Why an Interaction Score Depends on the Endpoint
A combination can look synergistic on one response scale and non-synergistic on another without any change in the measured observations. The reason is that an interaction score is not a property of two drugs alone: it compares observed combination response with a specified reference prediction, and that prediction depends on the endpoint and its scale. A fractional inhibition endpoint and an absolute change endpoint need not imply the same reference prediction.
This distinction matters in QSP because experimental outputs may be transformed before comparison: a cell count may become a percent reduction, a concentration may become a log concentration, or a single measurement may be replaced by an area under a response curve. Baseline selection and measurement time can also change the calculated effect. These choices should be specified before interpreting a combination as synergistic or antagonistic.
For the examples below, let \(Y_0\) be the response under a matched control condition, and let \(Y_A\), \(Y_B\), and \(Y_{AB}\) denote responses with drug A, drug B, and their combination. We use an endpoint where a lower response indicates a stronger drug effect. Absolute reduction is \(D_i=Y_0-Y_i\), in the original response units. Fractional inhibition is \(e_i=(Y_0-Y_i)/Y_0\), provided \(Y_0>0\) and the denominator is appropriate for the experimental question.
The Bliss expression is appropriate when the fractional effects are interpreted as independently acting probabilities of preventing the measured response, under the assumptions behind that reference. Absolute additivity instead assumes that reductions add in the original units. Neither assumption follows merely from observing a dose matrix. Selection of a reference null is a scientific modeling decision; earlier material on Bliss independence, Loewe additivity, and reference-null selection provides the corresponding framework.
Why Absolute and Fractional Predictions Diverge
Suppose both drugs reduce the same positive control response, and the observed combination is additive in absolute units. Write \(e_A=D_A/Y_0\) and \(e_B=D_B/Y_0\). Absolute additivity then gives a combination fractional inhibition of \(e_A+e_B\), if that sum is within the endpoint’s feasible range. Bliss predicts a different value.
The gap is \(e_Ae_B\), which is positive when both monotherapies have positive fractional effects. Thus, a combination that is exactly additive in absolute units will exceed the Bliss prediction whenever both drugs reduce the response. Under the Bliss reference it will look synergistic, even though it has no departure from absolute additivity. This is not an arithmetic error; the two references encode different meanings of “no interaction.”
This short derivation also identifies a practical diagnostic: calculate the combination residual under each prespecified reference, in the same units as that reference. Do not infer a mechanism from the sign of a residual until the response scale and null model are clear.
Worked Example: Absolute Additivity Looks Synergistic Under Bliss
In a hypothetical assay, the matched vehicle response is 100 response units. Drug A alone gives 80 units, drug B alone gives 80 units, and the combination gives 60 units. Assume the assay is in range, the control is matched to the treated wells, and the values are the means of adequate replicate measurements. Compare the absolute-additivity and Bliss predictions.
Absolute scale. The reductions are \(D_A=100-80=20\) units and \(D_B=100-80=20\) units. Absolute additivity predicts \(D_{AB}=20+20=40\) units of reduction, or a combination response of \(100-40=60\) units. The observed combination is exactly 60 units, so its departure from absolute additivity is \(60-60=0\) response units.
Fractional scale. Each monotherapy inhibits \(20/100=0.20\), or 20%. Bliss predicts \(e_{AB}=0.20+0.20-(0.20)(0.20)=0.36\), or 36% inhibition. The observed combination inhibits \((100-60)/100=0.40\), or 40%. Its Bliss excess is \(0.40-0.36=0.04\), equivalent to 4 percentage points of inhibition.
The combination is exactly additive in absolute response units but exceeds the Bliss prediction by 4 percentage points. Calling it simply “synergistic” would omit the essential qualification: it appears synergistic relative to Bliss on the fractional-inhibition scale, while showing no departure from absolute additivity in this example.
Baseline Choice Changes Fractional Effects
A fractional endpoint is only as meaningful as its denominator. The denominator is often the matched vehicle response at the same time point, but other choices—such as a pretreatment value, a historical control, or a different control group—answer different questions. If \(Y_0\) changes while the treated values remain fixed, the fractional effects change, and so can the combination’s departure from a fractional reference.
A comparison is especially difficult to interpret if the baseline is not shared across treatment groups or is not matched in time. For example, a naturally declining control response can make a drug look more or less inhibitory depending on whether effects are measured against the initial value or against the contemporaneous vehicle. In longitudinal experiments, baseline choice should be stated alongside the endpoint definition rather than hidden in a normalization step.
Worked Example: A Different Denominator Reverses the Bliss Conclusion
Use the same hypothetical treated responses as before: A gives 80 units, B gives 80 units, and the combination gives 60 units. For a deliberately contrasting calculation, compare two possible denominators, 100 and 120 units. Suppose both are positive response references, but only 100 is the contemporaneous matched vehicle in the experiment.
