Skip to content

ZIP (Zero Interaction Potency)

ZIP compares a fitted combination surface, built from per-row and per-column dose–response curves, with the zero-interaction expectation of the two single drugs (Yadav et al. 2015). It is widely used in combination screening, including as the default in SynergyFinder.

Idea

  1. Take each drug's Hill curve from the matrix fit: its potency and slope come from its own single-agent edge (first row or column), with the two plateaus shared by both drugs (see the matrix fitting pipeline on the Bliss independence page).
  2. Map responses to fraction affected \(f = (\text{effect}_0 - y) / (\text{effect}_0 - \text{effect}_\infty)\). The zero-interaction expectation is Bliss independence on that scale: [ f_{\text{ZIP}} = f_h + f_v - f_h f_v ] (equivalently, the Bliss combination of the two single-drug Hill curves in response units). This is what the Reference view labels “ZIP reference (zero-interaction expectation)”.
  3. For each row (fixed vertical concentration, varying horizontal drug), fit a three-parameter Hill curve in fraction-affected space: baseline \(f_0\) fixed to the other drug’s single-agent fraction at that row, with free potency, Hill slope, and maximum effect (capped at full effect). For a 5-parameter (asymmetric) drug the slice curve keeps that drug's fitted asymmetry, so a combination with no interaction scores exactly zero. This is a synfit extension: the published method uses symmetric 4-parameter curves, and the two are identical for 4-parameter drugs.
  4. Do the same for each column (vertical drug vs concentration).
  5. The fitted combination fraction at cell \((i,j)\) is the average of the row slice and column slice predictions at that cell. The heatmap δ-score compares this fitted surface to \(f_{\text{ZIP}}\), normalised by the response scale.

This is intermediate between purely probabilistic nulls (Bliss, HSA) and strict dose additivity (Loewe): it allows potency and efficacy shifts between drugs without refitting full single-agent curves at every combination.

Delta score

\[ \Delta_{\text{ZIP}} = \frac{y_{\text{fitted}} - y_{\text{ZIP}}}{\text{effect}_0 - \text{effect}_{\infty}} \]

where \(y_{\text{ZIP}}\) is the zero-interaction expectation and \(y_{\text{fitted}}\) is the averaged slice-fit combination surface (both in response units). synfit uses the equivalent form \(-(f_{\text{fitted}} - f_{\text{ZIP}})\) in fraction-affected space.

Delta Interpretation
Negative Synergy — stronger effect than ZIP predicts
Near zero Additivity in the ZIP sense
Positive Antagonism — weaker effect than ZIP predicts

For inhibition assays, \(\text{effect}_0\) is the uninhibited response and \(\text{effect}_{\infty}\) the fully inhibited plateau; for activation, the asymptotes are swapped but the same formula and sign convention apply.

Comparison with other models

Property Bliss HSA Loewe CI ZIP
Null idea Independent probabilities Stronger single agent Dose additivity Zero interaction in \(f\), then slice-fit combination
Uses the fitted single-drug Hill curves? No (observed edges) No (observed edges) Yes Yes (for the zero-interaction expectation and the slice anchors)
Per-row/column extra fits? No No No Yes (3-parameter slices)
Undefined cells? No No Yes (outside single-agent range) Only where both the row and the column curve fit failed (a warning names the failed rows/columns)

Further reading

Implementation details

Core (synfit.zip):

  • zip_reference(...) — Bliss combination of the supplied single-drug Hill parameters in fraction-affected space, mapped back to responses (zero-interaction expectation).
  • zip_fitted_surface(...) — per-row and per-column 3-parameter slice fits in normalised space; fitted combination is the mean of the two slice predictions; maps to response units.
  • zip_delta(...) — \(\delta = -(f_{\text{fitted}} - f_{\text{ZIP}})\) (negative = synergy). Cells where either concentration is zero are NaN. If a slice fit fails, the cell uses the other direction's fit alone; if both fail the cell is NaN (unscored). zip_scores(...) returns all of this plus failed_rows, failed_cols and n_unscored, which the app shows as a warning above the synergy panel. Optional keyword-only asymmetry_hor / asymmetry_ver for 5p drugs.

Plotting (synfit.plotting):

  • zip_heatmap(...) — calls zip_delta(), then matrix_heatmap() with RdBu centred at 0 (same diverging scale as the Bliss deviation heatmap: negative δ toward red, positive toward blue).