AI peer review for clinical trial manuscripts

PerfectPaper reviews clinical and translational therapeutics manuscripts with a standing team of specialist agents, and lets you add up to three custom agents for conventions that team does not cover. This page carries five prebuilt briefs for experimental therapeutics: response criteria, early-phase safety, trial registration and endpoint pre-specification, drug synergy claims, and translational dose relevance. Each is text you paste into the agent field.

Early-phase oncology has a dense vocabulary of defined terms — dose-limiting toxicity, confirmed response, recommended phase 2 dose — where the definitions carry the meaning. Most reviewer objections in this literature are about definitions applied loosely rather than about analysis done wrongly. A response rate computed over the wrong analysis population is arithmetically perfect and still wrong. A dose-limiting toxicity count printed without the definition that produced it is a number with no unit. Neither defect is visible to a checker that only recomputes the percentages on the page.

The five agents

Response criteria discipline — RECIST and iRECIST conflated, unconfirmed responses counted, censoring rules absent, waterfall plots that disagree with the reported response rate.

Early-phase safety and dose reporting — dose-limiting toxicity window undefined, grading version unstated, safety denominators that shift between tables, recommended dose unsupported by the escalation data.

Trial registration and endpoint pre-specification — missing registration identifier, reported primary endpoint differing from the registered one, a secondary endpoint promoted to the headline.

Drug synergy claims — synergy asserted from a bar chart, no reference model named, combination index misapplied.

Translational dose relevance — in vitro concentrations far above achievable exposure, animal-to-human dose conversion without a stated basis, target engagement never shown at the effective dose.

What a clinical trial manuscript is reviewed against

A clinical trial manuscript is reviewed against four external documents that are not in the manuscript: the response criteria it names, the grading system it names, the reporting guideline its journal endorses, and its own registry record. The standing specialist team reads the paper against itself — structure, arithmetic, table consistency, figures, citations, statistics, causal inference. The five agents on this page read the paper against those four documents.

The reporting guidelines are specific and recent enough to matter. CONSORT 2010 is the parent statement, covered in the CONSORT checklist. Yap and colleagues published the CONSORT Dose-finding Extension, CONSORT-DEFINE, in BMJ 2023;383:e076387, adding 21 new checklist items and modifying 19 existing CONSORT items for early phase dose-finding trials — a standard that postdates the design of many dose-finding trials now reaching submission. RECIST 1.1 is Eisenhauer and colleagues, European Journal of Cancer 2009;45(2):228–247; iRECIST is Seymour and colleagues, Lancet Oncology 2017;18(3):e143–e152. Adverse event grading runs on the US National Cancer Institute’s Common Terminology Criteria for Adverse Events, where the National Cancer Institute has targeted implementation of CTCAE v6 for newly opened CTEP and DCP studies from January 1, 2026, so manuscripts submitted through 2026 and 2027 are split across CTCAE v5.0 and CTCAE v6 with non-identical grade boundaries.

None of these documents is quoted in a typical methods section. That is the gap the five agents fill: each finding names the clause, the dose level or the table, rather than reporting that something is unclear.

How custom agents work

PerfectPaper lets an author add up to three custom agents to a review. Each takes a name, instructions up to 6,000 characters, a work type, a skill level, and an optional tool selection — web search, scholarly search, code execution, and vision. PerfectPaper supplies the review contract and output format, so the brief describes what to look for.

The five briefs on this page each fit inside the 6,000-character limit with room to add your own items, and each names its suggested settings on its own page: work type domain_review for response criteria, safety and translational dose, work type verification for registration, skill level expert throughout. Tool choice is the setting authors get wrong most often. Code execution lets an agent recompute a percentage from its own printed numerator and denominator instead of judging arithmetic in prose. Web search is what allows the registration agent to open a registry record rather than only checking the manuscript against itself. Scholarly search is what allows the translational dose agent to find a reported human pharmacokinetic parameter for a named compound. Vision is what lets an agent count bars in a waterfall plot.

Writing your own variant, rather than pasting one of these, is covered in how to write a custom reviewer agent; the full catalogue across fields is at custom AI reviewer agents.

Choosing among them

For a first-in-human or dose-escalation paper, take early-phase safety and registration. For an efficacy report, take response criteria first, since that is where the definitional slippage lives. Synergy is worth a slot only when the paper claims synergy specifically rather than additivity or combination benefit. Translational dose relevance suits preclinical papers that argue for clinical development, and it is often the objection that decides whether that argument lands.

