Why Binding-Model Residuals Matter
A binding model summarizes how an observed signal depends on ligand concentration, time, and binding parameters. A good-looking curve is not enough to establish that the model describes the process adequately: different mechanisms can produce similar curves over a limited range. Residuals make discrepancies visible by comparing each observation with the model prediction at that observation’s conditions.
In this tutorial, a residual is defined as observed response minus predicted response. Its sign therefore has a direct interpretation: a positive residual means the model underpredicts the observation, and a negative residual means the model overpredicts it. The pattern of residuals across concentration or time can suggest which assumption deserves investigation, but it rarely identifies a mechanism by itself.
The basic one-site equilibrium model provides a useful reference. If \(C\) is the free ligand concentration, \(K_D\) is the equilibrium dissociation constant, and \(\theta\) is fractional occupancy, then
This expression assumes a single class of independent, reversible binding sites; equilibrium at the time of measurement; and a concentration that represents free ligand at the binding site. A measured signal may also include background or scale parameters, for example \(y=b+B_{\max}\theta\). The residuals must be examined on the scale of the measured response, with the assay’s uncertainty in mind.
Patterns Across Concentration
A concentration-response residual plot can reveal that the chosen curve shape is inadequate. Under the one-site model, occupancy rises hyperbolically toward a plateau. If the actual response rises more steeply around its midpoint, the one-site model can overpredict at lower concentrations and underpredict at higher concentrations. Conversely, a curve that rises more gradually may produce the opposite pattern. These are clues to a shape mismatch, not a unique diagnosis: cooperative binding, heterogeneous populations, concentration error, or other departures from assumptions can also produce structured residuals. The earlier material on cooperative binding and heterogeneous target populations treats those mechanisms.
A separate issue is the concentration used in the model. The occupancy equation uses free concentration, not necessarily the nominal or total concentration added to an assay. If ligand depletion is appreciable, substituting total concentration for free concentration can systematically distort predictions. Residuals may help flag this problem, but measurements of free ligand or a binding model that accounts for depletion are needed to evaluate it.
Worked Example: A One-Site Curve Misses a Steeper Response
Suppose fractional occupancy is measured at five free ligand concentrations. For illustration, assume the observations follow \(\theta_{\mathrm{obs}}=C^2/(100+C^2)\), while a proposed one-site model uses \(K_D=10\) nM and predicts \(\hat{\theta}=C/(10+C)\). The response is fractional occupancy, so it has no physical unit.
| \(C\) (nM) | Observed occupancy | One-site prediction | Residual \(r_i\) |
|---|---|---|---|
| 1 | 0.00990 | 0.0909 | -0.0810 |
| 3 | 0.0826 | 0.2308 | -0.1482 |
| 10 | 0.500 | 0.500 | 0.000 |
| 30 | 0.900 | 0.750 | 0.150 |
| 100 | 0.990 | 0.909 | 0.0810 |
At \(C=3\) nM, for example, the observed value is \(3^2/(100+3^2)=9/109=0.0826\), and the one-site prediction is \(3/(10+3)=0.2308\). Thus \(r=0.0826-0.2308=-0.1482\): the proposed model overpredicts occupancy there. At \(C=30\) nM, the observed value is \(900/1000=0.900\) and the prediction is \(30/40=0.750\), giving \(r=0.150\). The progression from negative residuals at lower concentrations to positive residuals at higher concentrations indicates that the one-site curve is too gradual through this range.
The comparison holds the one-site parameters fixed at the stated values; refitting them would change the individual residuals. The pattern still illustrates what to look for: systematic curvature or a run of residuals with the same sign is more informative than any single difference. The generating curve has a Hill exponent of 2, but this does not establish that the physical system has exactly two binding sites. A steep curve can have multiple explanations, so an independent mechanistic test is needed.
Omitted Nonspecific Binding
In many assays, measured signal can include both specific and nonspecific binding. A simple representation is
where \(b\) is background and \(k_{\mathrm{NS}}\) is a nonspecific-binding slope. This form assumes the nonspecific component is approximately linear over the concentrations considered. If the fitted model omits that component, its specific-binding term approaches a plateau while the data may continue to rise. Residuals that become increasingly positive at high concentrations are therefore a warning sign for omitted nonspecific binding, among other possibilities such as an incorrect upper plateau or concentration-dependent assay interference.
