A custom AI reviewer for drug synergy claims

PerfectPaper’s synergy agent reads a combination study for the most common overstatement in the therapeutics literature: describing a combination as synergistic when the evidence shows only that it works better than either agent alone. The agent checks whether a reference model is named, whether the experimental design can support the calculation, and whether the word synergy is used technically or loosely.

Synergy is a defined concept requiring a reference model for expected additive effect. A combination that outperforms both single agents may be synergistic, additive, or even antagonistic — the bar chart cannot distinguish them.

Copy the brief below into a custom agent slot.

Why “better than either agent” is not synergy

Synergy is excess over a reference model, and “better than either agent alone” is the weakest such model rather than a substitute for one. Three reference models cover nearly all published combination work, and each specifies a different arithmetic for the effect expected when two agents do not interact.

Highest single agent (HSA). The expected effect is the larger of the two single-agent effects. A combination that exceeds it has shown only that adding the second agent added something.

Bliss independence. Chester Bliss set this out in “The toxicity of poisons applied jointly” (Annals of Applied Biology 1939;26:585–615). For fractional effects E_A and E_B on a 0–1 scale, the expected combined effect is E_A + E_B − E_AE_B, the probability that at least one of two independent events occurs. Bliss assumes the agents act through independent mechanisms.

Loewe additivity. Loewe and Muischnek defined it by dose equivalence in “Über Kombinationswirkungen” (Archiv für Experimentelle Pathologie und Pharmakologie 1926;114:313–326): doses d_A and d_B are additive when d_A/D_A + d_B/D_B = 1, where D_A and D_B are the single-agent doses producing the same effect. Loewe is the model that passes the sham-combination test: a drug combined with itself must come out additive, and Loewe delivers that where Bliss generally does not.

Greco, Bravo and Parsons compared the competing formalisms in “The search for synergy: a critical review from a response surface perspective” (Pharmacological Reviews 1995;47:331–385), and Berenbaum’s “What is synergy?” (Pharmacological Reviews 1989;41:93–141) remains the standard statement of why the question has no model-independent answer. The consequence for a manuscript is narrow and decisive: with no named model there is no expected value, and no result can exceed it.

When to use this agent

• Your manuscript describes a drug combination
• The abstract or discussion uses the word synergy or synergistic
• Your evidence is a three-bar comparison of drug A, drug B, and both
• You computed a combination index and are unsure whether the design supports it
• A reviewer has previously asked which synergy model you used
• You analysed a dose matrix in SynergyFinder or CompuSyn and reported one model’s score
• Your combination result is in vivo and the discussion calls it synergy

The agent brief

The brief below is what PerfectPaper’s synergy agent runs against a manuscript; paste it into a custom agent slot to run the same check yourself. Suggested settings: work type analysis, skill level expert, tools code execution.

Name: Drug synergy and combination claims

You are reviewing claims about drug combinations. Your task is to determine
whether each claim of synergy is supported by evidence capable of establishing
synergy, as distinct from evidence of combination benefit.

Determine and report:

1. The claim. Identify every use of synergy, synergistic, potentiation and
   similar terms, with location. For each, determine whether the term is used
   technically or as a loose synonym for better. Report loose usage and
   propose a precise alternative such as greater than either single agent.
   Treat this as the primary finding of this brief.

2. Reference model. For every technical synergy claim, determine whether a
   reference model for expected additive effect is named: Bliss independence,
   Loewe additivity, highest single agent, or Chou-Talalay median effect.
   Report claims made with no reference model. State plainly that without one
   there is no definition of the additive expectation the result exceeds.

3. Design adequacy. Determine whether the experimental design supports the
   named model. A combination index from the Chou-Talalay method requires dose
   response data for each single agent and for the combination at fixed ratio.
   Bliss independence requires fractional effects for each agent. Report
   claims computed from single-dose data, and report a full dose matrix
   analysed only at one dose pair.

4. Combination index. Where a combination index is reported, determine whether
   the effect level it corresponds to is stated, since the index varies with
   effect level. Report values given without one. Report indices reported
   without any measure of uncertainty.

5. Consistency of conclusion. Determine whether the synergy conclusion holds
   across the doses and cell lines tested, or only at selected points. Report
   conclusions drawn from the most favourable condition where others are shown
   and disagree.

6. Single agent activity. Determine whether single agent activity is shown at
   the doses used in combination. Where one agent has no effect alone at that
   dose, note that several models behave differently in that regime and that
   the choice of model then determines the answer.

