Identifiability of IC50 and Hill Coefficient

Why IC50 and the Hill Coefficient Can Be Hard to Separate

In an indirect response model, drug concentration changes a production or loss rate, and the measured response then changes over time as the system turns over. Estimating the concentration-response parameters therefore involves two linked questions: whether the observed concentrations distinguish the concentration-effect curve, and whether response measurements capture the resulting indirect change. A rich response time course at one fixed concentration may describe the time course well while leaving the curve’s IC50 and Hill coefficient poorly determined.

The earlier material on Hill effects and nonlinear indirect responses introduced concentration-dependent rate modulation. Here the focus is identifiability: which parts of the concentration range and response trajectory contain information about IC50 and the Hill coefficient, and what patterns warn that those parameters cannot be estimated separately with useful precision.

Definition: Let $C(t)$ be drug concentration and let $q(C)$ be the fractional concentration-dependent inhibition of a production rate: $$ q(C)=\frac{C^h}{IC_{50}^h+C^h}, \qquad C\geq 0,\ IC_{50}>0,\ h>0. $$ Here $IC_{50}$ has concentration units and is the concentration at which $q=0.5$; $h$, the Hill coefficient, is dimensionless and controls the steepness of the curve. If $I_{\max}$ is the maximum fractional inhibition, the remaining fraction of the baseline production rate is $1-I_{\max}q(C)$, with $0\leq I_{\max}\leq1$.

For specificity, consider inhibition of response production:

\frac{dR}{dt}=k_{\mathrm{in}}\left[1-I_{\max}q(C(t))\right]-k_{\mathrm{out}}R.

The model assumes a stable $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$, a constant maximum inhibition, and a known concentration history that adequately represents the concentration driving the effect. In the absence of drug, the baseline is $R_0=k_{\mathrm{in}}/k_{\mathrm{out}}$, if the system is at steady state. The model includes turnover delay but no additional effect-site delay or time-varying sensitivity. If those assumptions are not adequate, apparent difficulty estimating $IC_{50}$ or $h$ may reflect model misspecification as well as limited information.

The parameters have different roles. $IC_{50}$ locates the curve horizontally: increasing it shifts the response transition toward higher concentrations. The Hill coefficient changes the curve’s steepness around that transition. At very low concentrations, inhibition is small; at very high concentrations, inhibition approaches $I_{\max}$. Observations confined to either extreme tend to show little about the transition’s location or steepness.

Concentration-Response Sensitivity

A direct way to see which concentrations inform each parameter is to differentiate the Hill function. Write $q(C)$ as a logistic function of the log concentration ratio:

q(C)=\frac{1}{1+\exp[-h\log(C/IC_{50})]}.

For positive $C$, its sensitivities to the log of $IC_{50}$ and to $h$ are:

\frac{\partial q}{\partial \log(IC_{50})}=-h q(1-q), \qquad \frac{\partial q}{\partial h}=q(1-q)\log(C/IC_{50}).

These derivatives explain the design problem. The sensitivity to $IC_{50}$ is small when $q$ is near 0 or 1, and is largest near $q=0.5$, the transition region. The sensitivity to $h$ is exactly zero at $C=IC_{50}$, because all Hill curves with the same $I_{\max}$ pass through $q=0.5$ there. Estimating steepness therefore requires observations on the flanks of the transition, not just at its midpoint. Concentrations on both sides help distinguish a shift in location from a change in steepness.

This is a sensitivity argument, not a guarantee that a particular design will yield precise estimates. Measurement error, uncertainty in $I_{\max}$, concentration error, and the indirect response dynamics also affect the information available. Nevertheless, observations only far below or far above $IC_{50}$ are generally a poor basis for estimating both parameters.

Worked Example: A Concentration Range That Spans the Transition

Consider a hypothetical inhibitory model with $IC_{50}=4$ mg/L, $h=2$, $I_{\max}=0.80$, and baseline response $R_0=100$ response units. Suppose each listed concentration is maintained long enough for the response to reach steady state. Calculate the fractional inhibition and steady-state response at 1, 2, 4, 8, and 16 mg/L.

At steady state, setting $dR/dt=0$ gives $R/R_0=1-I_{\max}q(C)$. For $C=2$ mg/L, for example:

q(2)=\frac{2^2}{4^2+2^2}=\frac{4}{20}=0.2000, \qquad R=100[1-(0.80)(0.2000)]=84.00\ \text{response units}.

