Uncertainty in Translating Potency Across Species

Why Potency Translation Needs More Than a Ratio

A potency value is not a species-independent property that can be copied directly from one assay into a human prediction. A measured value may reflect the assay’s temperature, buffer, substrate concentration, and protein content, as well as the compound’s binding to proteins in that assay. Moreover, equal receptor affinity does not guarantee equal functional response: receptor abundance, signaling, and other species-specific features can change the concentration needed to produce an effect.

The useful question is therefore specific: what free concentration in each species corresponds to a defined pharmacological effect, under comparable conditions, and how uncertain is the comparison? The calculations below address potency translation at the concentration or target-sensitivity level. They do not convert potency into a dose, exposure profile, or clinical outcome.

Definition: Receptor affinity describes binding between a compound and a target, often summarized by a dissociation constant \(K_d\); lower \(K_d\) indicates tighter binding under the stated conditions. Functional potency describes the concentration associated with a specified response in a functional assay, commonly an \(EC_{50}\). Free potency is a potency expressed as unbound concentration, after accounting for the unbound fraction in the assay. These quantities are related but are not interchangeable.

Put Potency on a Comparable Concentration Basis

Suppose a functional assay reports \(EC_{50,\mathrm{total},s}\) for species \(s\), using total compound concentration in the assay well. If the unbound fraction in that assay is \(f_{u,s}\), then the unbound concentration at the reported \(EC_{50}\) is \(f_{u,s}EC_{50,\mathrm{total},s}\). This adjustment assumes that the measured fraction applies at the relevant concentration and that free concentration is the relevant driver of the response.

Assays may also differ in conditions that affect the measured potency. Let \(q_s\) be a justified correction factor that converts the measured result to a specified common condition. It must be estimated from appropriate calibration data or a validated mechanistic relationship; it is not a generic correction for “species differences.” Define the standardized free potency as:

P_s = EC_{50,\mathrm{total},s}\, f_{u,s}\, q_s .

Here, \(P_s\) has concentration units. If conditions are already matched, use \(q_s=1\). If no defensible condition adjustment is available, report the condition mismatch and avoid presenting an arbitrary corrected value as if it were measured. When the assay reports an unbound \(EC_{50}\) directly, do not apply the assay unbound-fraction correction a second time.

For a simple equilibrium binding model, receptor occupancy is \(\theta=C_u/(K_d+C_u)\), where \(C_u\) is free concentration. Solving for concentration gives \(C_u=K_d\theta/(1-\theta)\). Thus, for the same occupancy in two species, the free concentrations are proportional to their \(K_d\) values, provided equilibrium and the same occupancy definition apply. This is a binding comparison, not proof that the same occupancy produces the same functional response.

Worked Example: Correcting Functional Potency for Assay Binding

In a hypothetical matched functional assay, a rat preparation has a total-concentration \(EC_{50}\) of 120 nM and assay unbound fraction \(f_{u,\mathrm{rat}}=0.25\). The human preparation has an \(EC_{50}\) of 45 nM and \(f_{u,\mathrm{human}}=0.60\). Assume the assay conditions are matched, so both condition factors are 1. Calculate the free-potency ratio, human divided by rat.

For rat, \(P_{\mathrm{rat}}=120\ \mathrm{nM}\times0.25\times1=30.0\ \mathrm{nM}\). For human, \(P_{\mathrm{human}}=45\ \mathrm{nM}\times0.60\times1=27.0\ \mathrm{nM}\). Therefore:

R=\frac{P_{\mathrm{human}}}{P_{\mathrm{rat}}} =\frac{27.0\ \mathrm{nM}}{30.0\ \mathrm{nM}}=0.900 .

On the total-concentration results alone, the human \(EC_{50}\) is \(45/120=0.375\) of the rat value. After assay binding is considered, the ratio is 0.900: the human free \(EC_{50}\) is estimated to be 90.0% of the rat free \(EC_{50}\). The direction and magnitude of a comparison can change substantially after normalization. The arithmetic assumes the unbound fractions are applicable to these assays and that the two functional readouts are sufficiently comparable.

Quantifying Uncertainty in the Cross-Species Ratio

A ratio is often easiest to analyze on the log scale, where multiplication becomes addition and reciprocal comparisons change sign. Let \(R=P_{\mathrm{human}}/P_{\mathrm{animal}}\). From the definition of standardized free potency:

\log R = \log\!\left(\frac{EC_{50,\mathrm{total,human}}} {EC_{50,\mathrm{total,animal}}}\right) +\log\!\left(\frac{f_{u,\mathrm{human}}}{f_{u,\mathrm{animal}}}\right) +\log\!\left(\frac{q_{\mathrm{human}}}{q_{\mathrm{animal}}}\right).