7. In vivo claims. For animal combination experiments, determine whether a
   synergy claim rests on tumour growth curves alone. Report that group
   comparison of curves shows combination benefit rather than synergy unless
   a formal model is applied.

Use code execution to check any reported index or fractional effect against
the underlying numbers where the manuscript provides them.

Where synergy is not supported, propose the accurate claim the data do
support rather than recommending removal of the finding.

What this agent catches

Worked example: the three-bar figure

The three-bar figure is the most common evidence offered for synergy and the least capable of establishing it. Take a figure where drug A at 1 μM reduces viability 30%, drug B at 500 nM reduces it 25%, the combination reduces it 60%, and the legend reads “A and B act synergistically (p < 0.001)”. Three reference models give three different verdicts on those same three bars.

Under HSA the expectation is 30% and the observation is 60%, so the combination clears the reference. Under Bliss the expectation is 0.30 + 0.25 − (0.30 × 0.25) = 0.475, so 60% clears it by 12.5 percentage points. Under Loewe the expectation cannot be computed at all: the figure contains no dose-response curve, so there is no D_A or D_B at the 60% effect level to put in the denominators. Two answers and one undefined, from one figure.

The p-value contributes nothing to this question. It tests the combination bar against a single-agent bar, which is the combination-benefit comparison. No test between two bars can compare an observation against a reference-model expectation, because that expectation is not one of the bars — it is a number computed from two of them. Significance and reference-model excess are different quantities, and treating the first as evidence for the second is the substitution the agent reports, alongside the broader overclaim check.

The accurate claim those data support is that the combination reduced viability more than either agent alone, by 30 percentage points over drug A. That sentence is publishable, defensible, and no less important than the one it replaces.

Common questions about synergy versus combination benefit

If my drug combination beats each drug alone, can I just call it synergy?

No — that comparison only shows combination benefit. Highest single agent (HSA), the weakest reference model, already asks whether the combination exceeds the better single agent, so clearing it is a precondition for synergy, not a demonstration of it. Bliss independence or Loewe additivity, computed from the same numbers, can still call the identical result additive or even antagonistic.

What is the difference between synergy and additivity?

Additivity is the effect expected from combining two agents given their individual effects, defined by a reference model. Synergy is an effect exceeding that expectation. Without a model there is no expectation, so there is nothing for the result to exceed.

Is combination benefit a weaker finding?

Not necessarily. For a therapeutic question, whether the combination works better is often the more relevant claim, and it is one the data usually support. A fixed-dose combination can be approved under 21 CFR 300.50(a) on evidence that each component contributes to the effect, with no synergy claim required at all. The problem is only the mismatch between the word and the evidence.

The combination index, and what makes one interpretable

A combination index is interpretable only when three things travel with it: the effect level at which it was computed, the design that produced it, and a measure of uncertainty. The index comes from the Chou-Talalay median-effect method (Chou and Talalay, Advances in Enzyme Regulation 1984;22:27–55), whose protocol Chou restated in Cancer Research (2010;70:440–446), fitting f_a/f_u = (D/D_m)^m to each single agent and to the combination.

CI below 1 is read as synergy, CI of 1 as additivity, CI above 1 as antagonism. The number is not a constant. It moves with f_a, the fraction affected, so “CI = 0.7” with no effect level attached is uninterpretable — the same combination can be synergistic at f_a = 0.90 and antagonistic at f_a = 0.50. Report the index at stated effect levels, conventionally f_a = 0.50, 0.75 and 0.90, or as a CI–f_a plot across the measured range.

The method emits its own diagnostics: D_m, the median-effect dose; m, the shape parameter, where m = 1 is hyperbolic and m > 1 sigmoidal; and r, the linear correlation coefficient of the median-effect plot. A manuscript that reports CI without D_m, m and r has published the conclusion and withheld the fit. The method is also contested in the peer-reviewed literature rather than only by reviewers: Ashton published a critique as a letter in Cancer Research (2015;75:2400), and Wooten and colleagues, deriving the MuSyC framework, report a mathematical error in the original derivation that biases the index whenever the single-agent slopes depart from m = 1 (Nature Communications 2021;12:4607). That is a second reason to publish m: it tells a reader whether the reported CI sits in the regime where the criticism bites.

