What a Kp Discrepancy Says—and Does Not Say
A measured tissue-to-plasma partition coefficient, \(K_p\), that differs from a model prediction is a diagnostic clue, not an explanation. The size and direction of the difference can help prioritize checks, but a residual alone cannot identify whether the cause is a unit conversion, a tissue-property input, a binding assumption, pH, or a process absent from the model. A useful diagnosis therefore begins by making sure the measured and predicted quantities are genuinely comparable.
For this tutorial, \(K_p\) means the ratio of total drug concentration in a specified tissue to total drug concentration in plasma, with both concentrations expressed on compatible bases. A comparison is interpretable only when the drug, species, tissue definition, concentration basis, and relevant experimental or prediction conditions match. If a discrepancy remains after those checks, examine tissue-specific inputs and assumptions systematically.
The log-ratio is helpful because it treats reciprocal discrepancies symmetrically: a measurement twice the prediction and one half the prediction have residuals of equal magnitude and opposite sign. It also expresses the mismatch as a multiplicative factor, which is often more informative than a raw difference when \(K_p\) values vary widely across tissues.
A Short Derivation: Reading the Residual Scale
Let \(F=K_{p,\mathrm{meas}}/K_{p,\mathrm{pred}}\) be the measured-to-predicted fold ratio. By the definition of \(r\), \(r=\log_{10}(F)\), so exponentiating both sides gives \(F=10^r\). If the measurement is twice the prediction, \(F=2\) and \(r=\log_{10}(2)\approx0.3010\). If it is half the prediction, \(F=0.5\) and \(r=\log_{10}(0.5)\approx-0.3010\). The equal and opposite residuals follow from \(\log_{10}(1/F)=-\log_{10}(F)\).
This residual is a descriptive diagnostic, not a statistical test. It does not by itself include uncertainty in either the measured value or model inputs, and it does not establish that a mechanistic cause is present. It is most useful when calculated consistently across tissues and paired with a clear record of the inputs used.
Worked Example: Correcting a Concentration-Basis Mismatch
Suppose a hypothetical experiment reports total drug concentration in liver as \(10.0\) nmol/g of wet tissue. Plasma concentration is \(2.00\) nmol/mL. A prediction is \(K_p=5.10\), defined using tissue and plasma concentrations on a compatible volume basis. Assume, for this illustration, that liver density is \(1.05\) g/mL and that the tissue concentration is representative of the wet tissue.
Convert before forming the ratio. There are \(1.05\) g of liver per mL, so the tissue concentration on a volume basis is $$ 10.0\ \mathrm{nmol/g}\times1.05\ \mathrm{g/mL}=10.5\ \mathrm{nmol/mL}. $$ Thus the compatible measured partition coefficient is \(10.5/2.00=5.25\). The residual is $$ r=\log_{10}\left(\frac{5.25}{5.10}\right)=\log_{10}(1.02941)=0.01259. $$ The measured value is about \(1.029\)-fold the prediction, a small positive discrepancy on this scale.
If the units were divided without conversion, the numerical ratio would be \(10.0/2.00=5.00\), and the residual would be \(\log_{10}(5.00/5.10)=-0.008600\). That apparent negative residual is not evidence of a biological mismatch: it results from treating nmol/g and nmol/mL as if they were the same concentration basis.
A Practical Diagnostic Sequence
Before interpreting a residual mechanistically, apply the checks in a deliberate order. These checks are especially important when a residual appears large or differs in sign among organs. The sequence below is a way to organize the investigation, not a claim that every discrepancy has one uniquely identifiable cause.
Confirm the same tissue, species, drug, total or unbound concentration definition, and plasma reference. Check units and whether tissue concentration is reported per mass, per volume, or on another basis.
Review whether the prediction used tissue-specific composition appropriate to the species and tissue represented by the measurement. Composition inputs affect the predicted partitioning and can create organ-specific residuals.
Check the binding quantities and ionization or pH inputs actually used. Determine whether changing a plausible input moves the prediction in the direction and approximate magnitude needed, rather than changing inputs indiscriminately.
If a discrepancy persists after input and basis checks, ask whether a tissue-specific transport or other process absent from the prediction could contribute. Treat this as a hypothesis to evaluate, not as a conclusion from the residual alone.
