Reference Surfaces (Null Predictions)¶
Every synergy model in synfit works by comparing the observed combination matrix against a reference surface — the response matrix you would expect under the model's null hypothesis of no interaction. Deviations from that reference are what the synergy heatmap actually displays.
The Deviation / Reference toggle in the Dataset Detail view lets you inspect either half of that comparison side-by-side for all four models.
What each reference represents¶
| Model | Null hypothesis | Reference surface (what the app shows) |
|---|---|---|
| Bliss Independence | Drugs act independently in fractional-effect space | Outer product of the observed single-agent edges (mean_matrix[0, :] and mean_matrix[:, 0]), rescaled with the fitted asymptotes via bliss_reference — same marginals as the data at zero-concentration rows/columns. |
| HSA (Highest Single Agent) | The combination is no stronger than the better single drug | Elementwise stronger of the two observed marginals — lower response for inhibition, higher for activation (hsa_reference on the same edges). |
| Loewe Additivity | Dose equivalence: \(\dfrac{c_\text{hor}}{A(E)} + \dfrac{c_\text{ver}}{B(E)} = 1\) | Per-cell root find for the effect \(E^*\) that satisfies the equation above, using the inverse Hill curves \(A(E)\), \(B(E)\) of the two single-drug fits. Boundary cells reduce to the corresponding single-drug Hill prediction. |
| ZIP (Zero Interaction Potency) | Zero interaction in fraction-affected space | Bliss combination of the single-drug Hill parameters (zip_reference) — the zero-interaction expectation \(y_{\text{ZIP}}\). The δ-score heatmap compares the fitted combination surface (averaged row/column slice fits) to that expectation, not the average of slice surfaces. |
Why Loewe is special¶
Bliss, HSA and ZIP all define their null directly as a response surface — a matrix of "expected \(y\) under no interaction" that you can read off a grid. Loewe's null is instead a constraint on the doses at each cell (\(\text{CI} = 1\)), so the deviation heatmap shows the CI value itself rather than a response deviation.
If you asked "what would the CI matrix look like if the null held?" the answer is degenerate: a matrix of ones. That's not a useful picture. So the Loewe reference surface shown here is the effect \(E^*\) that the null predicts at each concentration pair — found by solving the dose-equivalence equation per cell — giving a proper response surface directly comparable to the other three references.
Reading the two views¶
- Deviation (default) —
BWRdiverging palette centred on the null. Blue / red indicate synergy / antagonism. This is the same heatmap the page has always shown. - Reference — plasma palette on the observed-matrix range, so the reference renders on exactly the same colour scale as the Preview heatmap above. Flip between the Preview, the Reference, and the observed matrix to judge visually how far the combination departs from each model's null prediction.
A flat reference that looks nothing like the observed matrix is a strong qualitative signal of interaction — often clearer than a single summary number.
Implementation¶
Reference surfaces for the Dataset Detail API are built in share_payload.py: Bliss and HSA from observed marginals (bliss_reference, hsa_reference), Loewe from fitted marginals (loewe_reference), ZIP from the zero-interaction expectation (zip_reference). Deviation matrices use the same direction-neutral effect_0 / effect_inf asymptotes as the core *_deviation helpers.