The same calculation at the other concentrations gives:

\begin{aligned} q(1)&=\frac{1}{17}=0.05882, & R(1)&=100[1-(0.80)(0.05882)]=95.29,\\ q(4)&=\frac{16}{32}=0.5000, & R(4)&=100[1-(0.80)(0.5000)]=60.00,\\ q(8)&=\frac{64}{80}=0.8000, & R(8)&=100[1-(0.80)(0.8000)]=36.00,\\ q(16)&=\frac{256}{272}=0.9412, & R(16)&=100[1-(0.80)(0.9412)]=24.71. \end{aligned}

The concentrations include a low-effect point, the midpoint, and points on both flanks, with a higher concentration approaching the inhibition plateau. Such coverage can help separate the curve’s location from its steepness, provided response measurements are precise and the model assumptions are reasonable. These are model predictions for a hypothetical example, not observations from a study.

Why Indirect Response Sampling Matters

A concentration does not instantaneously set the measured response in this model. It changes the production rate, after which turnover moves $R$ toward a new level. For a constant concentration applied at time zero, with $R(0)=R_0$, the normalized response $x(t)=R(t)/R_0$ satisfies:

\frac{dx}{dt}=k_{\mathrm{out}}\left[1-I_{\max}q(C)-x\right].

Solving this first-order equation gives a response that approaches its drug-specific steady state gradually. For time-varying concentration, the response reflects the preceding exposure history as well as the current concentration:

x(t)=e^{-k_{\mathrm{out}}t}x(0) +k_{\mathrm{out}}\int_0^t e^{-k_{\mathrm{out}}(t-s)} \left[1-I_{\max}q(C(s))\right]\,ds.

This expression makes the sampling issue explicit: recent rate effects receive more weight than distant ones, with the weighting controlled by $k_{\mathrm{out}}$. A response measurement taken soon after a concentration change may show only part of the eventual effect. Conversely, a long exposure at one fixed concentration may characterize its response level and the turnover trajectory, but it does not by itself reveal where the concentration lies on the Hill curve.

Worked Example: A Delayed Response After a Concentration Step

Suppose a hypothetical system starts at baseline, with $I_{\max}=0.80$, $IC_{50}=4$ mg/L, $h=2$, and $k_{\mathrm{out}}=0.20$ per hour. At time zero, concentration steps to 4 mg/L and remains there. Determine the expected normalized response at 1, 5, and 20 hours.

At 4 mg/L, $q=0.5$, so the drug-specific steady-state normalized response is $1-(0.80)(0.5)=0.60$. Since $x(0)=1$, the step-response solution is:

x(t)=0.60+(1-0.60)e^{-0.20t} =0.60+0.40e^{-0.20t}.

Substitution gives $x(1)=0.60+0.40e^{-0.2}=0.9275$, $x(5)=0.60+0.40e^{-1}=0.7472$, and $x(20)=0.60+0.40e^{-4}=0.6073$. Thus the response remains close to baseline at 1 hour even though the concentration is at $IC_{50}$; it approaches the steady-state value only over time.

If the 1-hour response were incorrectly treated as a steady-state observation, the inferred fractional inhibition would be $(1-0.9275)/0.80=0.0906$, far below the actual Hill-function value $q=0.5$. The discrepancy is due to turnover lag, not evidence that the Hill function has a different $IC_{50}$.

What Different Sampling Patterns Can Identify

A single concentration level generally provides only one value of $q(C)$, even if many responses are recorded over time. If $I_{\max}$ is known and that value is $q_*$ at concentration $C_*$, then every positive $h$ can be paired with an $IC_{50}$ that gives the same inhibition:

IC_{50}=C_*\left(\frac{1-q_*}{q_*}\right)^{1/h}.

The time course may help estimate the magnitude of the rate change and turnover, subject to the model and data, but repeated observations at that fixed concentration do not supply new concentration levels. The equation shows directly why $IC_{50}$ and $h$ can trade off.

A narrow concentration range can also leave the parameters highly correlated in practice. If all concentrations are far below the transition, then $q(C)$ is small and the response is close to baseline; if all are far above it, then $q(C)$ is near its plateau. In either case, modest changes in $IC_{50}$ or $h$ can have little effect on predicted responses over the observed range. A broad range spanning both flanks is more informative, particularly when $I_{\max}$ is also estimated.

Response sampling must complement concentration coverage. Measurements at a range of concentrations should be timed so their delayed consequences are distinguishable. For an exposure that changes quickly relative to turnover, early response observations may be dominated by prior exposure and the initial state. Measurements that extend toward the new response level can help characterize the size of the rate effect, while observations during the transition can inform its timing. If concentration changes over time, the model should use the exposure history; pairing each response only with the simultaneous concentration can be misleading.