A broad shift in residuals across many tissues may point first to a shared convention or input, such as the plasma reference or a common property used by the prediction. A discrepancy concentrated in one tissue makes a tissue-specific composition, binding, pH, or transport issue more plausible. These patterns prioritize checks; they do not prove the cause, because multiple errors or mechanisms can produce similar residual profiles.
Worked Example: A Tissue-Composition Input Changes the Residual
Consider a hypothetical drug for which a composition-based prediction gives \(K_{p,\mathrm{pred}}=4.00\) in one tissue, while the compatible measured value is \(5.10\). The initial residual is $$ r=\log_{10}\left(\frac{5.10}{4.00}\right)=\log_{10}(1.275)=0.1055. $$ The measurement is \(1.275\)-fold the prediction.
An audit then finds that the prediction used a tissue-composition profile not appropriate for the species in the measurement. With the appropriate profile supplied to the same model, the prediction becomes \(5.20\). The revised residual is $$ r=\log_{10}\left(\frac{5.10}{5.20}\right)=\log_{10}(0.980769)=-0.008439. $$ The composition correction substantially reduces the discrepancy, but this numerical improvement alone does not prove that composition was the only source of error. The corrected input must be independently justified, and other assumptions still need to be checked.
Binding and pH: Check the Direction as Well as the Size
Binding assumptions matter because \(K_p\) is a total-concentration ratio, whereas many mechanistic relationships are expressed through unbound concentrations. Under a specified equilibrium framework, let \(K_{p,uu}=C_{u,\mathrm{tissue}}/C_{u,\mathrm{plasma}}\), where \(C_u\) denotes unbound concentration, and let \(f_{u,\mathrm{tissue}}\) and \(f_{u,\mathrm{plasma}}\) be the unbound fractions in tissue and plasma. Since total concentration equals unbound concentration divided by the unbound fraction, the corresponding relationship is $$ K_p=K_{p,uu}\frac{f_{u,\mathrm{plasma}}}{f_{u,\mathrm{tissue}}}. $$ This relation applies only when the concentrations and unbound fractions refer to compatible conditions and the unbound ratio is defined on the same tissue and plasma basis. It shows why an assumed tissue binding term can shift predicted total \(K_p\), even if the unbound tissue-to-plasma ratio is unchanged. It does not justify changing a binding input unless the revised value is supported.
pH is another tissue-specific input to check for ionizable drugs. If the model accounts for a tissue-to-plasma pH gradient, verify that the pH values and the drug’s ionization assumptions correspond to the intended conditions. A change in predicted \(K_p\) after updating pH is evidence that the prediction is sensitive to that input; it is not independent confirmation that pH caused the original measured value.
Worked Example: Testing a Binding Assumption
Suppose a hypothetical model predicts \(K_p=8.00\), while a compatible measurement is \(12.0\). The residual is $$ r=\log_{10}\left(\frac{12.0}{8.00}\right)=\log_{10}(1.50)=0.1761. $$ The measurement is \(1.50\)-fold the prediction.
For a simple diagnostic calculation, suppose \(K_{p,uu}\) and \(f_{u,\mathrm{plasma}}\) are held fixed, and the model’s tissue unbound fraction was \(0.0300\). If a supported alternative estimate is \(0.0200\), the relationship \(K_p=K_{p,uu}f_{u,\mathrm{plasma}}/f_{u,\mathrm{tissue}}\) implies a \(0.0300/0.0200=1.50\)-fold increase in predicted total \(K_p\). The revised prediction would therefore be \(8.00\times1.50=12.0\), giving \(r=\log_{10}(12.0/12.0)=0\).
This arithmetic demonstrates how the assumed binding quantity could account for the size of the discrepancy under the stated relationship. It does not establish that the alternative fraction is correct, or rule out other explanations. The binding input requires evidence independent of the fact that it makes the prediction agree.
When an Unmodeled Process Is Plausible
A persistent residual may reflect a limitation of the model structure rather than an incorrect value for an existing input. For example, tissue-specific uptake or efflux transport may influence distribution in ways not represented in a passive or composition-based prediction. A mismatch that is localized to a tissue where such transport is biologically plausible can motivate a targeted investigation. However, the same pattern can also result from an incorrect tissue composition, binding estimate, pH assumption, or measurement basis.
