Bliss Independence¶
The Bliss independence model is the default null hypothesis for drug combination analysis in synfit. It predicts the combined response of two drugs assuming they act independently on the same biological target.
The Model¶
If drug A alone produces fractional response \(E_A\) and drug B alone produces \(E_B\), the Bliss-predicted combined response is:
where each \(E\) is the fractional response from the single-drug Hill curve (1 = no effect, 0 = full inhibition).
The predicted response matrix is the outer product of the two single-drug response vectors, rescaled to \([\text{bottom}, \text{top}]\).
Activation assays¶
For activators, work with the fractional effect \(E = (y - \text{baseline}) / (\text{maximum} - \text{baseline})\) (0 = no effect, 1 = full activation). The Bliss expectation is the probability that at least one drug produces the effect:
This is the same model as the inhibition form above, written in effect space: synfit uses the fitted baseline and maximum as the two asymptotes, so both directions share one formula and one sign convention. For example, single-agent activation of 40 % and 60 % gives a Bliss expectation of 76 %.
Bliss needs a meaningful normalisation
Bliss treats the normalised response as a probability of effect. That holds when the baseline and maximum are real plateaus of the assay (untreated control and full effect). Rescaling an arbitrary signal, such as raw fluorescence without a defined maximum, does not give it that meaning, so check that the fitted asymptotes are the assay's true floor and ceiling before reading Bliss scores.
Deviation from Bliss¶
The Bliss deviation at each concentration pair compares the observed response to the Bliss prediction:
where \(\text{effect}_0\) is the zero-drug response and \(\text{effect}_{\infty}\) the saturating response (inhibition or activation).
| Deviation | Interpretation |
|---|---|
| Negative | Synergy — stronger effect than Bliss independence predicts |
| Near zero | Independence — drugs act independently |
| Positive | Antagonism — weaker effect than Bliss independence predicts |
The Deviation Heatmap¶
The Bliss deviation heatmap visualizes \(\Delta_{\text{Bliss}}\) for every concentration pair in the matrix. Rows correspond to one drug's concentrations, columns to the other's.
The app colours the heatmap on a diverging scale centred at zero: negative deviations (synergy) are blue, positive deviations (antagonism) are red, and white is near independence. (The synfit library's own matplotlib plot uses RdBu, where the colours are the other way round.)
Strong synergy or antagonism at specific concentration ratios reveals the pharmacologically interesting region of the combination space.
Matrix Fitting Pipeline¶
- Edge extraction — the first row and first column of the matrix (where one drug is at zero concentration) are the single-drug dose-response data. The combination cells are not used for fitting.
- Joint single-agent fit — both edges are fitted together: each drug gets its own potency (C50) and Hill slope (and asymmetry for a 5-parameter model), while the two plateaus (top and bottom) are shared, because both drugs are measured on the same plate against the same controls. Nothing about the drug-drug interaction enters this fit.
- Bliss prediction — the Bliss-predicted matrix is computed via
bliss_referencefrom the observed single-agent edges (the matrix's zero-concentration row and column), normalised with the fitted plateaus as the zero-drug and saturating responses. - Deviation — the normalised difference between observed and predicted values, as above.
The fitted Hill curves from step 2 are used by Loewe and ZIP; Bliss and HSA use the observed edges directly.
Further Reading¶
- Bliss, C.I. (1939) "The toxicity of poisons applied jointly" — the original paper defining the independence model
- Wikipedia: Drug combination — overview of combination analysis approaches
Implementation details
Bliss module (synfit.bliss):
bliss_reference(resp_horizontal, resp_vertical, effect_0, effect_inf)→np.ndarray— preferred for analysis views. Takes the observed single-drug marginals (typicallymean_matrix[0, :]andmean_matrix[:, 0]) and returns the outer-product Bliss prediction. Normalized survivals are clipped to \([0, 1]\) before multiplication so noisy or super-inhibited marginals cannot push the reference outside the response window. Likehsa_reference, it is built from the observed single-agent edges. Returns shape(len(resp_vertical), len(resp_horizontal)).bliss_independence(conc_hor, conc_ver, c50_hor, c50_ver, hill_hor, hill_ver, effect_0, effect_inf, asymmetry_hor=1.0, asymmetry_ver=1.0)→np.ndarray— model-based variant used as the forward model inside the joint matrix likelihood (MatrixFit._log_prob_data). Evaluateshill_curve(5-parameter ifasymmetry_* ≠ 1) at the concentration grids, clips to \([0, 1]\), and multiplies. Returns shape(len(conc_ver), len(conc_hor)).bliss_deviation(data, bliss, effect_0, effect_inf)→np.ndarray— normalized deviation:(data - bliss) / (effect_0 - effect_inf). Negative = synergy (stronger effect than Bliss predicts); positive = antagonism.
Matrix fitter (synfit.matrix):
MatrixFit(replicates, conc_horizontal, conc_vertical, valids=None, error_model="lognormal")— constructor. Validates replicate shapes, computesmean_matrix, extracts edge slices via_edge_slice(), pre-fits edges withSingleDrugFit.MatrixFit._edge_slice(axis)— for"horizontal", uses the first column of each replicate (rep[:, 0]) withconc_vertical; for"vertical", uses the first row (rep[0, :]) withconc_horizontal, building a tidy DataFrame forSingleDrugFit.MatrixFit._PARAM_NAMES = ("log_c50_hor", "log_c50_ver", "hill_hor", "hill_ver", "effect_0", "effect_inf")— the 6 jointly optimized parameters.MatrixFit.fit()→MatrixFitResult— runsscipy.optimize.minimizeon the joint objective (sum of log-likelihoods over all replicates under the Bliss model).MatrixFitResult— containshorizontalandverticalFitResultobjects plus sharedeffect_0,effect_inf,success,message.