Why Use a Nonlinear or Piecewise Level A Mapping?
A Level A in vitro-in vivo correlation (IVIVC) links dissolution to absorption point by point. After any needed time scaling or lag-time alignment, a simple starting point is that a given fraction dissolved corresponds to the same fraction absorbed. That linear mapping is not always adequate: the fraction absorbed may rise slowly at first and then more rapidly, or the relationship may change slope over a range of dissolution. A nonlinear or piecewise mapping can represent such patterns, but its extra flexibility is useful only when the data can support it and its predictions remain credible.
Write \(D_j(t)\) for the fraction dissolved in vitro for formulation \(j\) by time \(t\), and \(A_j(t)\) for its fraction absorbed in vivo by that time. Both are fractions between 0 and 1. The mapping concerns the relationship between values of \(D\) and \(A\), not a direct relationship between dissolution time and plasma concentration. Once a mapping \(A=g(D)\) is specified, its implications for absorption rate can be combined with drug disposition to predict a plasma concentration-time profile.
The endpoint conditions are modeling choices that should match the product and the data. If only a fraction of the dose can be absorbed, the appropriate endpoint may differ from 1; forcing \(g(1)=1\) in that situation would misrepresent the process. The examples below use the complete-absorption convention. They also assume that \(D\) and \(A\) are measured or estimated on compatible dose-fraction scales, the in vitro method represents release relevant to absorption, and the in vivo estimates are sufficiently informative to support estimation of the mapping. Time alignment is treated as established; the mapping should not be used to conceal a timing mismatch.
Nonlinear Mappings
A compact nonlinear model can be useful when the data show smooth curvature. One simple form is the power mapping \(g(D)=D^\gamma\), where \(\gamma>0\). It respects both endpoints and is nondecreasing. If \(\gamma=1\), it reduces to the linear identity mapping. If \(\gamma>1\), then \(g(D)<D\) for intermediate fractions: absorption accumulates more slowly than dissolution on the fraction scale. If \(0<\gamma<1\), then \(g(D)>D\) for \(0<D<1\).
Worked Example: A Power-Law Level A Mapping
Suppose hypothetical, time-aligned data include a point with \(D=0.25\) and estimated \(A=0.125\). Fit the one-parameter power mapping through this point, then predict absorption when \(D=0.64\).
Under \(A=D^\gamma\), the observed pair gives \(0.125=0.25^\gamma\). Taking logarithms, \(\gamma=\log(0.125)/\log(0.25)=1.5\), since \(0.25^{1.5}=0.125\). The fitted mapping is therefore \(g(D)=D^{1.5}\).
At \(D=0.64\), the predicted fraction absorbed is \(g(0.64)=0.64^{1.5}=0.64\sqrt{0.64}=0.64(0.8)=0.512\). Thus the model predicts 51.2% absorbed at this dissolution fraction. The shape is consistent with the assumptions: for \(\gamma=1.5>0\), the mapping is increasing, and \(g(0)=0\), \(g(1)=1\). One paired point is enough to solve this idealized one-parameter equation, but it is not enough to establish that a common power law adequately describes several formulations or the range of interest.
A useful justification follows from the chain rule. If \(A(t)=g(D(t))\) and the curves are differentiable, then the instantaneous fraction absorbed per unit time is \(dA/dt=g'(D(t))\,dD/dt\). Thus a nonlinear mapping changes absorption rate according to both the dissolution rate and the local slope of the mapping. For the power model, \(g'(D)=\gamma D^{\gamma-1}\) for \(D>0\). A mapping should therefore be judged not just by how closely it follows paired fractions, but also by whether the implied absorption-rate pattern is plausible and supports useful plasma predictions when combined with disposition.
Continuous Piecewise-Linear Mappings
A piecewise-linear mapping allows different slopes below and above a prespecified dissolution fraction \(c\), called the breakpoint. For \(0<c<1\), one continuous form is \(g(D)=m_1D\) for \(D\le c\), and \(g(D)=m_1c+m_2(D-c)\) for \(D>c\). Continuity is built in: both expressions equal \(m_1c\) at \(D=c\). To impose \(g(1)=1\), the slopes must satisfy \(m_1c+m_2(1-c)=1\). A nondecreasing mapping additionally requires \(m_1\ge0\) and \(m_2\ge0\).
Worked Example: Fitting a Constrained Two-Slope Mapping
Consider hypothetical paired fractions with dissolution values \(D=0.2, 0.4, 0.6, 0.8, 1.0\) and corresponding estimated absorption values \(A=0.15, 0.31, 0.49, 0.75, 1.00\). For illustration, choose \(c=0.6\) and \(m_1=0.8\), then enforce the endpoint condition to determine \(m_2\).
The endpoint equation gives \(0.8(0.6)+m_2(1-0.6)=1\), so \(m_2=(1-0.48)/0.4=1.3\). Both slopes are positive. The mapping is \(g(D)=0.8D\) when \(D\le0.6\), and \(g(D)=0.48+1.3(D-0.6)\) when \(D>0.6\). It is continuous at \(0.6\), where both expressions give \(0.48\), and reaches 1 at \(D=1\).
The fitted values at the observed dissolution fractions are \(0.16, 0.32, 0.48, 0.74,\) and \(1.00\). Subtracting the observations gives residuals \(0.01, 0.01, -0.01, -0.01,\) and \(0\), so the sum of squared residuals is \(0.01^2+0.01^2+(-0.01)^2+(-0.01)^2+0^2=0.0004\). This small discrepancy illustrates the arithmetic of fitting the mapping; by itself, it does not establish that the breakpoint or the piecewise form is scientifically justified.
