Why Exposure Coverage Matters
An escalation study can show a biomarker changing across dose cohorts without showing what happens at higher exposures. The response may still be approaching its maximum, or the maximum effect itself may be small. Those interpretations lead to different development questions, so the first task is to ask whether the observed exposures cover the part of the exposure-response curve that distinguishes them.
Dose levels alone do not answer that question. Participants at the same dose can have different exposures, and exposure distributions from adjacent cohorts can overlap. A useful assessment therefore examines observed exposure, the biomarker response at the relevant sampling time, and whether the range plausibly covers both the response transition and its upper, nearly saturated region.
Coverage Under a Direct Emax Model
For a direct, monotone exposure-response relationship, a standard Emax form is \(R(C)=R_0+E_{\max}C/(EC_{50}+C)\), where \(C\) is the relevant exposure, \(R_0\) is the response at zero exposure, \(E_{\max}\) is the signed maximum change from baseline, and \(EC_{50}\) is the exposure producing half of that maximum change. If the biomarker decreases, \(E_{\max}\) is negative; the fraction of the maximum response reached is still \(C/(EC_{50}+C)\).
This model is useful for coverage assessment only when it is a reasonable description of the measurement being analyzed. In particular, the exposure metric and biomarker sampling time should be relevant to the effect; the response should be adequately represented by a monotone direct relationship; and baseline, time-course effects, and repeated measurements should be handled appropriately. Earlier material on direct-effect Emax models, turnover, effect compartments, and repeated biomarker measurements addresses those modeling choices. If a delayed response is sampled at an unsuitable time, or a changing baseline is mistaken for drug effect, the apparent exposure coverage can be misleading.
The fraction of the maximum effect predicted at exposure \(C\) is \(f(C)=C/(EC_{50}+C)\). Solving for exposure at a chosen fraction \(q\) gives \(q=C/(EC_{50}+C)\), so \(qEC_{50}=C(1-q)\), and therefore \(C=qEC_{50}/(1-q)\). Thus, 50% of the maximum occurs at \(C=EC_{50}\); 80% occurs at \(C=4EC_{50}\); and 90% occurs at \(C=9EC_{50}\). These are descriptive landmarks, not universal decision thresholds.
The implication is important: a maximum observed exposure near \(EC_{50}\) can show a substantial effect while leaving considerable room for further increase. A plateau is most plausible when response changes little across exposures already well above the transition. Conversely, if the transition location is unknown, the observed exposure range cannot by itself prove that the plateau was reached.
Why Low Coverage Can Hide the Maximum Effect
When \(C\) is much smaller than \(EC_{50}\), the Emax equation has the approximation \(R(C)-R_0 \approx (E_{\max}/EC_{50})C\). This follows because \(EC_{50}+C\) is then approximately \(EC_{50}\), so \(E_{\max}C/(EC_{50}+C)\approx E_{\max}C/EC_{50}\). In this region, the data primarily inform the ratio \(E_{\max}/EC_{50}\), the low-exposure slope. They may not separately determine the maximum effect and the exposure at which half of it occurs.
That limitation explains a common ambiguity. A modest observed change at low exposures could arise from a small \(E_{\max}\), or from a larger \(E_{\max}\) with a much higher \(EC_{50}\). Both can produce a similar initial slope. Calling the effect “weak” from those observations alone treats an unobserved asymptote as if it had been measured.
Worked Example: Cohorts That Approach a Plateau
Consider hypothetical cohort summaries from a study measuring a biomarker in units. Suppose the response change is modeled as \(R(C)-R_0=-24C/(3+C)\), where exposure \(C\) is in mg/L. The negative sign represents a decrease. Assume the exposure metric and sampling time are appropriate for this direct-effect model, and that baseline and time-course changes have been accounted for.