Give the safety and response agents code execution, since both benefit from recomputing rates and reconciling denominators.

Three slots and five agents means two are always left out, so choose by what the paper claims rather than by what it is about. A phase 1 report that also carries a dose-expansion efficacy cohort claims two things and needs safety, response criteria and registration. A combination paper running from cell lines to a xenograft claims synergy and development potential, so synergy and translational dose take two slots and registration is irrelevant. A randomised phase 3 with an independent data monitoring committee has its safety reporting handled by the trial itself; registration and response criteria are the slots that pay there, because outcome switching and censoring rules are what reviewers of a phase 3 actually challenge.

Definitions are the science in an early-phase paper

A dose-limiting toxicity is whatever the protocol says it is, which is why a count reported without the definition cannot be interpreted. A complete definition has four parts: a grade threshold, usually grade 3 or higher for non-haematological events; the organ systems in scope; the named exceptions; and the assessment window. The exceptions carry most of the between-trial variation — grade 3 nausea resolving to grade 1 within 48 or 72 hours on standard supportive care, grade 3 neutropenia lasting fewer than 7 days without fever, isolated lymphopenia, alopecia at any grade. Two trials of the same molecule with different exception lists derive different maximum tolerated doses from identical patient data.

The arithmetic downstream is unforgiving. Under a 3+3 design, 0 of 3 patients with a dose-limiting toxicity escalates, 1 of 3 expands the cohort to 6, 2 or more stops escalation, and the maximum tolerated dose is conventionally the highest dose at which no more than 1 of 6 patients had one. Whether a single grade 3 transaminase rise that resolved in four days counts therefore decides the dose the paper recommends.

Response criteria behave the same way. RECIST 1.1 permits a maximum of five target lesions and no more than two per organ, defines partial response as a decrease of at least 30% in the sum of diameters, and defines progressive disease as a relative increase of at least 20% and an absolute increase of at least 5 mm. The 5 mm floor is the clause most often dropped. A patient whose target sum reaches a nadir of 15 mm and then measures 19 mm has grown 26.7%, clearing the 20% threshold, but has grown only 4 mm, so has not progressed under RECIST 1.1. A methods section writing “progression was defined as a 20% increase” has described a more sensitive rule than the one it cites, and every survival curve downstream inherits the difference.

Immunotherapy adds a distinction RECIST 1.1 alone does not make. Seymour and colleagues’ iRECIST treats a first progression call as unconfirmed progressive disease, iUPD, rather than as progression outright, and requires a confirmatory scan, normally 4 to 8 weeks later, before it becomes confirmed progressive disease, iCPD — because a minority of patients on checkpoint inhibitors show pseudoprogression, an apparent increase in tumour burden that a later scan shows was not true growth. A manuscript reporting RECIST 1.1 progressive disease on a checkpoint-inhibitor arm without stating whether iRECIST confirmation was applied has made a provisional call read as a final one.

Common questions about definitions and denominators in early-phase reporting

Does a grade 3 event that resolves in a few days still count as a dose-limiting toxicity?

A dose-limiting-toxicity count depends on the protocol’s exception list, not on the grade alone. Many phase 1 protocols exclude grade 3 nausea resolving to grade 1 within 48–72 hours on standard antiemetics, or grade 3 neutropenia lasting under 7 days without fever, from the count. State which exceptions applied and cite the protocol clause, not just the CTCAE grade.

Is a 25% increase in tumor size progressive disease under RECIST 1.1?

Not by itself. RECIST 1.1 requires a relative increase of at least 20% in the sum of target-lesion diameters and an absolute increase of at least 5 mm from the nadir. A 25% increase measured from a 12 mm nadir is only 3 mm, so it clears the percentage threshold but misses the absolute floor and is not progression under the criteria as written.

Why does my response rate look different in the abstract than in Table 2?

A response rate changes with its denominator, not its numerator. The abstract may report responses over all treated patients while Table 2 reports the same responses over response-evaluable patients only, which excludes anyone who never had a post-baseline scan. State the analysis population next to every rate, in the abstract as well as the tables, so the figures reconcile instead of silently disagreeing.

Reviewer 2 says my sample size is too small — how do I answer that for a dose-escalation cohort?

Name the design, not just the count. A 3+3 or BOIN dose-escalation cohort is sized to characterise toxicity at each dose level, not to estimate a response rate precisely; state that 3 to 6 patients per level cannot support a rate reported to one decimal, and that any efficacy signal is hypothesis-generating pending a dedicated expansion or randomised cohort.

