Choosing Direct, Turnover, Or Tolerance Response SDEs

Choosing a Response Structure From Its Time Course

A drug concentration can rise quickly while the measured response changes slowly, adapts during continued exposure, or remains altered after concentration has fallen. These patterns are clues to the response mechanism, but they are not, by themselves, proof of one mechanism. The model should separate the exposure profile, the response mechanism, and stochastic variation, then be checked against observed trajectories.

This tutorial compares three candidate structures. A direct-effect model has a response mean determined by current exposure, with stochastic fluctuations around that mean. An indirect-response or turnover model lets the drug alter production or loss of a response that turns over over time. A tolerance model lets the effect of a given exposure diminish as an adaptive state changes. These are structural distinctions: changing the noise term alone does not turn one structure into another.

Definition: Let $C_t$ be the concentration driving the response and $Y_t$ the observed response. A direct-effect response SDE has a response mean $m(C_t;\theta)$ that depends on current concentration, with any additional stochastic state not introducing a drug-driven delay. A turnover response SDE has a latent response $R_t$ whose drift contains production and loss terms, at least one of which depends on $C_t$. A tolerance response SDE includes an adaptive state that changes the response to a given concentration as exposure continues and recovers when exposure wanes.

The descriptions are operational model definitions, not claims that every pharmacological mechanism belongs to exactly one category. For example, a direct-effect model can have serially correlated background variation, and a turnover model can include stochastic fluctuations. The distinction is whether drug-related delay or adaptation is represented in the response mechanism.

Direct Effect: Current Exposure Sets the Mean

A useful direct-effect specification is

Y_t=m(C_t;\theta)+Z_t,\qquad dZ_t=-\lambda Z_t\,dt+\sigma\,dW_t,\qquad \lambda>0.

Here $m(C_t;\theta)$ might be a baseline plus a saturable exposure effect, $Z_t$ is a mean-zero Ornstein–Uhlenbeck fluctuation, and $W_t$ is Brownian motion. The fluctuation can create correlated departures from the concentration-response curve, but its drift does not encode a delayed drug effect. If $C_t$ changes abruptly, the conditional mean $m(C_t;\theta)$ changes at that time.

The model assumes the concentration reaching the relevant effect site is adequately represented by $C_t$. If effect-site equilibration is needed, adding a lag state changes the structure: the response mean then depends on a filtered exposure rather than current concentration. A direct-effect model is most plausible when response onset and offset track exposure closely, allowing for sampling resolution and measurement noise.

Worked Example: A Direct Saturable Response

Suppose the mean response is $m(C)=50+10C/(2+C)$, with $C$ in concentration units. At $C=2$, the drug-related increment is $10(2)/(2+2)=5$, so the mean response is $55$. At $C=1$, it is $50+10/3=53.33$. If concentration falls from 2 to 1, the model predicts an immediate change in the mean from 55 to 53.33, not a gradual relaxation between these values.

Suppose the stochastic fluctuation starts at $Z_0=0$, and $\lambda=0.5$ per hour and $\sigma=0.3$ response units per square-root hour. The OU solution has conditional mean $E[Z_t\mid Z_0=0]=0$ and variance

\operatorname{Var}(Z_t\mid Z_0=0) =\sigma^2\int_0^t e^{-2\lambda(t-s)}\,ds =\frac{0.09}{1.0}(1-e^{-t}).

At $t=4$ hours this is $0.09(1-e^{-4})\approx0.08835$, and the standard deviation is approximately $0.2972$. This uncertainty describes stochastic fluctuation around the current mean; it does not create a systematic concentration-dependent delay.

Turnover: Drug-Modified Production or Loss

An indirect-response model represents a response that is continually produced and removed. A simple stochastic form is

dR_t=\{k_{\mathrm{in}}(C_t)-k_{\mathrm{out}}(C_t)R_t\}\,dt+\sigma_R\,dW_t, \qquad k_{\mathrm{out}}(C_t)>0.

The drift is production minus loss. Without drug, if both rates are constant, the baseline equilibrium is $R_0=k_{\mathrm{in},0}/k_{\mathrm{out},0}$. Drug action can inhibit or stimulate production, or inhibit or stimulate loss. These alternatives can produce different response directions and recovery patterns; they should be chosen to match the pharmacological hypothesis and checked against data. With additive diffusion, the model is a local approximation when negative response values would otherwise be implausible; a positive-state formulation may be needed for responses constrained to be positive.

