Selecting a Reference Null for Combination Synergy

Why the Reference Null Matters

A combination response is not, by itself, evidence of synergy. The observed response must be compared with a defined reference null: the response predicted if the drugs interact according to a particular no-interaction model. Different references formalize “no interaction” differently. They can therefore classify the same dose pair differently without any arithmetic error.

For clarity, let $E_A(d_A)$ and $E_B(d_B)$ denote the effects of drugs A and B given alone at doses $d_A$ and $d_B$, and let $E_{AB}(d_A,d_B)$ be the observed combination effect. Assume throughout that effects are normalized to the interval from 0 to 1, where 0 is no effect and 1 is the maximum possible effect, and that a larger value means a greater response of interest, such as greater fractional inhibition. A positive difference from a reference prediction is then an excess effect under this convention. For endpoints where a smaller value is beneficial, first define an effect scale on which larger values represent more of the effect being analyzed.

The basic assumptions matter. Single-agent and combination measurements should use a consistent endpoint, exposure duration, and normalization; the doses in the single-agent data should match those used in the combination matrix; and the response scale must suit the chosen reference. Experimental uncertainty also matters: a small difference from a reference prediction is not automatically reliable evidence of interaction.

Four Reference Models

Definition: The Bliss independence reference predicts the combination effect from independent action on the unaffected fraction: $$ E_{\mathrm{Bliss}}=E_A+E_B-E_AE_B=1-(1-E_A)(1-E_B). $$ The Bliss excess is $E_{AB}-E_{\mathrm{Bliss}}$. Positive values indicate more effect than predicted by independence; negative values indicate less.

Bliss is natural when the drugs are treated as acting independently on the fraction not yet affected by either drug. It uses the fractional effect scale and does not require that the drugs have dose-equivalent actions. Its independence assumption is substantive: it is not a universal definition of additivity.

Definition: The highest-single-agent (HSA) reference predicts the larger of the two single-agent effects: $$ E_{\mathrm{HSA}}=\max\{E_A,E_B\}. $$ The HSA excess is $E_{AB}-E_{\mathrm{HSA}}$. A positive value means the combination outperforms the better single agent at those doses.

HSA asks a direct practical question: does the combination do better than either component alone? It does not model how two effects should combine, and passing this benchmark is a relatively permissive criterion for calling an excess effect. HSA is useful when outperforming either monotherapy is the relevant benchmark, but it should not be mistaken for a mechanistic additivity model.

Definition: The Loewe additivity reference treats the drugs as dose-equivalent contributors to a common effect. At a target effect $E$, let $D_A(E)$ and $D_B(E)$ be the doses of each drug alone that produce $E$. The combination index is $$ CI(E)=\frac{d_A}{D_A(E)}+\frac{d_B}{D_B(E)}. $$ The Loewe additive condition is $CI(E)=1$; $CI(E)<1$ indicates that less dose is needed than the additive isobole predicts, and $CI(E)>1$ indicates more.

Loewe requires single-agent dose-response information that permits dose equivalence at the effect being evaluated. It is especially interpretable for agents with similar or substitutable actions. If one drug cannot produce the combination effect alone, its dose $D_i(E)$ is not available, and the standard dose-equivalence calculation is undefined there. Applying Loewe in that case requires an explicit extension or alternative model, not an invented dose.

ZIP means zero interaction potency. It asks whether the combination response departs from a reference surface in which neither drug shifts the potency of the other. In practical use, the reference is constructed from fitted single-agent dose-response curves under a specified ZIP implementation. Unlike the simple pointwise formulas for Bliss and HSA, ZIP is not a single universal formula independent of implementation: fitting, response normalization, and surface construction can differ. A ZIP result therefore needs the fitting and reference-surface convention to be stated. Its defining question is potency preservation, rather than Bliss independence, Loewe dose equivalence, or simply beating the better single agent.