The index also takes one of two forms, and a manuscript rarely states which it used. Chou and Talalay’s generalised equations define a mutually-exclusive case, for agents with a similar mode of action, and a mutually-nonexclusive case for agents acting independently, which adds the cross term d_Ad_B/(D_AD_B) (European Journal of Biochemistry 1981;115:207–216). That term is positive, so the nonexclusive form always returns the larger CI from identical data and the two cases can return opposite verdicts around CI = 1. In practice the case is usually settled by default rather than by argument: Chou recommends the mutually-exclusive form (Pharmacological Reviews 2006;58:621–681) and CompuSyn implements it, so a bare CI is most likely the exclusive one. Naming the case, and the mechanistic reason for it, is a one-sentence addition to the methods that removes the ambiguity rather than leaving a reader to assume the default.

Design constrains the arithmetic. The constant-ratio design requires the combination to be dosed at a fixed ratio of A to B across a dose range, so that a median-effect plot exists for the mixture. A CI computed from one dose of A plus one dose of B has no such plot behind it, whatever the software returns.

Common questions about the combination index

Why did my combination index change when I picked a different effect level?

Because the Chou-Talalay combination index is a function of f_a, the fraction affected, not a single fixed number. The same drug pair can compute to CI = 0.6 (synergy) at f_a = 0.90 and CI = 1.3 (antagonism) at f_a = 0.50. Report the index across the measured range — conventionally at f_a = 0.50, 0.75 and 0.90, or as a CI–f_a plot — rather than at one arbitrarily chosen point.

Is a combination index of 0.7 automatically synergy?

Only with the information that makes it interpretable attached: the effect level it was computed at, a dose-response design for each single agent and the combination at fixed ratio, and a measure of uncertainty. A bare “CI = 0.7” with none of those attached is a number without an interpretation, not evidence of synergy on its own.

Do I need a combination index to publish a combination study?

No. A combination study reporting benefit over each monotherapy needs no combination index, because it makes no synergy claim. The index becomes necessary only when the manuscript uses the word synergy technically, and then the Chou-Talalay requirements — dose-response data for each single agent and for the combination at fixed ratio — apply in full.

When the reference models disagree

Reference models disagree on the same dose-response matrix routinely, which is why the model belongs in the protocol rather than in the analysis. Bliss and Loewe answer different questions — Bliss asks whether the agents behave as independent probabilistic events, Loewe asks whether one behaves as a dilution of the other — and one matrix can be Bliss-synergistic and Loewe-additive at once without either result being wrong.

SynergyFinder 3.0 (Ianevski, Giri and Aittokallio, Nucleic Acids Research 2022;50:W739–W743) makes that disagreement visible by design: it computes Bliss, Loewe, HSA and ZIP, the zero interaction potency model of Yadav and colleagues (Computational and Structural Biotechnology Journal 2015;13:504–513), on the same matrix and reports the four scores side by side. Its authors describe ZIP as an integration of the Bliss independence and Loewe additivity models: it assumes non-interacting drugs leave each other’s fitted dose-response potency unshifted, an expectation that factorises into the Bliss product and that also returns zero for a sham combination, which is the property Loewe is justified by. A manuscript reporting a ZIP score alone has therefore not avoided a modelling choice — it has made one carrying both assumptions at once. A manuscript that ran the tool therefore has four numbers in hand, and the methods section should say which of them the claim rests on and why.

The failure mode the agent looks for is model selection after the result: four scores computed, the largest reported, the other three absent from the methods. The remedy is to report every model computed, or to pre-specify one and state the mechanistic reason. Where one agent is inactive alone at its combination dose, the models diverge hardest — HSA has almost nothing to exceed, Bliss expects the active agent’s effect unchanged, and Loewe is undefined because no D_B produces that effect. Brief item 6 exists for exactly that regime.

Questions about choosing a synergy model

Which synergy model should I use?

Bliss independence suits agents with independent mechanisms; Loewe additivity suits agents acting on the same target; highest single agent is the weakest but simplest. The agent reports the absence of a named model rather than choosing one, since the choice depends on mechanism.

Bliss says synergistic and Loewe says additive for the same data — which one is right?

Both can be correct at once, because they test different null hypotheses. Bliss asks whether the two agents behave as independent probabilistic events; Loewe asks whether one agent behaves as a dilution of the other. A drug pair with independent mechanisms can clear the Bliss threshold while sitting on the Loewe additive line — report which model the claim rests on and why, not the more favourable of the two.