Denominators are where trial manuscripts actually fail

Denominator drift is the defect this cluster turns up most often, and it presents as ordinary arithmetic rather than as an error anyone would defend. Two constructed but entirely ordinary cases show the shape.

An efficacy report randomises 120 patients, doses 114, obtains a post-baseline scan in 98, and observes 34 responses. The response rate is 28.3% on intention to treat, 29.8% on all treated, and 34.7% on response evaluable. That is a spread of 6.4 percentage points from a single numerator, and the largest of the three is produced by excluding the patients who deteriorated before their first scan.

A first-in-human report enrols 34 patients, treats 32, and describes 28 as evaluable for dose-limiting toxicity after four replacements. A table headed treatment-emergent adverse events reports grade 3 or higher neutropenia in 7 patients, 21.9%. The abstract states 25%. Recompute: 7/32 is 21.9% and 7/28 is 25.0%, so the abstract silently switched from the treated population to the evaluable one. Neither figure is wrong alone, no reviewer sees the switch without doing the division, and two rates for the same event now enter the literature for a later meta-analysis to choose between.

Precision is the second half of the same problem. With three to six patients per dose level, one event moves a rate by 17 to 33 percentage points, so a percentage carried to one decimal claims a precision six patients cannot support. Report counts with the denominator in the row or the caption.

A fourth population complicates this further. Many oncology protocols also define a modified intention-to-treat, mITT, population — commonly all randomised patients confirmed eligible on central review, excluding those randomised in error — which sits between intention-to-treat and all-treated and is not one of the three populations above by default. A rate labelled only “mITT” without the protocol’s own definition of that population is as uninterpretable as one with no population stated at all, because different protocols build different exclusions into the same three letters.

Registration and pre-specification is the check an editor runs first

Outcome switching is the finding most likely to change what a trial paper is allowed to claim, and it is measured rather than suspected. Mathieu, Boutron, Moher, Altman and Ravaud compared 147 adequately registered trials against their registry records in JAMA 2009;302:977–984 and found discrepancies between the registered and the published primary outcome in 31% of them; where the direction could be assessed, the change favoured a statistically significant result in 82.6% of cases, 19 of 23.

The COMPare project measured the same thing prospectively in the five journals least likely to have the problem. Across 67 trials published in the New England Journal of Medicine, The Lancet, JAMA, the BMJ and Annals of Internal Medicine, a mean of 76% of pre-specified primary outcomes were correctly reported, trials carried a mean of 5.4 undeclared additional outcomes each, and 58 of the 67 — 87% — had discrepancies warranting a correction letter (Trials 2019;20:118).

Two mechanical checks come before any of that. The ICMJE recommends that journals publish the trial registration number at the end of the abstract, so an identifier sitting only in a funding statement is functionally missing during technical check. And prospective registration has two different lines: the ICMJE recommends registration “at or before the time of first patient consent for enrollment”, while United States regulation at 42 CFR § 11.24(a) allows 21 calendar days after the first participant is enrolled. A sponsor who files on day 20 has met the regulation and missed the journal policy. That 21-day allowance itself applies only to what the regulation defines as an applicable clinical trial; 42 CFR Part 11 excludes small feasibility device studies and, for drugs, trials described as phase 1, so a first-in-human dose-escalation study can sit outside the federal deadline entirely and answer to the journal’s ICMJE-based policy alone.

Questions about trial registration and endpoint changes

Why does a journal want the NCT number in the abstract and not just the methods section?

The ICMJE recommends the trial registration number appear at the end of the abstract specifically so a technical editor can check it during the initial screen, before the manuscript reaches peer review. A registration number sitting only in a funding statement or supplementary file is functionally invisible at that stage and can generate a request for revision before review even starts.

Is it a problem if my primary endpoint changed after I registered the trial on ClinicalTrials.gov?

It is answerable, not disqualifying, if you disclose it. Mathieu and colleagues found discrepancies between the registered and published primary outcome in 31% of trials, and where a direction could be assessed, 82.6% favoured a significant result — which is why an undisclosed switch reads as outcome-driven. State the change, the date it was made, and the reason, in the methods section itself.

Do I have to register a phase 1 dose-escalation trial within 21 days like a phase 2 trial?

