Skip to content

Hill Equation (4-Parameter)

The standard sigmoidal dose-response model — used for both inhibition (response falls with dose) and activation (response rises with dose). Given a drug concentration \([D]\), the expected response is:

\[ y = E_\infty + \frac{E_0 - E_\infty}{1 + \left(\frac{[D]}{C_{50}}\right)^{n_H}} \]

where \(E_0\) is the response at zero drug and \(E_\infty\) is the response at saturating drug. The same equation handles both directions — only the assignment of the two asymptotes to Top/Bottom flips.

Parameters

Parameter Meaning Typical range
C₅₀ Concentration at half-maximal effect — reported as IC₅₀ for inhibition fits, EC₅₀ for activation fits 0.01–1000 (drug-dependent)
Hill coefficient (\(n_H\)) Steepness of the sigmoidal transition 0.5–4
Top Higher of the two asymptotes (response at zero drug for inhibition; at saturating drug for activation) ~1.0 (normalized)
Bottom Lower of the two asymptotes (response at saturating drug for inhibition; at zero drug for activation) ~0.0 (normalized)
  • C₅₀ is the headline potency number. It is the same parameter in both directions; only the display label changes (IC₅₀ vs EC₅₀).
  • Hill coefficient reflects cooperativity. \(n_H = 1\) means simple binding; \(n_H > 1\) means steep, switch-like response.
  • Top and Bottom define the dynamic range. For normalized data, Top ≈ 1.0 and Bottom ≈ 0.0.

Log-space C₅₀

Internally, the optimizer works with \(\log_{10}(C_{50})\) rather than C₅₀ directly. C₅₀ values span orders of magnitude (e.g. 0.01 µM to 100 µM), and log-space optimization is numerically better behaved.

Inhibition vs Activation

The Hill equation is fully symmetric between the two directions — same parameters, same equation, same fitter. The only differences are practical:

  • Inhibition: response falls with dose. Low-conc asymptote is high (Top), high-conc asymptote is low (Bottom). C₅₀ is reported as IC₅₀.
  • Activation: response rises with dose. Low-conc asymptote is low (Bottom), high-conc asymptote is high (Top). C₅₀ is reported as EC₅₀.

Top and Bottom always refer to the higher and lower asymptote on the response axis; the direction tag tells the fitter which end of the dose axis each maps to.

5-Parameter (Asymmetric) Hill Equation

An extended model that adds an asymmetry parameter \(S\):

\[ y = E_\infty + \frac{E_0 - E_\infty}{\left(1 + \left(\frac{[D]}{\kappa}\right)^{n_H}\right)^S} \]

When \(S = 1\), this reduces to the standard 4-parameter model. \(S \neq 1\) allows the upper and lower transitions to have different shapes.

κ is not the half-max C₅₀ when S ≠ 1

In the 4-parameter form, the concentration parameter is the half-maximal C₅₀ (i.e. IC₅₀ for inhibition, EC₅₀ for activation) — the dose at which the response sits exactly between \(E_0\) and \(E_\infty\). Under the asymmetric form, that is no longer true: the parameter \(\kappa\) and the actual half-max concentration are related by

\[ C_{50} = \kappa \cdot \left(2^{1/S} - 1\right)^{1/n_H} \]

For \(S = 1\) the factor collapses to 1 and \(\kappa = C_{50}\). For \(S \neq 1\) the two can disagree by half a log unit or more (e.g. \(S = 2,\ n_H = 1\) gives \(C_{50} \approx 0.41\,\kappa\)).

That is why the 5-parameter fit reports the parameter as \(\kappa\) rather than IC₅₀ / EC₅₀ — to avoid implying a half-max interpretation that the asymmetric formula doesn't actually carry.

Further Reading

Implementation details

Core functions (synfit.hill):

  • hill_curve(conc, c50, hill=1.0, effect_0=1.0, effect_inf=0.0, asymmetry=1.0) — unified 4p/5p Hill equation. When asymmetry=1 (default), this is the standard 4-parameter curve; other positive values use the asymmetric (5p) exponent on the denominator. Works for inhibition or activation by swapping effect_0 and effect_inf.
  • log_wall(x, boundaries, steep=1) — soft bound penalty used as a log-prior during optimization (see Optimization).

Data classes (synfit.data):

  • FitConfig — initial log_c50, hill, effect_0, effect_inf, optional asymmetry; fitting_parameters lists which of these are optimized (include "asymmetry" for the 5p model); bounds (FitBounds); error_model; direction ("inhibition" or "activation"). There is no separate model enum — 4p vs 5p is determined solely by fitting_parameters.
  • FitResult — fitted c50, log_c50, hill, effect_0, effect_inf, optional asymmetry, plus r2, aic, bic, log_likelihood, rss. to_dict() exposes ic50 (always) and adds ec50 when direction == "activation".
  • FitBounds — per-parameter bound tuples. Defaults: log_c50=(-5, 5), hill=(0.1, 4.0), effect_0=(0, 2), effect_inf=(0, 2), asymmetry=(0.1, 10).

Fitter class (synfit.single):

  • SingleDrugFit(data, config=None) — main fitter. Takes a DataFrame with columns concentration, y, replicate. If no config is provided, initial parameters and bounds are auto-derived from the data via _init_config_from_data().
  • SingleDrugFit.fit(valids=None) → FitResult — runs scipy minimize, computes R², AIC, BIC.
  • SingleDrugFit.predict(result) → numpy array — evaluate the fitted curve at all data concentrations.
  • _init_config_from_data(data, direction) — heuristic IC₅₀/EC₅₀ concentration guess from the midpoint response; sets effect_0 / effect_inf and bounds from the observed range (with inhibition vs activation handling).