Synergy in vivo and in the clinic

An in vivo tumour growth experiment establishes combination benefit; it establishes synergy only when a reference model is applied to it. Four arms — vehicle, A, B, A+B — with the combination curve lowest is a group comparison, and the usual analysis, a repeated-measures ANOVA or a mixed model on log tumour volume, tests differences between arms rather than excess over an expectation.

Formal in vivo synergy analysis does exist. Demidenko and Miller set out a statistical test for Bliss-defined drug independence applied to tumour volume data (PLoS ONE 2019;14:e0224137), in which the single-agent arms define the independent-action expectation and the departure from it is the estimand. Running it requires all four arms and enough animals per arm to estimate that expectation with usable precision, which is the same in vivo rigour question of randomisation, blinding and per-arm n that governs the rest of the experiment, and the same preclinical reporting standards.

Clinically, the assumption that combination benefit implies drug interaction has been tested and found unnecessary. Palmer and Sorger showed in Cell (2017;171:1678–1691) that patient-to-patient variability in which single agent works can produce clinical combination benefit with no additivity or synergy at all: independent drug action is sufficient to explain it. A discussion section reasoning from a phase III survival advantage back to mechanistic synergy is making an inference that paper undercuts.

Regulators keep the two questions apart. 21 CFR 300.50(a) permits a fixed-combination prescription drug “when each component makes a contribution to the claimed effects” — contribution, not synergy — and the FDA’s June 2013 final guidance Codevelopment of Two or More New Investigational Drugs for Use in Combination frames the evidentiary task as demonstrating each drug’s contribution, consistent with that regulation.

How an unsupported synergy claim is detected in a manuscript

An unsupported synergy claim is detected by reading the figure legends against the methods, not by reading the statistics. Four checks find most of them.

Search the text for the word. List every occurrence of synergy, synergistic, synergism, potentiate and potentiation with its location. In many manuscripts the technical word appears only in the title and abstract while the results section says, accurately, “greater than either agent alone”.

Find the model in the methods. If no methods sentence names Bliss, Loewe, HSA, ZIP or Chou-Talalay, then no synergy claim in the paper has a reference. Naming software is not naming a model: “analysed in CompuSyn” or “analysed in SynergyFinder” identifies a tool, and SynergyFinder 3.0 alone computes four.

Check the design against the model’s stated requirements. Chou-Talalay needs dose-response data for each single agent and for the combination at fixed ratio. Bliss needs fractional effects for both agents at the combination doses. Loewe needs iso-effective doses, which means curves rather than points. Count the dose levels in the methods section and compare against that list.

Check consistency across every condition run. A 3 × 3 dose matrix in three cell lines yields 27 estimates. Where the abstract’s claim rests on one of them, the other 26 are the sensitivity analysis and they were already performed. Reporting the most favourable condition shares its structure with selective reporting under multiple testing, and the remedy is the same: show them all.

What reviewers say when a synergy claim is unsupported

Reviewers rarely write “this is not synergy” as their first sentence. They write the specific version, and these are the comments that force a revision.

“By which model is synergy defined here? Please state Bliss, Loewe or highest single agent in the methods.” “The combination index is reported without the effect level at which it was calculated.” “The Chou-Talalay method requires dose-response data for each single agent; the manuscript reports a single concentration of each.” “Figure 3 demonstrates that the combination outperforms both monotherapies, which is not the same claim as synergy.” “Synergy is reported in one of the three cell lines tested; please report the combination index for all three.” “No measure of uncertainty is given for the combination indices.” “Drug B has no measurable single-agent effect at 100 nM; how is the additive expectation defined in that regime?” “The in vivo experiment supports combination efficacy; a synergy claim would require a dose-response design.”

The last of those most often survives into a second round, because the requested experiment is substantial and authors reply that the claim is conventional in their field. It is conventional, and the comment still holds. The adjacent objection, that a mechanism is asserted rather than demonstrated, is covered in reviewer says the mechanism is not shown.

Frequently asked questions

Can an AI reviewer check whether a synergy claim is supported?

Yes, on the parts of the question the manuscript answers in writing. The agent finds every synergy claim, checks the methods for a named reference model, checks the reported design against that model’s data requirements, and reports each claim that fails one of those tests. Judging whether the chosen model suits the pharmacology remains the authors’ argument.

Can I claim synergy from an in vivo experiment?

