Skip to content

Bliss vs Loewe vs HSA vs ZIP: which synergy model?

"Synergy" means more effect than expected. Each model defines "expected" differently, so the same combination can be synergistic by one and antagonistic by another. Neither is a bug. The models answer different questions, and choosing between them is a statement about how you think the drugs act.

The one-paragraph version

Model Expected effect of the combination Implied mechanism Use it when
HSA (Highest Single Agent) The better of the two drugs alone None; a pragmatic floor Asking "is the combination better than just using the stronger drug?"
Bliss independence Probabilities combine: \(E_A + E_B - E_A E_B\) Different targets, acting independently Drugs hit different pathways (the usual case in combination screens)
Loewe additivity (CI) Doses add, as if one drug were a dilution of the other Same target or mechanism Drugs share a target, or you need a Chou–Talalay combination index
ZIP (Zero Interaction Potency) Bliss expectation, compared against re-fitted dose-response curves Different targets; looks at potency shifts Screening data where a single noisy well should not dominate; comparing with SynergyFinder

One example, three answers

Take two drugs with the same single-agent curve (IC₅₀ = 1 µM, Hill = 1). At 1 µM each, each drug alone gives 50 % inhibition. What should 1 µM + 1 µM give if nothing special happens?

Model Expected Observed 70 % is…
HSA 50 % (the better drug alone) synergy
Loewe 67 % (2 µM of "the same drug") synergy (CI ≈ 0.86)
Bliss 75 % (\(0.5 + 0.5 - 0.25\)) antagonism

ZIP starts from the same 75 % expectation, but scores a combination surface fitted across the whole row and column, so one well cannot settle its verdict.

The data did not change; only the definition of "nothing special". This is why synfit computes all four and shows them side by side.

Where each model breaks

HSA is the most lenient. Any combination that beats its best component looks synergistic, even when the two drugs merely add up. It is a useful first filter and a weak claim on its own.

Bliss assumes the drugs act independently. A drug combined with itself is the opposite of independent, and Bliss scores it as clearly non-additive (synergistic or antagonistic, depending on the slope), although nothing interesting is going on. Bliss also needs effects on a common 0–100 % scale; synfit gets that from the fitted top and bottom plateaus.

Loewe assumes one drug behaves like a dilution of the other. Two consequences:

  • When the two Hill slopes differ, a truly independent pair reads as synergistic or antagonistic depending on the dose. synfit warns when the slopes differ by 1.5× or more (slope mismatch).
  • CI is undefined where the combination reaches an effect neither drug alone can reach, which is exactly where strong synergy lives. Those cells appear grey.

ZIP inherits Bliss's independence assumption but scores the fitted combination surface, built from per-row and per-column curve fits, rather than raw wells. That smooths out single-well noise, and it can fail on rows or columns too flat to fit; synfit lists those. See ZIP.

On synfit's own synthetic test data, a pair built to be exactly Bliss-independent (Hill 1.5 and 1.2) shows Loewe CI below 0.7, apparent moderate synergy, in 7 of 25 cells. The "synergy" comes from the model choice, not from biology. More on this.

A practical rule

  1. Look at all four. If they agree, the conclusion does not hinge on one definition of "no interaction". That is robustness, not independent confirmation: all four score the same wells, so the replicates and their spread still carry the evidence.
  2. If they disagree, ask which assumption fits your drugs. Drugs on different pathways make Bliss and ZIP the more natural reference (different targets do not guarantee independent action, though). Drugs on the same target, or a dose-equivalence question, make Loewe the more natural one.
  3. Say which model a claim rests on. "Synergistic by Bliss and ZIP (δ ≈ −0.15 at the centre of the matrix), additive by Loewe" is reproducible. "Synergistic" alone is not.
  4. Look at regions, not cells. A coherent block of synergy across neighbouring doses means more than one extreme well.

What about MuSyC, BRAID and others?

Newer surface-fitting models separate synergy of potency (does one drug shift the other's IC₅₀?) from synergy of efficacy (does the combination reach a deeper maximum?). They are on synfit's roadmap; see Synergy models overview.

Try it

Run a combination analysis on your own matrix, or open the demo and compare the "Loewe sham (drug + itself)" and "no synergy" sample matrices: each fools one of the models.

Further reading