Theorem: Suppose $k_{\mathrm{in}}$ and $k_{\mathrm{out}}>0$ are constant over an interval, $\sigma_R$ is constant, and $R_0$ is the response at the start of the interval. Then, for elapsed time $t\geq0$, $$ E[R_t\mid R_0]=R_\infty+(R_0-R_\infty)e^{-k_{\mathrm{out}}t}, \qquad R_\infty=\frac{k_{\mathrm{in}}}{k_{\mathrm{out}}}. $$ If the initial value is fixed, the conditional variance is $$ \operatorname{Var}(R_t\mid R_0)= \frac{\sigma_R^2}{2k_{\mathrm{out}}} \left(1-e^{-2k_{\mathrm{out}}t}\right). $$

Proof. Write $k=k_{\mathrm{out}}$ and multiply the SDE by the integrating factor $e^{kt}$. The product rule gives

d(e^{kt}R_t)=k_{\mathrm{in}}e^{kt}\,dt+\sigma_R e^{kt}\,dW_t.

Integrating from 0 to $t$, then multiplying by $e^{-kt}$, yields

R_t=R_0e^{-kt}+\frac{k_{\mathrm{in}}}{k}(1-e^{-kt}) +\sigma_R\int_0^t e^{-k(t-s)}\,dW_s.

The Itô integral has conditional mean zero, so taking the conditional expectation gives the stated mean. By the Itô isometry, its variance is

\sigma_R^2\int_0^t e^{-2k(t-s)}\,ds =\frac{\sigma_R^2}{2k}(1-e^{-2kt}),

which proves the variance formula. $\square$

Worked Example: Turnover Inhibition and Recovery

Let the baseline response be $100$ units and $k_{\mathrm{out}}=0.2$ per day. Then $k_{\mathrm{in},0}=0.2(100)=20$ units per day. During constant exposure, suppose drug inhibits production by 50%, so $k_{\mathrm{in}}=10$, while $k_{\mathrm{out}}$ is unchanged. The on-drug equilibrium is $10/0.2=50$ units. Starting at 100, the expected response after five days is

50+(100-50)e^{-0.2(5)} =50+50e^{-1} \approx 68.39.

If exposure then stops and production returns to 20, the response moves toward the baseline 100 with the same time constant $1/0.2=5$ days. One day after stopping, starting from 68.39, the expected response is

100+(68.39-100)e^{-0.2(1)} \approx100-31.61(0.81873) \approx74.12.

Thus the response does not jump back to baseline when concentration vanishes. The lag and recovery follow from turnover, not from a persistent drug concentration in this example.

Tolerance: Effectiveness Changes During Exposure

A tolerance model is appropriate when the same concentration has a smaller effect after sustained exposure than it had initially. One parsimonious structure couples a response SDE to a deterministic adaptive state $Q_t$:

dR_t=\{k_R[R_0+E_{\max}h(C_t)Q_t-R_t]\}\,dt+\sigma_R\,dW_t, \qquad \frac{dQ_t}{dt}=k_{\mathrm{rec}}(1-Q_t)-k_{\mathrm{tol}}h(C_t)Q_t.

Here $h(C)$ is a dimensionless exposure function between 0 and 1, $Q_t$ is relative sensitivity, and $k_{\mathrm{rec}},k_{\mathrm{tol}}>0$. Under constant exposure, the adaptive state is pulled toward a lower equilibrium. When exposure is removed, $Q_t$ recovers toward 1. The response state $R_t$ can lag behind this changing target at a rate set by $k_R$. This formulation is a modeling choice; it does not imply that all tolerance mechanisms act through one measurable biological mediator.

Result: Under constant exposure with $h(C_t)=h$, where $0\leq h\leq1$, the tolerance state satisfies $$ Q_t=Q_\infty+(Q_0-Q_\infty)e^{-(k_{\mathrm{rec}}+k_{\mathrm{tol}}h)t}, \qquad Q_\infty=\frac{k_{\mathrm{rec}}}{k_{\mathrm{rec}}+k_{\mathrm{tol}}h}. $$

Derivation. For fixed $h$, collect the terms in the state equation:

\frac{dQ_t}{dt}=k_{\mathrm{rec}}- (k_{\mathrm{rec}}+k_{\mathrm{tol}}h)Q_t.

The equilibrium is the value making the right side zero, hence $Q_\infty=k_{\mathrm{rec}}/(k_{\mathrm{rec}}+k_{\mathrm{tol}}h)$. Subtract this equilibrium from both sides to obtain $d(Q_t-Q_\infty)/dt=-(k_{\mathrm{rec}}+k_{\mathrm{tol}}h)(Q_t-Q_\infty)$. Solving this scalar linear equation gives the stated expression. Since the exponential factor is positive, $Q_t$ moves monotonically from $Q_0$ toward $Q_\infty$ when those values differ. When exposure is removed, $h=0$, so $Q_t=1+(Q_0-1)e^{-k_{\mathrm{rec}}t}$. $\square$

Worked Example: Adaptation During Sustained Exposure

Suppose $h(C)=0.8$ during dosing, $k_{\mathrm{rec}}=0.1$ per day, $k_{\mathrm{tol}}=0.3$ per day, and $Q_0=1$. The effective adaptation rate is $0.1+0.3(0.8)=0.34$ per day, and the plateau is $Q_\infty=0.1/0.34\approx0.2941$. After five days,

Q_5=0.2941+(1-0.2941)e^{-0.34(5)} \approx0.2941+0.7059(0.18268) \approx0.4231.