Worked Example: Similar Fits Over a Low Concentration Range

Consider two hypothetical Hill curves with $I_{\max}=0.80$. Curve A has $h=1$ and $IC_{50}=4.95$ mg/L; curve B has $h=2$ and $IC_{50}=0.4975$ mg/L. Both give $q=0.0100$ at $C=0.050$ mg/L. Compare their predictions at three nearby low concentrations and the corresponding steady-state responses when $R_0=100$ units.

For curve A at $0.040$ mg/L, $q=0.040/(4.95+0.040)=0.008016$; at $0.060$ mg/L, $q=0.060/(4.95+0.060)=0.01198$. For curve B, $q=C^2/(0.4975^2+C^2)$, giving $q=0.006423$ at $0.040$ mg/L and $q=0.01434$ at $0.060$ mg/L. Both predict $q=0.0100$ at $0.050$ mg/L, to the shown precision.

Using $R/R_0=1-0.80q$, the predicted steady-state responses at 0.040, 0.050, and 0.060 mg/L are approximately 99.36, 99.20, and 99.04 units for curve A, and 99.49, 99.20, and 98.85 units for curve B. These curves have substantially different parameter values yet predict responses within about 0.64 units of each other across this narrow range. If response error is large relative to those differences, the data may not distinguish the parameter sets reliably. Adding measurements near and across the transition would be more informative than collecting many additional observations in the same low range.

Diagnostics, Alternatives, and Limitations

Weak separation of $IC_{50}$ and $h$ often appears as strongly correlated estimates, large standard errors, very wide uncertainty intervals, or estimates that move substantially when one concentration level is removed. Different parameter combinations may yield nearly indistinguishable fitted response trajectories over the observed range. A numerical optimizer returning a precise-looking point estimate does not establish that the data identify both parameters: inspect uncertainty and whether predictions change meaningfully when either parameter changes.

A useful design response is to expand the concentration range where scientifically and ethically appropriate, with levels below, near, and above the expected $IC_{50}$. If the Hill coefficient is not a primary target and the data cannot support its estimation, fixing $h$ to a justified value or simplifying the concentration-effect function may produce more stable estimates. This trades flexibility for stronger assumptions; results should be checked for sensitivity to those assumptions. More response samples at one concentration can clarify a time course, but they are not a substitute for concentration variation when the goal is to separate the two Hill parameters.

A direct concentration-response analysis of steady-state responses can be useful when steady state is genuinely attained at each concentration. For transient data, fitting the indirect response differential equation is more appropriate because it accounts for turnover and exposure history. Neither approach can recover information absent from the design. Further, uncertainty in measured or modeled concentrations, an incorrect effect compartment assumption, a misspecified turnover mechanism, or an uncertain $I_{\max}$ can make apparent $IC_{50}$-$h$ separation worse. Identifiability is therefore conditional on both the model and the available observations.

Common pitfall: “There are many response measurements, so $IC_{50}$ and $h$ must be well estimated.” This reasoning confuses the number of observations with the information they contain. Many measurements at one fixed concentration can characterize the indirect time course, but they repeatedly probe the same concentration-effect level. To distinguish curve location from steepness, the data generally need informative concentration variation as well as response sampling that captures turnover.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

What does $q=0.5$ mean at $C=IC_{50}$, and does that value alone determine the Hill coefficient?
Why are concentrations only far above $IC_{50}$ usually weak for estimating $h$?
A concentration is held fixed for a long time, and responses are sampled repeatedly. Why might this still fail to separate $IC_{50}$ from $h$?
In the step-response example, why is the normalized response at 1 hour not equal to its drug-specific steady-state value?
Name one diagnostic sign of weak practical separation between $IC_{50}$ and $h$, and one design remedy.

Solutions. (1) At $C=IC_{50}$, the fractional inhibition is one half, regardless of $h$; that point alone does not determine steepness. (2) At high concentrations, $q$ is near 1 and changes little as $h$ varies, so the response contains little information about steepness. (3) Repeated measurements at one concentration add temporal information but do not provide distinct levels of the concentration-effect curve; multiple $IC_{50}$-$h$ pairs can give the same $q$ there. (4) Turnover is gradual: $x(1)=0.9275$, whereas the drug-specific steady-state value is 0.60, so the response has not yet approached its limiting level. (5) Strong parameter correlation or wide uncertainty is a warning sign; adding concentration levels across the transition, including its flanks, is a possible remedy.