Matched vehicle denominator, 100 units. The fractional inhibitions are \(e_A=e_B=0.20\) and \(e_{AB}=0.40\). The Bliss prediction is \(0.20+0.20-(0.20)(0.20)=0.36\). The observed response exceeds that prediction by \(0.04\), or 4 percentage points: apparent Bliss synergy.
Alternative denominator, 120 units. The calculated effects become \(e_A=e_B=(120-80)/120=0.3333\), while \(e_{AB}=(120-60)/120=0.50\). The Bliss prediction is \(0.3333+0.3333-(0.3333)(0.3333)=0.5556\), to four significant figures. The observed 0.50 is \(0.0556\), or 5.56 percentage points, below that prediction: apparent Bliss antagonism.
The raw treated responses have not changed, but the apparent Bliss conclusion has reversed because the denominator changed. Since 120 is not the matched vehicle in this hypothetical experiment, that calculation would not be the primary analysis; it demonstrates why a normalization choice must be justified, not selected after seeing which result is preferred.
Ceilings and Measurement Time
Many assays have bounded outputs. Percent inhibition cannot exceed 100%, cell counts cannot be negative, and a fluorescent signal may saturate at the instrument’s upper range. A ceiling can compress observed responses: two underlying effects that add on a latent or unsaturated scale may be recorded as the same maximum value. At the same time, comparing the capped observation with a reference such as Bliss can produce a positive apparent interaction.
Worked Example: A Ceiling Creates a Bliss Excess
Consider a hypothetical assay that reports inhibition as a fraction from 0 to 1 and cannot record values above 1. Suppose each drug alone produces 0.55 inhibition. Assume the underlying, uncapped effects are additive on this inhibition scale, so their sum in combination would be \(0.55+0.55=1.10\). The reported combination is capped at 1.00.
Bliss predicts \(e_{AB}=0.55+0.55-(0.55)(0.55)=1.10-0.3025=0.7975\). The measured value of 1.00 is \(1.00-0.7975=0.2025\), or 20.25 percentage points, above the Bliss prediction. It therefore appears strongly synergistic relative to Bliss. But under the stated additive latent-response assumption, its uncapped value is simply the sum of the monotherapy effects; the endpoint ceiling, not an additional mechanistic interaction, explains why the reported combination reaches the maximum.
This calculation does not prove that the underlying effects are additive. Rather, it shows why a measurement at the maximum cannot distinguish a true interaction from a scale mismatch or censoring. A useful check is to determine whether single-agent and combination responses approach the assay limits, and to repeat measurements with a validated range that resolves those responses.
The time of measurement is another part of the endpoint. A combination may exceed a reference at one time and not another because drug exposure, target engagement, feedback, or response turnover evolves. A single time point cannot establish that an apparent interaction persists over the response trajectory. Conversely, an integrated summary such as an area under the curve can conceal a transient interaction or give disproportionate weight to a long-lasting modest effect.
Worked Example: The Size of Apparent Synergy Varies Over Time
In a hypothetical time-course assay, the time-matched vehicle responses are 100, 80, and 60 response units at 1, 4, and 8 hours. Assume the assay is unsaturated at all three times, each vehicle value is an appropriate positive denominator, and the absolute reductions from the vehicle are as follows.
| Time | Vehicle \(Y_0\) | A reduction \(D_A\) | B reduction \(D_B\) | Combination reduction \(D_{AB}\) |
|---|---|---|---|---|
| 1 hour | 100 | 20 | 10 | 30 |
| 4 hours | 80 | 16 | 16 | 32 |
| 8 hours | 60 | 6 | 12 | 18 |
At each time the combination reduction equals the sum of the two monotherapy reductions: \(20+10=30\), \(16+16=32\), and \(6+12=18\). Thus the combination is additive on the absolute scale at all three measurements.
At 1 hour, the fractions are \(e_A=20/100=0.20\), \(e_B=10/100=0.10\), and \(e_{AB}=30/100=0.30\). Bliss predicts \(0.20+0.10-(0.20)(0.10)=0.28\), so the excess is 0.02, or 2 percentage points. At 4 hours the fractions are 0.20, 0.20, and 0.40; Bliss predicts 0.36, giving an excess of 0.04. At 8 hours they are 0.10, 0.20, and 0.30; Bliss predicts 0.28, giving an excess of 0.02.
The fractional Bliss excess varies over time even though absolute additivity holds at every measured time. A report based only on 4 hours would suggest a larger Bliss excess than one based on 1 or 8 hours. The time-course and absolute endpoint show why that difference is a property of the chosen reference and time-specific fractions, not by itself proof that a mechanistic interaction peaks at 4 hours.