Why Bliss and Loewe Can Disagree

A short calculation shows why different nulls can give different answers. Suppose monotherapy effects are $E_A=0.40$ and $E_B=0.30$. Bliss predicts $0.40+0.30-(0.40)(0.30)=0.58$. HSA predicts $0.40$. The two reference effects differ because they answer different questions.

For a Loewe comparison, suppose each single-agent curve follows a unit-slope model $E(d)=d/(D_{50}+d)$, with its own $D_{50}$. Under this assumed model, the dose producing effect $E$ is $D(E)=D_{50}E/(1-E)$. The doses yielding the stated monotherapy effects therefore correspond to $d_A/D_{50,A}=0.40/0.60$ and $d_B/D_{50,B}=0.30/0.70$. These dose ratios permit a Loewe prediction at the combination point.

CI(E)=\frac{1-E}{E}\left(\frac{0.40}{0.60}+\frac{0.30}{0.70}\right)=1 \quad\Longrightarrow\quad E_{\mathrm{Loewe}}=\frac{1.0952}{1+1.0952}\approx 0.5227.

The derivation uses the single-agent inverse dose functions: at effect $E$, the fraction of each drug’s dose required is $d_i/D_i(E)=(d_i/D_{50,i})(1-E)/E$. Setting their sum to one gives the displayed equation. The result depends on the assumed curves and their shared unit slope; it is not a curve-free Loewe prediction.

Worked Example: One Matrix Point, Different Classifications

Consider hypothetical normalized effects $E_A=0.40$, $E_B=0.30$, and observed $E_{AB}=0.55$ at one dose pair. Assume the monotherapy curves and doses satisfy the unit-slope model just described. Compare the three pointwise references and suppose a prespecified ZIP fit predicts $E_{\mathrm{ZIP}}=0.56$ at this pair.

Bliss. The expected effect is $0.40+0.30-(0.40)(0.30)=0.58$. The excess is $0.55-0.58=-0.03$, so the combination is below the Bliss reference.

HSA. The expected effect is $\max(0.40,0.30)=0.40$. The excess is $0.55-0.40=0.15$, so it exceeds the better single agent.

Loewe. From the preceding derivation, the additive effect at these doses is approximately $0.5227$. The observed effect exceeds it by $0.55-0.5227=0.0273$, which is above the Loewe additive prediction.

ZIP. Under the stated fitted reference prediction, the difference is $0.55-0.56=-0.01$, below the ZIP prediction. This number is conditional on the specified ZIP fit; it cannot be reconstructed from the two monotherapy effects alone.

This point is below Bliss and ZIP, but above HSA and Loewe. Thus, a bare label such as “synergistic” omits essential information: the reference model, its assumptions, and the effect scale.

Worked Example: When Loewe Dose Equivalence Fails

Worked Example: A Combination Effect Outside One Monotherapy Range

Suppose hypothetical single-agent curves have maximum effects of $0.60$ for drug A and $0.90$ for drug B. At a selected dose pair, their measured effects are $E_A=0.50$, $E_B=0.40$, and the combination effect is $E_{AB}=0.72$. Assume these values are measured on the same normalized inhibition scale.

HSA. The reference is $\max(0.50,0.40)=0.50$. The combination exceeds it by $0.72-0.50=0.22$.

Bliss. The reference is $0.50+0.40-(0.50)(0.40)=0.70$. The combination exceeds it by $0.72-0.70=0.02$.

Loewe. Drug A’s maximum effect is $0.60$, so there is no dose $D_A(0.72)$ on its observed single-agent curve. The standard Loewe index at effect $0.72$ cannot be evaluated from these curves. The HSA and Bliss calculations remain defined, but they answer different questions; they do not repair the missing dose equivalence.

This limitation is important when drugs have markedly different maximum effects, partial agonism, or incompatible response ranges. Extrapolating a fitted curve beyond supported data may yield a numerical dose, but that does not make the dose well identified or the Loewe comparison trustworthy. State the range of effects over which the single-agent curves support the comparison.