Only if it counts as an applicable clinical trial under 42 CFR Part 11, and many phase 1 drug studies do not, since the regulation excludes trials described as phase 1. Where the federal deadline does not apply, the registration obligation that remains is the journal’s own ICMJE-based policy, which most editors still enforce at submission regardless of the regulatory question.

Synergy and translational dose are the two claims that outrun their evidence

Synergy is excess over a named reference model, and “better than either agent alone” is the weakest such model rather than a substitute for one. Take a figure in which drug A alone gives 30% killing, drug B alone gives 25%, and the combination gives 60%. Under highest single agent the expectation is 30%, so the combination clears it. Under Bliss independence the expectation is 0.30 + 0.25 − (0.30 × 0.25) = 0.475, so 60% clears it by 12.5 percentage points. Under Loewe additivity the expectation cannot be computed at all, because the figure contains no dose-response curve and therefore no iso-effective doses to put in the denominators. Two answers and one undefined, from one figure — which is why naming software is not naming a model. The most common software computing a single excess-over-expectation number is the Chou-Talalay median-effect method, which reports a combination index below 1.0 as synergy, at 1.0 as additivity and above 1.0 as antagonism — but the index itself requires a full dose-response curve for each single agent across the same effect range as the combination, not a reading taken from one combination point.

Translational dose relevance produces a shorter objection that is harder to answer: the effect requires ten micromolar and the achievable plasma concentration is two hundred nanomolar. The recurring technical error underneath it is comparing a total plasma concentration against an unbound in vitro concentration, which overstates achievable exposure by an order of magnitude or more for a substantially protein-bound compound. The underlying pharmacokinetic principle is that only the unbound fraction of a compound can engage its target, so a total-versus-unbound comparison is not merely imprecise, it is systematically biased toward the drug. Where both numbers exist, state the ratio. Where the effective concentration exceeds achievable exposure by roughly an order of magnitude, that ratio is the central limitation of the development argument and belongs in the paper rather than in the reviewer’s report. Whether the animal experiment behind it was randomised, blinded and adequately powered is a separate question, handled by in vivo rigour and the rest of preclinical review.

Common questions about synergy and translational dose claims

Is “better than either drug alone” enough to call two drugs synergistic?

No — that describes highest single agent, the weakest reference model, not synergy. A combination beating both monotherapies can still fall short of, meet, or exceed the additive expectation from Bliss independence or Loewe additivity; only clearing one of those named models, with the model stated, supports a synergy claim rather than a combination-benefit claim.

Reviewer says my in vitro concentration isn’t clinically achievable — what do I do?

Check whether you compared total plasma concentration against an unbound in vitro concentration; for a highly protein-bound compound that comparison overstates achievable exposure by an order of magnitude or more. Recompute using unbound concentrations on both sides, state the ratio between effective and achievable exposure, and frame the finding as mechanistic if a large gap remains.

How to detect these in your own manuscript before submitting

Five checks find most of what these agents find, and each takes under an hour with the manuscript open beside its registry record.

Recompute every percentage against its own printed numerator and denominator. Not the ones you doubt — all of them, including the abstract. The defect you are looking for is two correct percentages computed over different populations.

Name the analysis population next to every rate. Intention to treat, all treated, response evaluable, dose-limiting-toxicity evaluable. If any rate in the paper has no stated population, it has no meaning yet.

Copy the registry record’s primary outcome verbatim into a scratch file, then your methods sentence beneath it, then your abstract’s first result sentence beneath both. Compare measure, metric, timepoint and population across all three. A different timepoint on the same instrument is an endpoint change, and it is the version authors most often fail to recognise as one.

Count the bars. Every bar below the −30% line on a waterfall plot is a patient meeting the target-lesion criterion for partial response, so the count and the reported response rate must reconcile or the caption must explain why not.

Read your own discussion for the four words that need a definition behind them: synergy, clinically relevant, maximum tolerated dose, and confirmed. Each is a technical term with a specific requirement, and each is used loosely in most first drafts. That is also what overclaim checking and causal language discipline look for in the standing review.

What reviewers write when this is mishandled

Reviewers rarely name the class of defect. They write the specific version, and these are the comments that hold therapeutics papers in revision.