Write the three terms as \(A\), \(B\), and \(C\), respectively. The variance of their sum is:

\operatorname{Var}(\log R)= \operatorname{Var}(A)+\operatorname{Var}(B)+\operatorname{Var}(C) +2\operatorname{Cov}(A,B)+2\operatorname{Cov}(A,C)+2\operatorname{Cov}(B,C).

This identity follows by expanding \(\operatorname{Var}(A+B+C)=E[((A-E[A])+(B-E[B])+(C-E[C]))^2]\): the squared terms give the three variances, and each cross-product contributes twice a covariance. If the component estimates are independent, their covariance terms are zero and their variances can be added. Pairwise zero covariance is also sufficient for adding variances; independence is not necessary. If estimates share samples, calibration data, or a common assay, covariance may be important.

When the log ratio is approximately normally distributed with estimated standard error \(SE_{\log R}\), an approximate 95% uncertainty interval is obtained by exponentiating the log-scale interval. It describes uncertainty conditional on the specified inputs, covariance assumptions, and log-normal approximation. It is not a probability statement that a fixed parameter lies in the realized interval.

Worked Example: Propagating Uncertainty With Covariance

Continue with the hypothetical point estimate \(R=0.900\). Suppose the standard errors of the log total-potency ratio, log unbound-fraction ratio, and log condition-correction ratio are 0.180, 0.200, and 0.100, respectively. Suppose the covariance between the first two terms is \(-0.0120\), and the other two covariances are zero. These are illustrative uncertainty inputs, not measurements from a real study.

The variance is:

\operatorname{Var}(\log R) =(0.180)^2+(0.200)^2+(0.100)^2+2(-0.0120) =0.0324+0.0400+0.0100-0.0240 =0.0584 .

Thus \(SE_{\log R}=\sqrt{0.0584}=0.2417\). Since \(\log(0.900)=-0.1054\), the approximate 95% log-scale interval is:

-0.1054\pm1.96(0.2417) =[-0.5791,\ 0.3683].

Exponentiating both endpoints gives \([\exp(-0.5791),\exp(0.3683)]=[0.560,\ 1.445]\). Under the stated approximation, the interval for the human-to-animal free-potency ratio includes 1, so these inputs do not establish a clear potency difference at this level of uncertainty. The negative covariance reduces the variance relative to the sum of the three variances alone. Independence would be one sufficient reason for zero covariance, but is not required; assuming zero covariance should be justified, not inferred merely because dependence is unknown.

Represent Species-Specific Functional Sensitivity

Free \(EC_{50}\) is a useful summary, but it does not alone determine the concentration needed for every target response. Consider a Hill model for a baseline-subtracted response \(E_s\) in species \(s\):

E_s(C_u)=\frac{E_{\max,s} C_u^{\gamma_s}} {EC_{50,s}^{\gamma_s}+C_u^{\gamma_s}},

where \(E_{\max,s}\) is the maximum response on a common response scale, \(\gamma_s>0\) is the Hill coefficient, and \(EC_{50,s}\) is expressed as free concentration. This model assumes a monotone response over the relevant range and a suitable Hill-shaped concentration-response relationship. Solving for the free concentration at response \(E\), where \(0<E<E_{\max,s}\), gives:

C_{u,s}(E)=EC_{50,s} \left(\frac{E}{E_{\max,s}-E}\right)^{1/\gamma_s}.

The derivation is direct: divide the model by \(E_{\max,s}\), rearrange to obtain \(E/(E_{\max,s}-E)=(C_u/EC_{50,s})^{\gamma_s}\), and take both sides to the power \(1/\gamma_s\). This equation makes species-specific sensitivity explicit through \(EC_{50,s}\), \(E_{\max,s}\), and \(\gamma_s\). Comparing \(EC_{50}\) values is sufficient for comparing concentrations at the half-maximal response within each species’ own maximum, but not necessarily for matching the same absolute response when maxima or slopes differ.

Worked Example: Matching a Response With Different Species Sensitivity

Suppose hypothetical free-concentration response models have these parameters: animal \(EC_{50}=30.0\) nM, \(E_{\max}=80.0\) response units, and \(\gamma=1.20\); human \(EC_{50}=20.0\) nM, \(E_{\max}=90.0\) response units, and \(\gamma=1.00\). Assume response units are comparable between preparations and the models adequately describe the relevant range. Find the free concentration for a response of 60.0 units in each.