The hypothetical cohort mean exposures and observed mean changes are: 0.3 mg/L and −2.2 units; 3 mg/L and −12.0 units; 30 mg/L and −21.8 units; and 60 mg/L and −22.9 units. The model predicts, for example, at \(C=3\): \(−24(3)/(3+3)=−72/6=−12.0\) units. At \(C=30\), it predicts \(−720/33=−21.82\) units; at \(C=60\), it predicts \(−1440/63=−22.86\) units. The listed responses are these predictions rounded to one decimal place.
At 30 mg/L, the exposure is \(30/3=10\) times \(EC_{50}\), and the model’s fraction of maximum response is \(30/(3+30)=0.909\), or about 91%. At 60 mg/L, it is \(60/(3+60)=0.952\), or about 95%. The predicted additional decrease from 30 to 60 mg/L is only \(22.86-21.82=1.04\) units in magnitude, compared with the 9.82-unit increase in magnitude from 3 to 30 mg/L.
Here, the high-exposure cohorts show a flattening pattern while exposures are well above the fitted transition. Under the stated model, this is evidence consistent with approaching a plateau. It is not proof that every participant is saturated, and the conclusion depends on the model and the adequacy of the cohort summaries.
Worked Example: A Rising Response Without Evidence of a Plateau
Now consider hypothetical cohort mean exposures of 0.2, 0.5, and 1.0 mg/L, with biomarker changes of −0.47, −1.14, and −2.18 units. Suppose these values follow \(R(C)-R_0=-24C/(10+C)\). For the highest exposure, the calculation is \(−24(1)/(10+1)=−24/11=−2.18\) units. The response is still increasing in magnitude at the highest observed exposure.
At 1.0 mg/L, the exposure is only \(0.1EC_{50}\), and the fraction of maximum response is \(1/(10+1)=0.091\), or about 9.1%. The observations therefore cover only the low part of this model’s curve. They do not show whether the eventual maximum is large or small, and they do not establish that a plateau is absent at higher exposures.
The ambiguity can also be seen from the low-exposure slope. The model has \(E_{\max}/EC_{50}=-24/10=-2.4\) units per mg/L. A model with \(E_{\max}=-6\) units and \(EC_{50}=2.5\) mg/L has the same low-exposure slope, \(-6/2.5=-2.4\) units per mg/L. At the observed concentrations, the two curves may be difficult to distinguish. A small response over this restricted range is not enough to label the maximum effect weak.
The appropriate coverage conclusion is that the observed cohorts do not reach the transition region for the first model. To distinguish a small maximum effect from a higher \(EC_{50}\), one would need informative exposures nearer to, and ideally above, the plausible transition range, if those exposures are appropriate and feasible.
Worked Example: A Small Effect With Broad Coverage
Suppose a different hypothetical response follows \(R(C)-R_0=-3C/(2+C)\), with exposure in mg/L and response change in biomarker units. Cohort mean exposures are 0.2, 2, 20, and 60 mg/L, and observed changes are −0.27, −1.50, −2.73, and −2.90 units. For example, at 20 mg/L the model predicts \(−3(20)/(2+20)=−60/22=−2.73\) units.
Here \(EC_{50}=2\) mg/L. At 20 mg/L, the fraction of maximum response is \(20/(2+20)=0.909\), and at 60 mg/L it is \(60/(2+60)=0.968\). The predicted additional change in magnitude between those two high exposures is \(2.90-2.73=0.17\) units. The response is small even near the asymptote: the fitted maximum change is only −3 units.
If the observed range and the model are credible, this pattern differs from the preceding example: the high exposures cover the transition and approach saturation, yet the estimated maximum change remains small in biomarker units. That supports a small effect on this biomarker under the modeled conditions. It does not establish that the drug lacks other effects or that the biomarker is unimportant.
Assessing Cohort Coverage in Practice
A practical assessment begins with the individual exposure distributions, not only the cohort means. For each escalation cohort, inspect the range or relevant quantiles of exposure alongside the biomarker response at the intended time point. Note whether adjacent cohorts overlap, whether any cohort actually extends into a high-exposure region, and how much information is concentrated near the response transition. A high cohort mean can conceal participants whose exposures remain low; a wide within-cohort distribution can sometimes provide useful exposure variation, but it can also make a simple mean-response summary inadequate.