Only with a formal analysis. Comparing tumour growth curves across single agent and combination arms demonstrates combination benefit; establishing synergy means testing the observed combination against an expectation computed from a named reference model. That does not always require new dosing. Demidenko and Miller’s 2019 PLoS ONE method covers the longitudinal in vivo case directly, deriving a Bliss independence expectation from the single-agent arms of a four-arm tumour volume experiment, so the analysis is often available on data already collected. A Loewe-based or Chou-Talalay claim is the one that needs a dose-response design, which most efficacy studies do not include.

My reviewer says the tumour curves don’t prove synergy — what do I do now?

State the claim the curves actually support: combination benefit over each single agent, from the group comparison you already ran. If the reviewer specifically wants a synergy claim, that requires re-running the analysis with all four arms — vehicle, A, B, A+B — against a reference model such as Bliss, which a tumour-growth-curve comparison alone cannot supply.

How do I know whether my synergy claim will survive review?

Check three things before submission: that the methods name a reference model, that the design meets that model’s stated data requirements, and that the claim holds across every dose pair and cell line reported rather than at the best one. A claim passing all three survives the pharmacology objection.

What to do when the dose matrix is not available

When the design cannot support a synergy calculation, change the claim rather than searching for a model that flatters the data. Four options, in ascending order of effort.

Rewrite the claim to the measurement. Replace “synergistic” with what was observed: “greater than either agent alone”, “reduced viability by 60% against 30% for drug A and 25% for drug B”. This requires no new experiment and removes the objection completely. It is what the agent proposes by default.

Report the weakest defensible model, and name it. Where fractional effects exist for both single agents at the combination doses, Bliss excess is computable from data already in the manuscript and can be reported as excess over Bliss with its numeric value, framed as a descriptive quantity rather than a verdict.

Analyse the matrix you already ran. Many combination papers contain a full dose matrix that was analysed at a single dose pair. Fitting all of it is a reanalysis rather than a new experiment, and it converts the strongest objection on the list into a figure.

Add a matrix for one system. A single cell line dosed as a 6 × 6 constant-ratio matrix, in triplicate, produces a defensible combination index across the effect range and is a smaller undertaking than most authors expect before scoping it.

Bound what remains unresolved. A limitations sentence naming the reference model that would be required and stating that the present design cannot support it is stronger than a synergy claim an editor will strike. Naming the missing design is a specification; “further work is needed” is not.

What this agent does not do

The synergy agent does not choose a reference model and does not reconstruct the pharmacology behind a combination from first principles; it reports whether the synergy claim already made in the manuscript is supported by the evidence presented for it. Model choice depends on the mechanistic relationship between the two agents — same target, or independent pathways — and that is an argument the authors have to make in the discussion.

The agent reviews original research articles reporting drug combination experiments. It is not the right agent for a single-agent dose-response paper, for a review or meta-analysis, or for a combination mentioned only in a discussion section with no combination data behind it. For whether the animal model supports the conclusion at all, use mouse model relevance. Where the fractional effects feeding a combination index come from image-derived readouts, image quantification checks the measurement before any model consumes it. Writing your own variant of this brief is covered in how to write a custom reviewer agent.

How this differs from the built-in review

This synergy agent differs from PerfectPaper’s standing review team in what kind of knowledge it applies: pharmacological rather than logical. The standing team’s conclusion-validity and logic specialists compare claims against presented evidence and will flag an overreaching sentence, but they do not know that synergy has a technical definition requiring a named reference model, or that a three-bar figure cannot establish it however large the difference — recognising that gap takes domain knowledge the general-purpose specialists do not have.

The agent is briefed to propose the accurate claim rather than recommend deleting the finding, because combination benefit is usually still a real and publishable result. Full set: AI peer review for clinical trials.

PerfectPaper reads the abstract, the figure legends and the methods together, and reports every synergy claim that names no reference model or rests on a design the named model cannot support.

Review my manuscript

Related agents

These agents share a review with this one; you can run three at a time. See AI peer review for clinical trials for the full set of five therapeutics agents.

• Response criteria and endpoints — use it for confirmed responses, censoring rules and waterfall consistency.
• Early-phase safety and dose reporting — use it when the dose-limiting toxicity definition or its window is unstated.
• Registration and pre-specification — use it to catch a reported primary endpoint that differs from the registered one.
• Translational dose relevance — use it when the effective concentration may exceed achievable human exposure.