Skip to content

Synergy Models Overview

Drug combination analysis aims to determine whether two drugs interact synergistically (enhance each other), antagonistically (interfere), or independently. Several mathematical frameworks exist, each with different assumptions.

Currently Implemented

Model Null hypothesis Strengths Page
Bliss Independence Drugs act independently on the same pathway Probabilistic foundation; sensitive to moderate synergy Details
HSA (Highest Single Agent) Combination matches the better single drug Simple; most lenient synergy bar vs Bliss Details
Loewe Additivity / CI Drugs act as dose-additive dilutions of each other Most-cited in pharmacology; interpretable CI parameter Details
ZIP (Zero Interaction Potency) Per-row and per-column Hill fits, each anchored at the other drug's single-agent effect, compared with the zero-interaction (Bliss) expectation Default in SynergyFinder; explains potency shifts without undefined CI cells Details

Each model defines a reference surface — the response matrix predicted under its null hypothesis. synfit exposes these directly via the Deviation / Reference toggle on the matrix Dataset Detail view; see Reference Surfaces.

Same direction required

All four scores assume both drugs act in the same direction (both inhibition or both activation). If one drug is an inhibitor and the other an activator, there is no shared “expected effect” — synergy heatmaps are not shown; the dataset detail page explains why and links here.

Planned Models

Coming soon

The following models are on the development roadmap. They will appear here as they are implemented.

MuSyC (Multi-dimensional Synergy of Combinations)

A state-of-the-art parametric model that decomposes synergy into two orthogonal axes:

  • α (synergy of potency) — does drug B shift drug A's EC₅₀?
  • β (synergy of efficacy) — does the combination change the maximum achievable effect?

A combination can be potency-synergistic but efficacy-antagonistic, which Bliss and Loewe conflate.

Reference: Meyer, C.T. et al. (2019) "Quantifying drug combination synergy along potency and efficacy axes"

BRAID (Bivariate Response to Additive Interacting Doses)

Extends Loewe additivity to a full parametric 2D dose-response surface. The BRAID interaction parameter κ captures the direction and magnitude of the interaction in a single interpretable number.

Reference: Twarog, N.R. et al. (2016) "BRAID: A unifying paradigm for the analysis of combined drug action"

Comparison

Model Type Parameters Handles non-overlapping effect ranges? Separates potency from efficacy synergy?
Bliss Cell-wise 0 (uses single-drug fits) Yes No
HSA Cell-wise 0 Yes No
Loewe/CI Cell-wise 0 No (undefined outside range) No
ZIP Cell-wise 3-param fits per row/column Yes No
MuSyC Surface fit 9–11 Yes Yes
BRAID Surface fit 7 Yes Partially

Why the metrics can disagree

Bliss and Loewe are different null models, not just different reporting conventions. Bliss assumes the two drugs act on independent targets and combines their effects multiplicatively (E_combined = 1 − (1 − E_A)(1 − E_B) on normalised response). Loewe assumes the drugs act on the same target and combines their doses via the isobole

\[ \frac{[A]}{A_E} + \frac{[B]}{B_E} = \text{CI}. \]

These two formulations agree only under narrow conditions (essentially: both drugs Hill = 1 and identical effect transformations). For all other data — i.e. essentially all real pharmacology — they predict different no-interaction surfaces from the same single-agent curves.

The practical consequence is one of the most common interpretation pitfalls in combination pharmacology: a truly Bliss-independent combination produces a non-unity Loewe CI by construction, and vice versa. Running the analysis through Loewe alone could flag genuinely independent drugs as "synergistic" — and indeed, on synfit's own matrix_independent synthetic fixture (truly Bliss-additive, two drugs with Hill = 1.5 and Hill = 1.2), Loewe reads CI < 0.7 — looking like moderate synergy — at 7 of 25 cells, while Bliss deviation is essentially zero everywhere. The "synergy" is a framework artifact, not a biological signal.

This is why synfit shows all four metrics side-by-side rather than picking one. The interpretation rule:

  • Convergence across metrics — strong evidence of real interaction (or real independence). Trust it.
  • Divergence across metrics — diagnostic, not a stronger signal. The drugs probably violate one framework's assumption (independence vs same-target). Treat each metric as informative about a different question, and prefer the one whose null model matches the biology you have reason to believe.

Special case: same drug with itself. A drug combined with itself is by construction Loewe-additive (CI = 1 everywhere) — it is not Bliss-independent, since it shares its mechanism with itself. The matrix_loewe_sham synthetic fixture in synfit makes this concrete: Loewe CI ≈ 1 across the matrix, Bliss deviation strongly non-zero. Either tool used in isolation would mis-read this canonical reference case.

References:

  • Berenbaum, M.C. (1989) "What is synergy?" Pharmacological Reviews 41(2):93–141 — the original articulation that Bliss and Loewe diverge on identical data.
  • Greco, W.R., Bravo, G., Parsons, J.C. (1995) "The search for synergy: a critical review from a response surface perspective" Pharmacological Reviews 47(2):331–385 — landmark systematic comparison.
  • Foucquier, J., Guedj, M. (2015) "Analysis of drug combinations: current methodological landscape" Pharmacology Research & Perspectives 3(3):e00149 — modern review; recommends multi-framework reporting.