The breakpoint may be fixed using prior scientific reasoning or estimated from data. Estimating it adds flexibility and can make the fitting problem harder: the objective may change irregularly as the breakpoint moves, and a solution can depend on starting values or constraints. The breakpoint should lie within a range with observations on both sides. Otherwise the data provide little or no information about one of the slopes.
Identifiability, Overfitting, and Prediction
A parameter is identifiable when the available information can distinguish its value from plausible alternatives. A low fitting error does not guarantee identifiability. In the piecewise model, if all observed dissolution fractions lie below \(c\), their predicted absorption depends on \(m_1\) but not \(m_2\). The endpoint condition can calculate \(m_2\) for any chosen \(c\), but this algebraic constraint does not make \(c\) empirically identified: the data have not tested the upper segment.
Worked Example: Detecting an Unidentified Breakpoint
Suppose the only informative hypothetical pairs are \(D=0.2, A=0.16\) and \(D=0.4, A=0.32\). Each is exactly described by the lower-segment slope \(m_1=0.8\). Compare two candidate breakpoints, \(c=0.6\) and \(c=0.8\), with the upper slope chosen to satisfy \(g(1)=1\).
For \(c=0.6\), \(m_2=(1-0.8(0.6))/(1-0.6)=1.3\). For \(c=0.8\), \(m_2=(1-0.8(0.8))/(1-0.8)=1.8\). Both models exactly fit the two observed pairs, since both values of \(c\) exceed the largest observed \(D=0.4\). Yet their predictions at \(D=0.7\) differ: the \(c=0.6\) model gives \(0.48+1.3(0.1)=0.61\), while the \(c=0.8\) model gives \(0.8(0.7)=0.56\). The breakpoint and upper-segment behavior cannot be learned from these observations. More decimal places in the reported fit would not solve the information problem; data spanning the upper segment are needed.
Overfitting occurs when a flexible model follows peculiarities of the observed data that do not carry over to new formulations or profiles. Many time points from one formulation do not necessarily provide the same evidence as informative profiles from several formulations: points from a single profile are related, and they may cover only a limited range of dissolution. A complex spline or several breakpoints can produce a close in-sample fit while yielding unstable shapes or implausible predictions outside the observed range.
Predictive utility asks whether the fitted mapping, combined with the disposition model, predicts relevant in vivo behavior for data not used to fit it. When feasible, hold out an entire formulation rather than isolated time points from a profile: the latter can make the test too similar to the fitting data. Compare predicted and observed plasma profiles or prespecified exposure measures in context, and examine whether discrepancies matter for the intended use. A nonlinear mapping that reduces the fitting error but does not improve predictions is not automatically preferable to a simpler model.
Diagnostics, Alternatives, and Limits
Inspect paired \(D\)-and-\(A\) values and the fitted curve over the observed range. Check that the mapping respects the intended endpoints, does not decrease if monotonic absorption is assumed, and does not imply implausibly abrupt changes in absorption rate. Look for observations concentrated on one side of a breakpoint, a fitted breakpoint at the edge of the data, parameter estimates that change markedly when a formulation is removed, or competing parameter values with nearly identical fits. These are warnings about weak information or model sensitivity, not problems that can be repaired simply by adding more parameters.
The linear identity mapping is the simplest alternative and is preferable when it adequately represents the data and predictions. A smooth parametric curve, such as the power model, uses few parameters and avoids a slope discontinuity, but may impose the wrong curvature. A piecewise-linear model can represent a change in slope transparently, but its breakpoint and segment slopes need support across the range. More flexible monotone curves may represent complex shapes, but require still more information and careful constraints. These approaches are alternatives for the same mapping problem; none removes the need for a biologically plausible dissolution input, reliable absorption information, and informative formulation profiles.
A Level A mapping is limited by the quality and relevance of both its in vitro and in vivo inputs. Estimated fractions absorbed can carry uncertainty, and treating them as exact can make parameter certainty appear stronger than it is. A relationship learned over a narrow range should not be assumed to extrapolate to substantially different dissolution behavior. Further, an apparent nonlinear mapping can reflect timing misalignment, measurement error, or formulation-specific behavior rather than a universal dissolution-to-absorption mechanism. The model is a useful predictive construct only to the extent that its assumptions and intended range are supported.
Check Your Understanding
Answer each question, then compare with the solutions below.
- For \(g(D)=D^{0.7}\), is the mapping increasing, and is \(g(D)\) above or below \(D\) for intermediate fractions?
- A piecewise mapping has \(c=0.5\) and \(m_1=0.6\). What upper slope is required for \(g(1)=1\)?
- All observed dissolution fractions are below a candidate breakpoint. What can those observations tell you about the upper slope?
- Why is holding out an entire formulation usually more informative than holding out individual time points from profiles used to fit the mapping?
Solutions. (1) Since \(g'(D)=0.7D^{-0.3}>0\) for \(D>0\), it is increasing. For \(0<D<1\), raising \(D\) to an exponent below 1 gives \(D^{0.7}>D\). (2) The endpoint constraint is \(0.6(0.5)+m_2(0.5)=1\), so \(m_2=1.4\). (3) They inform the lower segment only; without observations above the breakpoint, the upper slope is not empirically identified, even if it is calculated from the endpoint constraint. (4) Time points within one profile are related and may be very similar to points used for fitting. Holding out a formulation better tests whether the mapping carries across formulations.