Why a Synergy Claim Needs a Reference
A drug combination is not synergistic or antagonistic in the abstract. The claim compares the combination's response with a specified reference for what would have happened without an interaction. Change that reference, the response scale, or the dose range used to construct it, and the conclusion may change. A large combination effect by itself does not establish synergy: two drugs can produce a large effect simply because each is active.
For clarity, suppose the response has been oriented so that larger values mean greater benefit. Let \(m(d_A,d_B)\) denote the expected response at doses \(d_A\) and \(d_B\), for a specified population and assessment time. A no-interaction reference model predicts \(m_0(d_A,d_B)\). The interaction contrast is the difference between the observed or estimated combination response and that reference:
With this orientation, a positive contrast means the combination response exceeds the selected reference; a negative contrast means it falls below it. If the outcome is an adverse response for which lower values are better, first define a benefit-oriented response or reverse the interpretation of the signs. The contrast is meaningful only when the response definition, units, population, time point, and reference are specified.
Different References Answer Different Questions
The earlier material on Bliss, Loewe, and highest-single-agent additivity introduced several common references. Their conclusions need not agree because they encode different meanings of “no interaction.” Highest-single-agent (HSA) compares the combination with the better of the two single-agent responses at the same component doses. Bliss independence combines fractional effects under an independent-action assumption. Loewe additivity asks whether the component doses are dose-equivalent to a dose of a single agent producing the same effect. Choosing among these is a scientific modeling decision, not a search for the one reference that gives the preferred result.
For example, with baseline-normalized fractional benefits \(E_A\) and \(E_B\), Bliss independence predicts \(E_0=E_A+E_B-E_AE_B\). This follows by treating the fractions unaffected by each drug as independent: the unaffected fraction under the combination is \((1-E_A)(1-E_B)\), so the affected fraction is \(1-(1-E_A)(1-E_B)\). The independence assumption, the fractional-effect scale, and the baseline normalization are part of the reference; the equation is not a universal definition of additivity.
Worked Example: One Combination, Two Reference Conclusions
In a hypothetical experiment, drug A produces 30% benefit and drug B produces 20% benefit at the tested doses. The combination produces 44% benefit. Compare it with HSA and Bliss references, assuming the percentages are baseline-normalized fractional benefits measured at the same time and in the same population.
HSA reference. The better single-agent response is \(\max(30\%,20\%)=30\%\). The combination contrast is \(44\%-30\%=14\) percentage points, so the combination exceeds the HSA reference.
Bliss reference. Convert percentages to fractions and calculate \(E_0=0.30+0.20-(0.30)(0.20)=0.44\), or 44%. The Bliss contrast is \(44\%-44\%=0\) percentage points. The combination is exactly at the Bliss reference.
Both calculations are correct for their stated references. The combination exceeds the stronger single agent, but the result is not greater than Bliss independence. Reporting simply “the drugs are synergistic” would hide this distinction. A defensible report names the comparison and explains why that reference is relevant.
The Response Scale Is Part of the Claim
Additivity is not invariant to transformations of the outcome. “Add the two effects” on a raw scale is not the same model as “multiply the remaining fractions” on a residual scale. A nonlinear transformation can change the size and even the sign of an interaction contrast. This is not a mathematical defect: it means that each scale expresses a different scientific statement.
The response scale should be chosen to represent the scientific question and measurement process. For example, a difference in biomarker units may be interpretable, while a proportional change may be more suitable when effects operate multiplicatively. For bounded outcomes, predictions must also remain within their possible range. A model that predicts an impossible response, or whose residual patterns vary systematically with dose, may be poorly suited to the chosen scale.
Worked Example: Raw-Scale Additivity Versus Multiplicative Residuals
Suppose a hypothetical biomarker has an untreated mean of 100 units. Drug A alone reduces it to 80 units, drug B alone reduces it to 80 units, and the combination mean is 64 units. Compare two stated reference models.
Additive reductions on the raw scale. Each drug reduces the biomarker by 20 units. The additive reference is \(100-20-20=60\) units. The combination contrast is \(64-60=4\) units: the observed biomarker is 4 units higher than the raw-additive prediction, which is less benefit under this outcome orientation.