A responsible diagnosis therefore separates three statements: the residual is observed; a proposed input or process could explain its direction and magnitude; and independent evidence supports that explanation. Only the first follows from calculating \(r\). Sensitivity analysis can help determine whether changing an input has enough influence to matter, but a sensitive input is not automatically the true cause.
Comparing Ways to Summarize the Mismatch
| Summary | Useful feature | Limitation |
|---|---|---|
| Raw difference, \(K_{p,\mathrm{meas}}-K_{p,\mathrm{pred}}\) | Retains the original \(K_p\) units and may be useful when absolute errors matter. | Its size depends strongly on the scale of the \(K_p\) values and is not symmetric for reciprocal fold errors. |
| Fold ratio, \(K_{p,\mathrm{meas}}/K_{p,\mathrm{pred}}\) | States directly how many times larger or smaller the measurement is. | Over- and underprediction are not centered around zero, and reciprocal ratios are harder to compare by subtraction. |
| Log-ratio residual, \(r\) | Centers agreement at zero and gives equal magnitudes for reciprocal fold errors. | Requires positive values and does not represent uncertainty or identify a mechanism by itself. |
| Residual profile across tissues | Can reveal shared shifts or tissue-localized discrepancies that guide input checks. | Patterns prioritize hypotheses but do not uniquely distinguish their causes. |
These summaries answer different questions. The raw difference is appropriate when an absolute concentration-ratio error is the target; the fold ratio is intuitive for multiplicative error; and the log-ratio is convenient for signed comparisons across tissues. None should be interpreted without checking that the compared values use a common definition and basis.
Common Pitfalls and Limits
- Comparing unlike concentration bases. A ratio formed from tissue concentration per gram and plasma concentration per milliliter is not automatically a valid \(K_p\). Convert or otherwise reconcile the bases first.
- Treating agreement as validation. A small residual does not show that the model’s mechanism is correct; compensating errors can also produce agreement.
- Changing inputs until the residual disappears. An adjustment is informative only if the revised input is independently plausible and the model’s assumptions remain appropriate.
- Assuming a tissue-specific residual proves transport. Transport is one possible explanation, not a diagnosis. Composition, binding, pH, and comparability checks remain necessary.
- Ignoring the conditions of measurement. The interpretation assumes that measured and predicted values refer to compatible conditions. If that comparability is in doubt, resolve it before assigning a mechanistic explanation.
The main limitation is identifiability: distinct combinations of input errors and missing processes can yield similar \(K_p\) residuals. A residual is most useful as part of an evidence chain that records the comparison basis, documents each input check, and states clearly which explanations remain uncertain.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A measured \(K_p\) is \(6.00\) and the prediction is \(3.00\). What is the log-ratio residual, and how should it be interpreted?
- A measured tissue concentration is reported per gram, but the plasma concentration is per milliliter. What must be checked before forming \(K_p\)?
- A tissue-specific residual becomes small after changing a composition input. Does that prove composition was the cause? Explain.
- Under the stated relationship \(K_p=K_{p,uu}f_{u,\mathrm{plasma}}/f_{u,\mathrm{tissue}}\), what happens to predicted \(K_p\) if \(f_{u,\mathrm{tissue}}\) is halved while the other terms are fixed?
- A large residual remains in one organ after basis, composition, binding, and pH checks. What can an unmodeled transport hypothesis explain, and what does the residual alone not establish?
Solutions. (1) \(r=\log_{10}(6.00/3.00)=\log_{10}(2)=0.3010\); the measurement is twice the prediction. (2) Confirm compatible concentration units and tissue basis, using an appropriate conversion or basis reconciliation before taking the ratio. (3) No. The revised input needs independent support, and other errors or compensating effects may remain. (4) Predicted \(K_p\) doubles because it is inversely proportional to \(f_{u,\mathrm{tissue}}\) when the other terms are fixed. (5) Transport may be a plausible missing tissue-specific process, but the residual alone cannot prove transport is present or exclude remaining measurement or model limitations.