Why Assay Observation Models Matter
In a target-mediated drug disposition (TMDD) model, free target and drug-target complex are latent biological states: they are quantities in the model, not automatically the values reported by an assay. An assay detects molecules through particular epitopes, reagents, and calibration procedures. It may therefore respond to both free target and complex, recover only part of the complex, or be affected by drug remaining in the sample. Treating the reported value as the exact concentration of one state can make a mechanistic model appear to fit poorly or can distort its parameter estimates.
An observation model describes how the model-predicted states generate the data. Its coefficients are assay properties, not binding or turnover parameters. These properties should be based on assay validation or estimated only when the data provide enough information to do so. The model also needs an explicit measurement-error distribution. The equations below use \(R\) for free target concentration, \(C\) for drug-target complex concentration, and \(D\) for free drug concentration. For one target molecule per complex, the biological total target concentration is \(R+C\); an assay need not report that total exactly.
Specify the Assay Before Interpreting Its Result
A useful starting point is to express the expected assay result in concentration-equivalent units. For a free-target assay, one possible model is
Here \(b_F\) is a background offset, \(\rho_F\) is the response to complex relative to free target, and \(q_F(D)\) is the recovery or response multiplier in the presence of drug. An ideal free-target assay has \(b_F=0\), \(\rho_F=0\), and \(q_F(D)=1\). In practice, these properties should be assessed over the relevant concentration ranges; a single constant coefficient may not describe an assay whose cross-reactivity changes with concentration.
A complex assay can be written as \(\mu_C=b_C+q_C(D)\{C+\rho_C R\}\), where \(\rho_C\) represents response to free target. A total-target assay may be represented as
The coefficients \(q_{T,R}\) and \(q_{T,C}\) are separate because an assay can recover free and complexed target differently. Only when both coefficients equal one and \(b_T=0\) does its expected reading equal the biological total \(R+C\). Assay results can instead be reported as raw signal; in that case, a calibration function maps the concentration-equivalent expressions to signal units.
For an additive Gaussian error model, an observation \(Y_j\) from assay \(j\) has conditional distribution \(Y_j\mid R,C,D\sim N(\mu_j,\sigma_j^2)\), where \(\mu_j\) is the assay-specific expected result and \(\sigma_j\) is its residual standard deviation. This is appropriate when errors are approximately symmetric and additive in the reported units. For positive measurements with variability that grows proportionally with concentration, a lognormal model is often more appropriate. If \(\log Y_j\mid R,C,D\sim N(\log\mu_j,\sigma_j^2)\), then \(\mu_j\) is the conditional median, not the arithmetic mean; the latter is \(\mu_j\exp(\sigma_j^2/2)\).
Worked Example: A Selective Assay With Small Cross-Reactivity
Suppose a hypothetical sample has latent concentrations \(R=4.00\) nM and \(C=8.00\) nM. A calibrated free-target assay has no offset, no drug interference, and \(\rho_F=0.050\). Assume additive Gaussian error with standard deviation \(0.30\) nM.
The expected assay-equivalent concentration is
If the observed result is \(Y_F=4.50\) nM, its residual is \(4.50-4.40=0.10\) nM, or \(0.10/0.30=0.333\) residual standard deviations. This is a modest discrepancy under the assumed error model. The assay result is not itself the free-target concentration: even at the expected value, the complex contributes \(0.40\) nM of assay-equivalent response.
When Cross-Reactive Assays Can Identify Both States
Two assays can, in principle, provide information about two latent states if their responses to those states are sufficiently different. Consider calibrated observations with no offsets or drug interference:
The coefficients \(a\) and \(b\) are cross-reactivity coefficients. For the moment, suppose the errors have mean zero. The expected observations are a linear system. Its coefficient matrix has determinant \(1-ab\); when \(1-ab\ne0\), the system has the algebraic solution
This inversion is unbiased under the stated linear model and zero-mean error, but it does not make the resulting estimates exact. It can amplify measurement noise, and it can produce a negative estimate even though the biological concentrations cannot be negative.
Worked Example: Correcting Two Cross-Reactive Measurements
Worked Example: Correcting Two Cross-Reactive Measurements
In a hypothetical sample, two calibrated assays report \(Y_1=5.00\) nM and \(Y_2=9.00\) nM. Their response model is \(Y_1=R+0.10C+\varepsilon_1\) and \(Y_2=0.20R+C+\varepsilon_2\). Assume zero offsets, no drug interference, and independent zero-mean errors with standard deviations \(0.20\) nM and \(0.30\) nM, respectively.
The determinant is \(1-(0.10)(0.20)=0.98\), not close to zero, so the algebraic correction is defined. Substitution gives
The estimated biological total is therefore \(\hat R+\hat C=4.1837+8.1633=12.3469\) nM, or \(12.35\) nM to four significant figures. This is an estimate, not a third independent observation.
The error variances also matter. Since \(\hat C=(-0.20Y_1+Y_2)/0.98\), independence gives
Similarly, \(\mathrm{Var}(\hat R)=\{(0.20)^2+(0.10)^2(0.30)^2\}/(0.98)^2=0.04259\ \mathrm{nM}^2\), so \(SE(\hat R)=0.2064\) nM. The two corrected estimates are correlated because each uses both assay results. In particular, their covariance is \(\{(1)(-0.20)(0.20)^2+(-0.10)(1)(0.30)^2\}/(0.98)^2=-0.01770\ \mathrm{nM}^2\). Ignoring this covariance would give the wrong uncertainty for their sum.