Response Scale, ZIP, and Matrix-Level Interpretation

Worked Example: Normalizing Viability Before Applying Bliss

Suppose hypothetical relative viabilities are $0.60$ with drug A, $0.70$ with drug B, and $0.45$ with the combination, all relative to the same untreated control. On an inhibition scale, the single-agent effects are $1-0.60=0.40$ and $1-0.70=0.30$, while the combination effect is $1-0.45=0.55$.

Bliss predicts remaining viability $0.60\times0.70=0.42$, equivalent to inhibition $1-0.42=0.58$. The observed viability $0.45$ is higher than the Bliss-predicted $0.42$, equivalently the observed inhibition $0.55$ is below $0.58$. Thus the result is below Bliss, even though the combination has lower viability than either single agent.

The calculation uses consistent control-normalized fractions. A raw fluorescence value or an arbitrary transformed score cannot be substituted without checking that the reference model is meaningful on that scale. The same caution applies when comparing ZIP fits: the endpoint, normalization, fitting range, and implementation should be reported.

A dose matrix usually contains many dose pairs. A pointwise excess surface can be summarized over a region, but the summary depends on which doses are included and how the surface is aggregated. A region-wide positive average can conceal antagonistic regions; a single positive cell can be unstable. Inspect the monotherapy curves and the full response surface alongside any summary score, and consider uncertainty and replicate variability before interpreting small differences.

The reference models are alternatives, not a ranking from incorrect to correct. Bliss is suited to an independent-action premise; Loewe to dose equivalence and substitutable effects; HSA to a “better than either alone” benchmark; and ZIP to a specified potency-preservation reference. The relevant choice follows from the biological question and available data. Agreement across references can strengthen a descriptive conclusion, but disagreement is informative about the assumptions driving it.

Diagnostics and Common Pitfalls

Calling any combination above monotherapy synergistic. This only establishes HSA excess. Compare with the reference that matches the intended null.
Using Loewe without dose-response support. Check that both drugs can produce the target effect and that the fitted curves are supported near it. Otherwise, report the Loewe comparison as unavailable or justify an explicit extension.
Treating Bliss as universal additivity. Bliss assumes independent fractional effects. Dependence between drug actions may make that reference biologically unsuitable.
Reporting a ZIP score without its convention. State the response normalization, single-agent curve fitting, ZIP implementation, and the dose region summarized.
Ignoring response direction or scale. Verify that larger normalized effect means more of the effect of interest, and that controls and endpoints are consistent across conditions.
Equating a reference difference with certainty. An excess is a point estimate. Replicate variation and model-fit uncertainty determine how confidently it can be distinguished from zero.

Key takeaway: A synergy classification is always relative to a null model. Name the reference, state its assumptions and response scale, verify that its data requirements are met, and interpret disagreement as model dependence rather than as a contradiction in the measurements.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

For $E_A=0.20$ and $E_B=0.50$, what are the Bliss and HSA reference effects?
An observed combination effect is $0.62$ for those same single-agent effects. What is the HSA excess, and what does it establish?
Why might a Loewe comparison be undefined if one drug’s maximum effect is below the observed combination effect?
What additional information is required to interpret a ZIP comparison beyond the two single-agent effects at one dose pair?
Why can a combination beat both single agents but remain below the Bliss reference?

Solutions. (1) Bliss is $0.20+0.50-(0.20)(0.50)=0.60$; HSA is $0.50$. (2) The HSA excess is $0.62-0.50=0.12$, establishing that the combination exceeds the better single agent at those doses, not that it exceeds every other null. (3) There is no single-agent dose producing that target effect, so the required dose equivalent is unavailable on the observed curve. (4) A specified ZIP implementation and fitted monotherapy dose-response information are needed to construct its potency-preserving reference surface. (5) HSA compares with the larger single-agent effect, whereas Bliss predicts the joint effect of independent action; the Bliss reference can be larger than either monotherapy effect.