Worked Example: Residuals Reveal a Rising High-Concentration Signal
Consider hypothetical assay responses in signal units with \(b=0\), \(B_{\max}=100\) signal units, \(K_D=10\) nM, and \(k_{\mathrm{NS}}=0.20\) signal units per nM. Suppose the proposed model includes the specific term but omits the nonspecific term. The values below are calculated from the stated parameters.
| \(C\) (nM) | Specific signal (units) | Nonspecific signal (units) | Observed total (units) | Omitted-term prediction (units) | Residual (units) |
|---|---|---|---|---|---|
| 1 | 9.091 | 0.200 | 9.291 | 9.091 | 0.200 |
| 10 | 50.00 | 2.000 | 52.00 | 50.00 | 2.000 |
| 100 | 90.91 | 20.00 | 110.91 | 90.91 | 20.00 |
At 100 nM, the specific signal is \(100(100)/(10+100)=90.91\) units, and the nonspecific signal is \(0.20(100)=20.00\) units. The observed total is \(110.91\) units; the model predicts only \(90.91\) units, so \(r=110.91-90.91=20.00\) units. The residual grows with concentration exactly as the omitted linear component does. A useful follow-up is to measure nonspecific binding under an appropriate blocking condition or to extend the model only when the assay design supports estimating that component.
Time-Course Residuals and the Equilibrium Assumption
An equilibrium model predicts the same occupancy whenever free concentration is held constant. That assumption is questionable if binding has not had enough time to approach equilibrium. For reversible one-site binding under constant free concentration \(C\), a kinetic model is
The first term describes association to unoccupied sites and the second term describes dissociation from occupied sites. With constant \(C\), the equilibrium occupancy is \(k_{\mathrm{on}}C/(k_{\mathrm{on}}C+k_{\mathrm{off}})=C/(K_D+C)\). Starting at occupancy \(\theta(0)\), the solution approaches that equilibrium exponentially:
This result follows by subtracting the equilibrium value from both sides of the differential equation. Let \(u(t)=\theta(t)-\theta_{\mathrm{eq}}\). Then \(du/dt=-(k_{\mathrm{on}}C+k_{\mathrm{off}})u\), whose solution is \(u(t)=u(0)e^{-(k_{\mathrm{on}}C+k_{\mathrm{off}})t}\). If the system starts below equilibrium, \(u(0)<0\), so occupancy rises toward equilibrium from below.
Worked Example: Equilibrium Model Overpredicts Early Occupancy
Suppose free ligand is held at \(C=10\) nM, initial occupancy is zero, \(k_{\mathrm{on}}=0.10\) nM\(^{-1}\) h\(^{-1}\), and \(k_{\mathrm{off}}=1.0\) h\(^{-1}\). These values imply \(K_D=1.0/0.10=10\) nM and equilibrium occupancy \(10/(10+10)=0.500\). Under the kinetic model, occupancy is \(\theta(t)=0.500(1-e^{-2t})\), with \(t\) in hours.
At 0.10 h, kinetic occupancy is \(0.500(1-e^{-0.2})=0.0906\). An equilibrium model predicts 0.500, so the residual is \(0.0906-0.500=-0.4094\). At 0.50 h, kinetic occupancy is \(0.500(1-e^{-1})=0.3161\), giving \(r=0.3161-0.500=-0.1839\). At 2.0 h, occupancy is \(0.500(1-e^{-4})=0.4908\), giving \(r=-0.0092\).
The early observations fall below the equilibrium prediction, so the equilibrium model overpredicts early occupancy; the negative residuals approach zero as equilibrium is approached. This pattern supports considering incomplete association, but it is not conclusive without checking concentration delivery, timing, and assay behavior. A delayed or changing free concentration, for example, can also produce time-structured residuals.
How to Diagnose Without Overinterpreting
A practical diagnostic starts with the signed residuals plotted against the variable linked to the assumption under review. For concentration-response data, inspect residuals against concentration and, when useful, against the model-predicted response. For time-course data, plot residuals against time and mark changes in exposure or assay handling. If measurement precision differs across conditions, raw residual size can be misleading; examine residuals relative to their expected standard errors as well. A standardized residual is \(r_i/SE_i\) when \(SE_i\) is the appropriate standard error for that observation.