Representing Drug Interference
Drug interference is an assay effect, not automatically a change in the biological target state. Residual drug in a sample may mask an epitope, alter capture or detection, or change complex recovery during sample processing. The direction and size of the effect are assay-specific. They should be characterized with validation experiments that vary drug concentration and the relevant target forms.
One possible empirical recovery function is \(q_F(D)=1/(1+D/K_I)\), where \(K_I\) is a fitted or validated assay-interference scale. This is a convenient model form, not a universal law of assay chemistry. It assumes that the drug concentration affecting the assay is known and that the specified function adequately describes recovery over the range used. If interference depends on sample dilution, incubation, or dissociation of complex, those conditions may require additional terms or a different observation model.
Worked Example: Adjusting a Drug-Interfered Free-Target Assay
Suppose a hypothetical sample has \(R=6.00\) nM and \(C=3.00\) nM. A free-target assay responds to complex with \(\rho_F=0.050\), and validation supports \(q_F(D)=1/(1+D/K_I)\) with \(D=30.0\) nM and \(K_I=30.0\) nM. Assume zero offset and additive Gaussian measurement error with standard deviation \(0.15\) nM.
The recovery multiplier is \(q_F=1/(1+30.0/30.0)=0.500\). Thus the expected result is
For a hypothetical observed result of \(3.10\) nM, the residual is \(3.10-3.075=0.025\) nM. If \(D\), \(K_I\), and \(C\) are treated as known for this simple correction, the corresponding free-target estimate is
The measurement-error contribution to the standard error of this corrected estimate is \(0.15/0.500=0.300\) nM, before accounting for uncertainty in \(D\), \(K_I\), or \(C\). If drug interference were ignored, the same observation would imply \(R=3.10-0.050(3.00)=2.95\) nM under the no-interference formula. That large difference illustrates why assay recovery must be represented rather than absorbed into a biological interpretation.
Fit the Observation Model and Diagnose Its Weaknesses
When assay results are collected over time or across subjects, the preferred analysis is usually to connect the observation model directly to the TMDD model’s predicted states at each sampling time. The resulting likelihood uses the assay-specific expected values and error distributions. This approach can account for uncertainty in the latent trajectory and can estimate parameters jointly when the data support them. Direct inversion is useful for explanation or exploratory summaries, but it treats calibrated coefficients as fixed and can be unstable when the assays respond similarly to the two states.
For the two-assay system above, the denominator \(1-ab\) reveals the core identifiability problem. If \(ab\) approaches one, small changes in either measured value can produce large changes in the corrected concentrations. If \(ab=1\), the coefficient matrix is singular: these two assays do not uniquely distinguish \(R\) from \(C\) under this model. A numerically available inverse is not, by itself, evidence that the states are estimated precisely.
Diagnostics should be performed on the observation model as well as on the state model. Examine residuals by assay, concentration, time, drug concentration, and sample-processing condition. Trends with drug concentration may indicate unmodeled interference; curvature may indicate concentration-dependent cross-reactivity; and increasing residual spread may call for a proportional or combined error model rather than constant additive error. Compare predictions with dilution, spike-recovery, and specificity validation results where available. A good fit to concentration-time data does not establish that an assay’s selectivity assumptions are correct.
A total-target assay can provide a useful additional constraint, especially when it has different and validated recoveries for free target and complex. An orthogonal assay, such as one using a different epitope or detection principle, may be more informative than simply adding another highly correlated assay. Conversely, forcing an idealized total-target relation \(Y_T=R+C\) onto an assay with poor complex recovery can bias inferred state concentrations. Assay validation and model structure must agree before that measurement is treated as a direct total.
Common Mistake and Practical Limits
Wrong reasoning: “The kit is called a free-target assay, so its reported concentration is \(R\); any discrepancy from the model must be biological.” Correction: The label does not establish zero cross-reactivity, complete recovery, or absence of drug interference. Specify the response to each relevant analyte and the error distribution, then compare the assay prediction with the observation.
Linear selectivity models are approximations. They may fail if binding or dissociation occurs during sample preparation, if response saturates, if the assay changes the equilibrium between free and bound forms, or if interference depends on the particular drug concentration and processing protocol. Calibration coefficients can also be uncertain. Treating them as known can understate uncertainty; estimating many assay coefficients from the same clinical data can instead create weak identifiability with the biological parameters. Where possible, use external validation to fix or constrain assay properties, and report limitations when they cannot be separated from biological effects.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A total-target assay has expected response \(\mu_T=R+0.70C\). When does its expected value equal the biological total target concentration?
- For \(Y_1=R+0.20C\) and \(Y_2=0.30R+C\), what is the determinant of the selectivity matrix, and can the two states be algebraically separated?
- A free-target assay has \(q_F(D)=0.80\), \(\rho_F=0.10\), \(C=2.00\) nM, and observation \(Y_F=4.00\) nM. Under the stated model, calculate the corrected estimate of \(R\).
- Why can direct inversion of two assays yield a negative estimate of a target concentration?
- What residual pattern would suggest that a constant assay-recovery model is inadequate?
Solutions. (1) It equals \(R+C\) only if \(0.70=1\), which is false for nonzero \(C\); in this model it equals the biological total only when \(C=0\). (2) The determinant is \(1-(0.20)(0.30)=0.94\), which is nonzero, so the equations can be inverted algebraically; precision still depends on measurement error and conditioning. (3) \(R=Y_F/q_F-\rho_F C=4.00/0.80-0.10(2.00)=4.80\) nM. (4) Measurement noise can move the observations outside the nonnegative combinations predicted by the noiseless model, and inversion maps those noisy values to negative estimates. (5) Residuals that systematically increase, decrease, or curve with drug concentration suggest that recovery changes with drug concentration and that a constant-recovery model may be insufficient.