against the host. Legends are where model mismatch survives editing, because they are written early, copied between drafts and read last. The three recurring defects are a legend naming a strain that the methods do not use, a legend reporting an immune population the host does not have, and a legend whose n counts tumours where the methods count animals — the last being a pseudoreplication problem rather than a model problem, and worth separating in your own head before you answer it.

What a reviewer writes when the model is wrong

Reviewers rarely write “the model is not relevant”. They write the specific version, and the objection usually arrives in one of these forms.

“The authors conclude that the compound restores anti-tumour immunity, but the experiments were performed in athymic nude mice, which lack mature T cells; this conclusion cannot be drawn from these data.” “Subcutaneous implantation does not recapitulate the pancreatic tumour microenvironment; the claims regarding stromal remodelling require an orthotopic model.” “The human cell line was grown in a mouse host, so the reported tumour-stroma cytokine axis assumes a cross-species interaction the authors have not demonstrated.” “No cell line authentication is reported; please provide an STR profile and the date of testing.” “The dose used is not related to any clinically achievable exposure, and the phrase clinically promising in the abstract is not supported.” “The mechanism is proposed but not demonstrated in a system capable of showing it.” “The single cell line does not support generalisation to the disease named in the title.”

The last one is worth anticipating because it is answered badly more often than any other. A one-cell-line experiment does not become a two-cell-line experiment through argument, and a reviewer who asks for a second line is usually satisfied by either a second line or a title that names the line, which is the same negotiation described in what reviewers mean by more controls and in orthogonal validation.

What to do when the right model is unavailable

The narrower sentence is the remedy in most cases, and it is available immediately, without new funding, new animals or a new ethics approval. The agent is built to propose it first because it is usually both the honest answer and the fastest one.

When you cannot run the syngeneic repeat. State the compartment. “In athymic nude mice, which lack mature T cells, compound X reduced tumour growth, indicating a T cell independent component of activity” is a stronger sentence than a suppressed immune claim, because it converts a limitation into a specification about mechanism. Move the immune hypothesis into the discussion, labelled as a hypothesis, and name the experiment that would test it.

When you cannot run the orthotopic model. Name the site in every claim sentence and confine microenvironment language to what subcutaneous tissue supports. Where human tissue data exist — a public expression dataset, an immunohistochemistry series, a published cohort — a correlation between your mechanism and human tissue is an available and legitimate support, provided it is labelled as correlative rather than causal.

When you have no humanised model. Say what the xenograft demonstrates, which is usually direct effect on tumour cells in vivo, and say what it does not. An explicit statement of the disconnected circuits reads as command of the system; an omission reads as an oversight and draws the comment anyway.

When the reviewer asks for the experiment regardless. Answer in three parts: agree with the biological point, state the specific claim you are narrowing and quote the new sentence, and give the bounded reason the experiment is not being performed for this manuscript. Vague acknowledgement is what triggers a second round; a rewritten sentence quoted in the response letter usually ends it. The limitations paragraph is where the direction of the limitation belongs — which way the model would bias the result, not that results should be interpreted with caution.

Questions about answering reviewer 2 on model relevance

Do I have to repeat the experiment in a better model?

Usually not. Where a claim exceeds the model, the agent prefers to propose a narrower sentence that the data already support, because that is normally the honest remedy and the far quicker one. A repeat is warranted only when the narrowed claim would no longer be the paper’s finding.

Reviewer 2 says my mouse model doesn’t reflect human disease — what do I even say to that?

Say what the model does show before defending what it does not. Agree with the biological point, state the specific claim you are narrowing and quote the replacement sentence, then give a bounded reason the additional experiment is out of scope for this manuscript. A vague defence of the model invites a second round; a rewritten sentence usually ends the exchange.

How do I answer a reviewer who says the model is not appropriate?

Agree with the biological point, quote the narrowed sentence you have written in its place, and give a bounded reason the additional experiment is out of scope for this manuscript. A rewritten sentence in the response letter settles the comment more often than an argument about the model does.

How this differs from the built-in review

The standing team includes a conclusion-validity specialist that compares claims against the evidence presented in the paper. This agent asks a prior question: whether the experimental system could generate that evidence at all, however clean the data. A result can be internally consistent, statistically sound, and still unable to support the sentence written about it.

PerfectPaper reads the methods for the host strain and the implantation site first, then reads the abstract and discussion for claims that require a compartment the host does not have, and reports the specific sentence alongside the missing compartment. Pairs with the overclaim check, which catches the same sentences from the language side, and with in vivo experimental rigour, which checks whether the experiment that did run was well designed.

The remedy it proposes is usually a rewritten sentence rather than a new experiment, which is why it is most useful before submission. Full set: AI peer review for preclinical cancer research.

Review my manuscript

Related agents

Four neighbouring agents share a review with this one, and three custom agents run together on a single manuscript. See AI peer review for preclinical cancer research for the full set of five preclinical agents.

• Unit of analysis and pseudoreplication — use it when a figure legend counts cells, tumours or wells rather than animals.
• In vivo experimental rigour — use it for randomisation, blinding, exclusions and the tumour volume formula.
• Flow cytometry reporting — use it when a percentage has no named parent population.
• Image quantification integrity — use it when a representative image is standing in for a quantification.