Multiplication of remaining fractions. Each drug leaves \(80/100=0.80\) of the biomarker. The multiplicative reference is \(100(0.80)(0.80)=64\) units. The combination is exactly at that reference.
Thus, the same hypothetical means are below additive benefit on one reference and exactly at the other. The raw-scale conclusion does not disprove the multiplicative model, and the multiplicative conclusion does not establish that multiplication is scientifically correct. The scale and mechanism motivating the reference must be justified before interpreting the contrast.
Dose Extrapolation Can Manufacture a Reference
Some references require estimates of monotherapy responses at doses or effects that were not adequately observed. Loewe calculations, for example, depend on dose-equivalent values at the combination effect. If that effect lies beyond the measured monotherapy range, a fitted curve may extrapolate it. The resulting interaction estimate can then be driven more by curve shape and parameter uncertainty than by observed combination data.
A fitted curve can appear smooth and still be weakly supported at its extremes. Sparse doses, a narrow range, or a plateau not reached in the data may leave the maximum effect and potency poorly determined. A highly precise-looking derived interaction can be misleading if uncertainty in those monotherapy parameters is ignored. Separate checks should ask whether the combination doses and the reference effect lie within the data-supported region, whether the curve form is plausible, and whether plausible alternative curves change the conclusion.
Worked Example: A Loewe Conclusion Driven by Extrapolation
Consider a hypothetical monotherapy experiment. Each drug has measured benefit of 0%, 20%, and 30% at doses 0, 1, and 2 mg, respectively. A combination of 2 mg of each produces 40% benefit. The monotherapy data stop at 30%, so the dose required for 40% benefit, often denoted \(D_{40}\), is not observed.
Suppose an analyst extends a straight line through the 1 mg and 2 mg observations. Its slope is \((30-20)/(2-1)=10\) percentage points per mg. Extrapolating from 30% at 2 mg gives an estimated \(D_{40}=2+(40-30)/10=3\) mg for each agent.
The Loewe dose-fraction sum for the combination is then \(2/3+2/3=1.333\). Under the usual interpretation, a sum greater than 1 suggests that more total dose was used than the additive reference would require, indicating antagonism relative to that reference. But this result relies on an extrapolated \(D_{40}\), not on observed 40% monotherapy doses.
If additional monotherapy data instead supported \(D_{40}=4\) mg for each drug, the sum would be \(2/4+2/4=1.000\), consistent with Loewe additivity. The measured combination response has not changed; the reference changed because the dose-response information improved. With the original data, the appropriate conclusion is that a Loewe classification is sensitive to unsupported dose extrapolation, not that antagonism has been established.
Uncertainty Is Part of the Evidence
An estimated interaction contrast is not a fixed truth. Sampling variation, model parameter uncertainty, variability between individuals, and uncertainty in concentrations or exposures can all affect it. An interval for \(\Delta\) describes uncertainty under the fitted model and analysis; it does not repair a poorly chosen reference, unsupported extrapolation, or an inappropriate response scale.
For a prespecified contrast with an approximately normal sampling distribution and estimated standard error \(SE\), a two-sided 95% confidence interval is commonly calculated as \(\hat{\Delta}\pm1.96SE\). This approximation requires that the estimator's sampling distribution be adequately approximated by a normal distribution. When interaction estimates are derived from nonlinear models, resampling or model-based intervals may be preferable, provided the method propagates relevant parameter uncertainty.
Worked Example: An Estimate That Does Not Establish Synergy
In a hypothetical study, the estimated combination benefit exceeds a prespecified reference by 8 response units, with \(SE=5\) units. Assume the contrast is approximately normally distributed and the standard error adequately represents its sampling uncertainty. Test \(H_0:\Delta=0\) against \(H_a:\Delta\ne0\) at \(\alpha=0.05\).