For the animal preparation:

C_{u,\mathrm{animal}}(60) =30.0\left(\frac{60.0}{80.0-60.0}\right)^{1/1.20} =30.0(3)^{0.8333} =74.9\ \mathrm{nM}.

For the human preparation:

C_{u,\mathrm{human}}(60) =20.0\left(\frac{60.0}{90.0-60.0}\right)^{1/1.00} =20.0(2) =40.0\ \mathrm{nM}.

Although the maxima and slopes differ, the requested response is below both maxima, so both calculations are defined. Under these hypothetical models, 40.0 nM free concentration corresponds to 60.0 response units in the human preparation, compared with 74.9 nM in the animal preparation. This is a functional-sensitivity comparison, not a claim that a particular clinical dose will achieve either concentration. If response scales are not comparable or the fitted models do not cover the target response, the numerical comparison is not justified.

Diagnostics, Alternatives, and Limitations

Before relying on a translated potency, check whether each input supports the interpretation assigned to it. Relevant diagnostics include:

Assay comparability: verify species identity, assay format, temperature, pH, incubation time, substrate or agonist conditions, and response definition. A correction factor is credible only when supported by calibration or a validated relationship.
Binding correction: check that unbound fractions refer to the assay matrix and relevant concentration range. Fractions near zero or one can be difficult to estimate precisely, and a small absolute error may produce a substantial ratio error.
Concentration-response adequacy: inspect whether the data cover both the response region and the plateau needed to estimate \(EC_{50}\), \(E_{\max}\), and \(\gamma\). If the maximum is not observed, sensitivity calculations at high response can be unstable.
Uncertainty structure: identify shared assay batches, shared calibration factors, and paired measurements. These can induce covariance; omitting it can either inflate or understate uncertainty depending on its sign.
Biological interpretation: distinguish a species difference in binding from a difference in functional response. A receptor-affinity ratio alone cannot account for receptor reserve, signaling, or other causes of different functional sensitivity.

The main alternatives answer different questions. Comparing raw total-concentration \(EC_{50}\) values is simplest, but can confound potency with assay binding and conditions. Comparing \(K_d\) values focuses on binding affinity and is appropriate when binding is the intended endpoint, but does not by itself predict functional sensitivity. Comparing standardized free \(EC_{50}\) values removes specified assay-level differences, while matching a response with a species-specific concentration-response model additionally incorporates maximum response and slope. None of these summaries replaces an in vivo model when the intended prediction concerns time-varying exposure or clinical effect.

A common mistake is to say, “The human total \(EC_{50}\) is lower, so the compound is proportionally more potent in humans.” That conclusion ignores the assay unbound fraction and potentially unmatched conditions. The correction is to state the measured endpoint, normalize only with defensible inputs, quantify uncertainty in the ratio, and specify whether the comparison concerns binding, free functional potency, or concentration for a matched response.

Key takeaway: A defensible cross-species potency comparison starts with a defined endpoint and comparable assay conditions, expresses potency on a free-concentration basis when appropriate, and carries uncertainty and covariance through the ratio. Species-specific functional sensitivity may require comparing concentrations at a specified response rather than comparing \(EC_{50}\) values alone.

Check Your Understanding

Answer each question, then compare with the solution provided.

A hypothetical animal assay reports total \(EC_{50}=80\) nM with \(f_u=0.50\); a human assay reports total \(EC_{50}=60\) nM with \(f_u=0.40\). Assume matched conditions. What is the human-to-animal free-potency ratio?
Solution: Animal free \(EC_{50}=80(0.50)=40\) nM; human free \(EC_{50}=60(0.40)=24\) nM. The ratio is \(24/40=0.600\).
What is the correct variance formula for the sum of two uncertain log-potency components \(A\) and \(B\)?
Solution: \(\operatorname{Var}(A+B)=\operatorname{Var}(A)+\operatorname{Var}(B)+2\operatorname{Cov}(A,B)\). Independence implies zero covariance, but zero covariance is sufficient for adding the variances and does not require independence.
In a Hill model, why can the same \(EC_{50}\) ratio fail to describe the concentration ratio for a matched absolute response?
Solution: The concentration at response \(E\) also depends on \(E_{\max}\) and \(\gamma\): \(C_u(E)=EC_{50}[E/(E_{\max}-E)]^{1/\gamma}\). Different maxima or slopes change the concentration needed for that response.
A binding assay has a different temperature from the functional assay, but no calibration data are available. Should a numerical condition correction be applied?
Solution: No. Report the mismatch and its potential relevance, or obtain a justified calibration. An unsupported correction factor gives a precise-looking value without a defensible basis.