Diagnostics and Alternative Comparisons
A defensible analysis makes the response definition explicit and checks whether the reference assumptions are plausible. The following sequence helps separate a scale-dependent score from a robust pattern in the data.
State whether higher or lower values indicate effect, the units, the transformation, and the measurement time or time summary.
Use a control appropriate to the design and time point. State the denominator for any normalized response and check that it is positive and measured on a comparable basis.
Explain why absolute additivity, Bliss independence, Loewe additivity, or another reference matches the question. Do not switch references solely to obtain a preferred label.
Look for values near assay floors or ceilings, control drift, and transient effects. Verify that the observed response is not censored or dominated by a single measurement time.
Report the raw or absolute response alongside the transformed endpoint and its reference prediction. If conclusions differ, state that dependence rather than presenting one scale as definitive.
These reference approaches have different assumptions and are not substitutes for one another. Absolute additivity is directly interpretable in the original units but may be inappropriate for bounded responses or agents with dose-equivalent effects. Bliss uses fractional effects and independence as its reference, so the effect definition and denominator are central. Loewe additivity is designed for dose equivalence and requires suitable single-agent dose-response information. A mechanistic combination model can represent specific pathways or dynamic processes, but its interaction parameters and assumptions must be supported by the design and data; a different scale alone does not identify a mechanism.
The introductory material on selecting a reference null for combination synergy establishes why the null is part of the question being asked. Here the additional practical requirement is to test sensitivity to reasonable endpoint definitions. If a synergy label disappears under a modest baseline change, an alternative valid scale, or a nearby measurement time, the result should be reported as endpoint-dependent. That sensitivity can guide follow-up experiments, but it is not a reason to choose the most favorable endpoint after analysis.
Common Mistakes and Limits
- Calling synergy scale-free. A positive Bliss excess does not imply a positive departure from absolute additivity. Name the reference and scale in the conclusion.
- Normalizing to an arbitrary baseline. A denominator that is not matched to the experimental question can alter fractional effects and even reverse their apparent interaction direction.
- Treating a ceiling value as precise evidence of a large effect. A capped measurement does not reveal how far the uncapped response would have gone; it may mask differences or inflate apparent excess over a reference.
- Reporting only the most favorable time point. A selected time may overstate or understate an interaction. Prespecify the primary time or summary and show the trajectory when available.
- Equating an endpoint discrepancy with mechanism. An interaction score is a comparison with a model-based reference. Mechanistic interpretation requires additional evidence and a model appropriate to the biology.
There are also limits to endpoint-sensitivity checks. An alternative scale is not automatically more valid, and agreement across scales does not prove a mechanism. Replicate variability, dose selection, and uncertainty in the single-agent responses affect the precision of an interaction estimate. When responses are noisy, the calculated excess should be accompanied by uncertainty from an analysis appropriate to the experimental design, rather than being treated as an exact property of the drugs.
Check Your Understanding
Answer each question using the stated endpoint and reference; solutions follow the questions.
- A vehicle response is 200 units. A and B each reduce it by 30 units, and the combination reduces it by 60 units. Is the combination additive on the absolute scale? What is its Bliss excess?
- Why can a combination that is additive in absolute units exceed the Bliss prediction when both fractional monotherapy effects are positive?
- An assay has a maximum reportable inhibition of 1.00. Each drug alone gives 0.60 inhibition, and the combination is reported as 1.00. Calculate the Bliss prediction and explain one limitation of interpreting the excess.
- In a time-course study, absolute reductions are additive at every time, but the fractional Bliss excess is larger at one time point. What should be reported before interpreting that time-specific difference?
Solutions. (1) Absolute additivity predicts \(30+30=60\) units, exactly the observed reduction. Each monotherapy has fractional inhibition \(30/200=0.15\), while the combination has \(60/200=0.30\). Bliss predicts \(0.15+0.15-(0.15)(0.15)=0.2775\). The observed Bliss excess is \(0.30-0.2775=0.0225\), or 2.25 percentage points, despite exact absolute additivity. (2) Absolute additivity gives \(e_A+e_B\), whereas Bliss subtracts \(e_Ae_B\). Their difference is \(e_Ae_B>0\) when both effects are positive. (3) Bliss predicts \(0.60+0.60-(0.60)(0.60)=0.84\). The reported excess is \(1.00-0.84=0.16\), or 16 percentage points. Because the combination is at the measurement ceiling, its uncapped value is unknown; the apparent excess does not by itself establish a mechanistic interaction. (4) Report the endpoint and units, time-matched baseline and denominator, the prespecified reference, the absolute and fractional results, and the response trajectory or selected time-point rationale. Then describe the difference as time- and scale-dependent unless additional evidence supports a mechanistic interpretation.