“Please state which version of RECIST was used and whether responses were confirmed.” “The response rate in Table 2 is computed over response-evaluable patients while the abstract reports it as the rate in the treated population; please reconcile.” “Dose-limiting toxicity is reported but never defined; please state the grade threshold, the organ systems, the exceptions and the assessment window.” “The adverse event grading version is not specified.” “The recommended phase 2 dose is stated without pharmacokinetic or pharmacodynamic support.” “The abstract reports a rate that I cannot reproduce from any table.” “The reported primary outcome differs from the registered one; please explain when and why it changed.” “Figure 3 demonstrates that the combination outperforms both monotherapies, which is not the same claim as synergy.” “The comparison of survival between responders and non-responders is subject to immortal time bias; a landmark analysis is required.” “Drug B has no measurable single-agent effect at 100 nM; how is the additive expectation defined in that regime?”

Almost every one of these is answerable from data the authors already hold, which is the argument for finding them before submission rather than answering them afterwards. The exceptions are the recommended-dose objection and the achievable-exposure objection, which sometimes cannot be answered at all.

When the obvious remedy is unavailable

Most findings in this cluster arrive after the data are fixed, and the remedy is then to bound the claim rather than to restate it.

Central review was not performed and the scans cannot be re-read. State that assessment was investigator-determined, name the criteria and version, and state the expected direction. Investigator assessment runs optimistic in a measurable direction, and in a randomised comparison the optimism largely cancels between arms; in a single-arm trial nothing cancels, because the investigator-assessed rate is the result.

No pharmacokinetic sampling was performed in the escalation. The recommended dose cannot then be justified on exposure, and the honest form is a recommendation labelled provisional with a named route to confirming it.

The protocol never specified a dose-limiting toxicity definition tighter than the one reported. State what was applied in practice, state that it rested on investigator judgement, and use “highest dose administered” where “maximum tolerated dose” is not accurate.

The dose matrix for a synergy claim does not exist. Report excess over Bliss as a descriptive quantity with its numeric value where fractional effects for both agents are already in the manuscript, or drop the word synergy and claim combination benefit, which the data do support.

The effective concentration exceeds achievable human exposure. That limits the translational claim, not the mechanistic finding. Report the comparison with its ratio and frame the work as mechanistic. A limitations section that names the direction and magnitude of each defect is stronger than one that lists them.

Censoring depended on how the patient was doing. Report the observed distribution of assessment intervals rather than the intended schedule, and treat it as a competing risks and censoring problem rather than an administrative footnote.

What this page is not for

The five agents on this page review original research articles reporting clinical or translational therapeutics studies. They are not the right agents for a systematic review or meta-analysis, for an observational cohort or registry analysis, for a qualitative study, or for a protocol paper with no results. An epidemiological cohort belongs with the epidemiology agents; a purely mechanistic bench paper making no development argument needs none of these five.

The oncology dependency is worth stating precisely, because two of the five carry it and three do not. Response criteria and dose escalation rest on RECIST and on oncology dose-finding conventions, so they transfer poorly outside solid tumours and haematological malignancy. Registration, synergy and translational dose apply across therapeutic areas without modification.

PerfectPaper is expert-led and AI-assisted: the review it produces is a pre-submission read of a manuscript, not a substitute for journal peer review, and none of these agents can re-read a scan, open a sealed protocol, or verify a number that was never recorded. What they do is find the findings a reviewer would raise while you can still act on them. The general case for the standing team is at AI peer review.

Frequently asked questions

What does AI peer review check in a clinical trial manuscript?

PerfectPaper checks a trial manuscript against the four documents outside it: the response criteria it names, the adverse event grading version it names, the reporting guideline its journal endorses, and its registry record. On top of the standing review of structure, statistics and arithmetic, the custom agents on this page check the definitional and reconciliation failures specific to therapeutics — confirmed responses, dose-limiting toxicity definitions, analysis populations, and outcome switching.

Can I add my own reviewer agents to a trial manuscript review?

Yes, up to three. They run alongside the standing specialist team, which already covers statistics, methodology, causal inference, tables, figures and citation verification, so a custom agent should target a convention the standing team does not — a named reporting guideline, a field-specific definition, or a registry check that requires web search.

Are these agents specific to oncology?

The response criteria and dose escalation agents are oncology-specific, since they rest on RECIST and on oncology dose-finding conventions, so they transfer poorly outside solid tumours and haematological malignancy. Registration, synergy and translational dose apply across therapeutic areas without modification, because none of the three depends on a tumour-specific response or toxicity-grading system.

Related

Response criteria · Early-phase safety · Registration and endpoints · Drug synergy · Translational dose · Custom AI reviewer agents

Review my manuscript