For $E_{\max}=20$, the drug-related target increment is initially $20(0.8)(1)=16$ units and at day five is $20(0.8)(0.4231)\approx6.77$ units. If $R_0=100$, the corresponding response targets are 116 and about 106.77. These are targets in the drift, not exact response means at every time: $R_t$ approaches the moving target at rate $k_R$, and its SDE also contributes variability. After dosing stops, $Q_t$ recovers toward 1 at rate $0.1$ per day, so sensitivity can recover more slowly than it declined.

Using Observed Patterns to Choose and Check a Model

Observed pattern	Structure to examine first	Important qualification
Response mean tracks current concentration with little systematic lag	Direct effect	Check whether sampling is frequent enough to detect a delay.
Response rises or falls gradually and recovers after concentration declines	Turnover	Different production and loss mechanisms can have similar trajectories.
Effect weakens during sustained exposure and returns after washout	Tolerance	Accumulation, changing concentration, or a delayed response can mimic adaptation.
Response-concentration loops differ on rising and falling exposure	Turnover or tolerance; possibly effect-site delay	Hysteresis alone does not identify its cause.

The main alternative to selecting one structure is to fit several scientifically plausible candidates and compare their predictive performance. The earlier material on calibration and model comparison provides tools for that task; the structural checks here help define candidates worth comparing. Useful diagnostics include plotting observed response against concentration with arrows showing time direction, inspecting onset and washout segments separately, and checking whether residuals retain systematic time patterns after fitting. A model that fits the overall response range but consistently misses early onset or late recovery is structurally suspect.

Do not attribute every discrepancy to stochastic noise. A larger diffusion coefficient can widen predictive intervals, but it does not correct a systematic delayed response, exposure-dependent attenuation, or biased recovery trajectory. Conversely, a few noisy observations do not establish tolerance: the proposed adaptation should improve prediction across repeated exposure and washout patterns, not only explain one trajectory.

Common Pitfalls and Limitations

Calling every lag “tolerance.” Turnover and effect-site delay can also create hysteresis. A tolerance claim needs evidence that sensitivity changes with exposure history after accounting for concentration and response turnover.
Confusing the noise state with a drug-effect state. Correlated OU fluctuations can create smooth deviations, but they do not represent drug-driven adaptation unless their dynamics explicitly depend on exposure.
Ignoring rate identifiability. Sparse observations around onset and washout may not distinguish a slow turnover rate from a delayed effect site. Sampling design and repeated dosing periods matter.
Over-interpreting fitted parameters. A fitted production or tolerance parameter describes the chosen model. It need not equal a separately measured biological rate or mechanism.
Using an unsuitable state space. Additive Gaussian noise can assign nonzero probability to impossible response values. Consider whether a positive or bounded state formulation is required, and check simulated trajectories for boundary violations.

These models are idealizations. Real responses can combine direct effects, turnover, and adaptation; concentration may be measured with error; and the relevant effect-site concentration may be unobserved. The practical goal is not to force every dataset into one label. It is to choose a structure whose assumptions are explicit, whose predicted timing is compatible with the data, and whose uncertainty is adequately calibrated.

Check Your Understanding

Answer each question, then check the worked solutions below.

A response mean changes as soon as measured concentration changes, while correlated residual fluctuations persist. Which candidate structure is most consistent, and why do the fluctuations not imply turnover?
A turnover response starts at 80 units, has a new equilibrium of 40 units, and $k_{\mathrm{out}}=0.1$ per hour under constant exposure. What is its expected value after 10 hours?
For a tolerance state, $k_{\mathrm{rec}}=0.2$, $k_{\mathrm{tol}}=0.4$, and $h=0.5$. Find the adaptation equilibrium and rate of approach.
Why does a response-concentration hysteresis loop not, by itself, prove tolerance?

Solutions. (1) A direct-effect model with a correlated stochastic fluctuation is consistent: its mean follows current concentration, while the OU fluctuation accounts for serially correlated departures. The fluctuation is not turnover because it does not make the drug alter response production or loss. (2) Apply the constant-coefficient mean formula: $R_{10}=40+(80-40)e^{-0.1(10)}=40+40e^{-1}\approx40+14.72=54.72$ units. (3) The rate is $k_{\mathrm{rec}}+k_{\mathrm{tol}}h=0.2+0.4(0.5)=0.4$ per time unit; the equilibrium is $0.2/0.4=0.5$. Thus $Q_t=0.5+(Q_0-0.5)e^{-0.4t}$. (4) Turnover or effect-site delay can also make response lag concentration; a loop records time dependence but does not uniquely identify its mechanism. Repeated exposure and washout data, with adequate sampling and model checks, are needed to support tolerance.