Record whether residuals are predominantly positive or negative and where that occurs. Positive residuals mean underprediction; negative residuals mean overprediction.
Curvature, long runs of one sign, or residuals that grow with concentration or time are more concerning than isolated deviations. Replicate measurements help distinguish structure from noise.
Consider curve shape or concentration definition for concentration-related structure; consider equilibrium and timing for time-related structure; consider nonspecific signal when high-concentration residuals rise beyond the predicted plateau.
Use relevant controls, measurements, or a mechanistically justified alternative model. Check whether the change improves residual behavior without creating implausible parameter estimates or unsupported complexity.
Residuals are not independent proof of a particular mechanism. A fitted model’s parameter estimates can absorb some misspecification, and the residual pattern depends on which concentrations and times were sampled. For example, omitting a modest nonspecific component may mostly shift the estimated plateau if data cover a narrow range; extending measurements to high concentrations may reveal the continuing rise. Likewise, too few early time points can conceal slow association. Consider uncertainty in the measured response and in occupancy estimates, including assay-error propagation, when deciding whether a pattern is meaningful.
The main alternatives to residual inspection answer different questions. Residual plots locate where predictions fail and are useful for diagnosing structure. Comparing candidate models can assess relative fit, but a more flexible model may fit better simply because it has more parameters; mechanistic plausibility and parameter identifiability still matter. Assay controls can isolate a component such as nonspecific binding, while additional time points can test whether the equilibrium prediction is reached. Neither a plot nor a model-selection statistic substitutes for design information.
Common Mistakes and Limitations
- Reversing the residual sign. With \(r_i=y_i-\hat{y}_i\), a negative residual means the observation is below the prediction: the model overpredicts. Do not describe it as model underprediction.
- Treating one residual as a diagnosis. A single large difference can arise from ordinary assay noise or a data issue. Look for a systematic, reproducible pattern and inspect the measurement process.
- Calling a steep curve proof of two binding sites. Curve shape can suggest a mismatch with a one-site model, but it does not uniquely establish stoichiometry. Other binding mechanisms or assay effects may look similar.
- Assuming a plateau despite rising data. If signal continues to increase at high concentration, an omitted nonspecific component is one possibility; also check background, interference, and whether the assumed response scale is appropriate.
- Using total concentration as free concentration without justification. The equilibrium occupancy equation uses free ligand. When depletion matters, a model based on nominal concentration can be systematically wrong.
The limits are important: residual diagnostics can reveal that a model is inadequate over the observed design, but they cannot guarantee that the model is correct elsewhere or determine the unique cause of failure. A sound conclusion states the observed pattern, the assumption it challenges, and what evidence would distinguish competing explanations.
Check Your Understanding
For each question, give your answer before reading the concise solution immediately below it.
- If \(y=0.42\) and \(\hat{y}=0.50\), what is the residual and what does its sign mean?
- A one-site model has negative residuals at low concentrations and positive residuals at intermediate-to-high concentrations. What is the most defensible interpretation?
- At high concentrations, measured signal keeps rising while a fitted specific-binding model approaches a plateau. Name one plausible omitted component and one follow-up check.
- Under constant free concentration, an equilibrium model predicts 0.60 occupancy, while early observed occupancy is 0.20. What is the residual, and which direction is the model’s error?
- Why does a concentration-response curve with a Hill exponent of 2 not by itself prove two physical binding sites?
Solutions. (1) \(r=0.42-0.50=-0.08\); the observation is below the prediction, so the model overpredicts by 0.08 response units. (2) The pattern suggests that the one-site curve shape may be too gradual across the observed range, but it does not uniquely identify the mechanism. (3) A linear nonspecific-binding component is plausible; measure binding under an appropriate blocking condition or otherwise validate the nonspecific contribution. (4) \(r=0.20-0.60=-0.40\); the equilibrium model overpredicts early occupancy. (5) A Hill-shaped response is a mathematical description of curve shape, not a unique structural model; other mechanisms or assay effects can produce similar steepness.