The standardized statistic is \(z=\hat{\Delta}/SE=8/5=1.60\). The two-sided p-value is approximately \(2P(Z\ge1.60)=0.110\), where \(Z\) has a standard normal distribution. The 95% confidence interval is \(8\pm1.96(5)=8\pm9.8\), or \((-1.8,17.8)\) response units.
Because the interval includes zero and the p-value exceeds 0.05, this analysis does not provide strong evidence against the no-interaction reference. It does not prove that the interaction is absent: values ranging from a small contrast below the reference to a substantial positive contrast are compatible with this interval. The conclusion is also conditional on the chosen reference, model, and assumptions.
Diagnostics, Comparisons, and Common Pitfalls
Before interpreting a synergy or antagonism estimate, check whether the analysis is answering the intended question. Inspect the monotherapy and combination observations together, not just a summary interaction score. Look for inadequate dose coverage, poor fit or systematic residual patterns, predictions outside the possible response range, and sensitivity to plausible model forms. If the response is measured repeatedly or varies across individuals, ensure that the uncertainty calculation reflects the design rather than treating dependent observations as independent.
For exposure-mediated combinations, also distinguish administered-dose comparisons from exposure-controlled comparisons. If one drug changes the concentration of the other, a response difference at the same administered dose may reflect altered exposure rather than a direct pharmacodynamic interaction. The estimand and the analysis should make clear which comparison is intended. This is why interaction contrasts need a clearly defined dose or exposure condition and why uncertainty in estimated concentrations may matter.
| Potential issue | What to check | Safer interpretation |
|---|---|---|
| Reference-model choice | Does the named null match the scientific question? | Report the model-specific contrast, not unqualified “synergy.” |
| Response scale | Would another scientifically plausible scale change the result? | Justify the scale and state the units of the contrast. |
| Dose extrapolation | Are reference doses and effects supported by observed monotherapy data? | Label extrapolations and assess sensitivity to curve assumptions. |
| Uncertainty | Does the interval include zero, and are important uncertainties represented? | Distinguish evidence against the reference from proof of no interaction. |
| Exposure changes | Could one drug have changed the other drug's exposure? | State whether the claim concerns administered doses or exposure-controlled effects. |
Common mistake: “The combination response is larger than either single-agent response, so the drugs are synergistic.” This reasoning compares the combination with HSA but treats that comparison as if it were the only possible no-interaction reference. The correction is to name the reference, calculate its predicted response at the tested doses, and compare the combination with that prediction. A combination can exceed HSA while exactly matching Bliss independence, as the first example showed.
A second frequent error is to report a positive point estimate as conclusive synergy while omitting its interval or the fact that the reference depends on extrapolation. A positive estimate describes its direction; it does not by itself establish that the effect differs from zero or that the model is credible. Report the estimate and its units, uncertainty, reference, dose range, and any material sensitivity to assumptions. If several doses, response scales, or reference models were explored, make the exploratory nature clear rather than presenting the most favorable result as if it had been the only planned comparison.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Two drugs produce fractional benefits of 0.40 and 0.25. What is the Bliss reference benefit, and what assumption does it encode?
- A combination exceeds HSA but equals the Bliss prediction. Is it correct to call it synergistic without qualification? Explain.
- Why is a Loewe conclusion uncertain if the combination effect exceeds every measured monotherapy effect?
- An estimated interaction contrast is \(-3\) units with \(SE=2\) units. Calculate an approximate 95% confidence interval and interpret it.
Solutions. (1) The Bliss reference is \(0.40+0.25-(0.40)(0.25)=0.55\), or 55% benefit. It encodes independent action on the baseline-normalized fractional-effect scale, so the unaffected fractions multiply. (2) No. It exceeds the HSA reference but is not above the Bliss reference; the report should state which reference is being used and avoid an unqualified claim. (3) The required single-agent dose for the combination effect is not observed, so the dose equivalence must be estimated by extrapolation. The result may depend strongly on the assumed monotherapy curve. (4) The interval is \(-3\pm1.96(2)=-3\pm3.92\), or approximately \((-6.92,0.92)\) units. It includes zero, so the estimate does not establish a nonzero contrast at the approximate 95% confidence level; it also does not prove the contrast is zero.