Specify the exposure metric, biomarker measure, analysis time, and direction of change. Confirm that these match the pharmacological question and the model assumptions.
Compare the distribution across cohorts with plausible \(EC_{50}\) values or an estimated transition. Ask whether observations include exposures below, around, and above that range.
Check whether exposures several times the plausible \(EC_{50}\) are observed and whether responses change little across those high exposures. A high dose without high measured exposure does not establish upper-range coverage.
Distinguish “the observed range does not locate the plateau” from “the response appears to approach a small asymptote.” Do not present either as certain if model or data limitations are substantial.
Dose-based summaries remain useful for describing the escalation design and operational decisions, but they cannot substitute for exposure-based assessment when exposure varies appreciably within or across cohorts. Conversely, exposure-response analysis does not eliminate the need to consider the design: a narrow exposure range, confounding between cohort and time, or changes in sampling conditions can make a fitted curve difficult to interpret.
Common Mistakes, Diagnostics, and Limitations
- Mistake: “The response is small at the top dose, so the drug has a weak maximum effect.” The correction is to check whether measured exposure reaches the response transition and upper region. A small change below the transition may reflect a high \(EC_{50}\), not a small \(E_{\max}\).
- Mistake: “The response is still increasing, so there is no plateau.” This overstates what the data show. If the maximum observed exposure is well below plausible \(EC_{50}\) values, the plateau is unobserved; it has not been ruled out.
- Mistake: “The highest dose is high enough to show saturation.” Nominal dose is not exposure. Check measured exposure distributions and the exposure-response scale before making that claim.
Warning signs of inadequate coverage include a maximum exposure below plausible transition values, no high-exposure observations, a response that continues to change at the range boundary, or markedly overlapping cohort exposure distributions that provide little extension of the range. A fitted \(E_{\max}\) that changes greatly when the model or exposure range changes is also a reason to qualify the interpretation, not a reason to claim a precise maximum effect.
Coverage cannot be judged independently of uncertainty, measurement quality, or the biological meaning of the response. Even exposures above a fitted \(EC_{50}\) may not support a reliable plateau conclusion if the biomarker is noisy, the high-exposure cohort is small, or the assumed direct-effect shape is inappropriate. And no finite observed range rules out every possible curve with a transition beyond that range. The conclusion should therefore be conditional: the data cover, or do not cover, the transition and upper region under the stated model and plausible parameter range.
The central distinction is straightforward but consequential: an unobserved plateau is a coverage limitation; a weak maximum effect is a statement about the asymptotic response. Cohorts can support the second interpretation only when their exposures and response information give a credible view of that asymptote.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Under an Emax model, what exposure corresponds to 80% of the maximum response in terms of \(EC_{50}\)?
- A cohort’s maximum exposure is \(0.2EC_{50}\). What fraction of the maximum response does the model predict there, and what can you conclude about plateau coverage?
- Why can low-exposure data fail to distinguish a small \(E_{\max}\) from a large \(EC_{50}\)?
- At exposures of \(10EC_{50}\) and \(20EC_{50}\), calculate the fractions of maximum response and explain what a nearly unchanged response across these levels might suggest.
Solutions. (1) From \(C=qEC_{50}/(1-q)\), \(C=0.8EC_{50}/0.2=4EC_{50}\). (2) The fraction is \(0.2/(1+0.2)=0.167\), about 17%. The data are far below the upper region and do not establish whether the maximum effect is small. (3) When \(C\) is much smaller than \(EC_{50}\), the response is approximately \((E_{\max}/EC_{50})C\), so the observed slope may be similar for different combinations of \(E_{\max}\) and \(EC_{50}\). (4) The fractions are \(10/11=0.909\) and \(20/21=0.952\). If the response changes little between these high exposures, that is consistent with approaching a plateau under the model, but the conclusion still depends on the data and model assumptions.