Building a Level A IVIVC From Multiple Formulations

From Formulation Profiles to a Point-to-Point Relationship

A Level A in vitro-in vivo correlation (IVIVC) links the progress of drug release measured in vitro to the progress of drug input into the body. Its defining feature is a point-to-point comparison: each in vitro dissolution value is paired with a corresponding in vivo input value. With multiple formulations that release at different rates, the key question is whether one relationship describes all of those paired profiles—not merely whether faster-dissolving formulations also tend to be absorbed faster.

For each formulation, the in vivo input profile is estimated from plasma concentration data, commonly by deconvolution using the methods covered in the earlier material on Wagner–Nelson and Loo–Riegelman methods. The result is a cumulative input or absorbed fraction over time. This tutorial treats those estimates as available inputs to the construction; it does not re-derive them. In vitro dissolution and in vivo input must be expressed on compatible fraction scales, with the dose basis and normalization stated.

The fractions are ordinarily expressed between 0 and 1 (or, consistently, as percentages between 0 and 100). A perfect identity relationship has \(a=0\) and \(b=1\): the in vitro fraction dissolved equals the in vivo fraction entered. A fitted relationship need not be the identity, but its intercept, slope, predicted range, and agreement across formulations must make scientific sense. A line that fits one formulation but not the others is not a common multiple-formulation mapping.

Conditions for a Meaningful Fit

The regression calculation is straightforward; establishing that the paired data deserve a shared mapping is the harder part. Before fitting, examine whether the following conditions are plausible.

• Comparable quantities. Dissolution and in vivo input use a common dose basis and consistent fraction scale. Differences in bioavailability or dose normalization must not be mistaken for differences in release.
• Meaningful pairing. Each \(x_{ij}\) must correspond to the in vivo input value \(y_{ij}\) at the intended observation. The time basis must be defensible. If a systematic lag or time shift prevents meaningful pairing, an unshifted point-to-point fit can be misleading; time alignment is a separate modeling question.
• Valid input estimates. Deconvolved profiles are estimates, not direct observations of absorption. Their quality depends on the disposition reference, sampling, and assumptions of the deconvolution method.
• Formulation changes act through release. The interpretation is strongest when formulations principally change release rate and do not independently change absorption, disposition, or the fraction ultimately available. Otherwise, the fitted association may combine multiple mechanisms.
• A shared relationship is plausible. The profiles should cover a useful range of release rates and fractions, and their paired trajectories should be compatible with one mapping. The data should not be dominated by a single formulation or a narrow range of values.

The points from one formulation are repeated observations along a profile. They are not automatically independent replicates. Ordinary least squares can describe a pooled line, but standard regression standard errors and tests that treat every time point as independent do not follow simply from calculating that line. The purpose here is to assess and construct the relationship; uncertainty requires methods that respect the profile and formulation structure.

Fitting a Common Linear Mapping

For a descriptive linear fit, pool the paired observations across formulations and minimize the sum of squared vertical differences between observed input and predicted input. The least-squares estimates are:

Here \(\bar{x}\) and \(\bar{y}\) are the means over the paired observations included in the fit. These equations give the line whose squared residuals, \(\sum_{i,j}(y_{ij}-a-bx_{ij})^2\), are smallest among all lines.

A short derivation explains the estimates. For fixed \(b\), differentiate the sum of squares with respect to \(a\) and set the derivative to zero:

Rearranging gives \(a=\bar{y}-b\bar{x}\). Substituting this expression into the sum of squares and differentiating with respect to \(b\) gives \(\sum_{i,j}(x_{ij}-\bar{x})(y_{ij}-\bar{y})-b\sum_{i,j}(x_{ij}-\bar{x})^2=0\), which yields the stated slope. This derivation establishes the least-squares line; it does not establish that a linear common relationship is scientifically adequate.

Rank Order and Profile Shape

Two complementary checks help reveal whether the pooled relationship represents the formulation profiles. Rank agreement asks whether formulations that dissolve more rapidly in vitro also show more rapid in vivo input. Shape agreement asks whether the paired trajectories follow a similar point-to-point pattern over the range of fractions. Rank agreement is useful but weaker: formulations can preserve their order while differing substantially in the detailed relation between dissolution and input.

A practical display plots each formulation’s in vivo input fraction against its in vitro dissolved fraction, using a distinct line or symbol for each formulation and marking the fitted common mapping. Also inspect the original profiles against time. A pooled scatterplot can hide a formulation-specific pattern if points from different profiles overlap or if one formulation supplies most of the observations.

Why Correlation Alone Is Not Enough

A large correlation or \(R^2\) can occur when both measures rise over time, even if their detailed relationship is biased or curved. Nor does an overall pooled fit guarantee that each formulation follows the same mapping. Assess the fitted intercept and slope against the identity line, inspect predicted values over the observed range, and compare residual patterns by formulation and fraction. Predictions outside the 0-to-1 fraction range are especially problematic for a fraction mapping. A nonzero intercept or a slope far from 1 is not automatically disqualifying, but it needs a plausible interpretation and should not conceal systematic disagreement.

Alternatives, Limitations, and Common Pitfalls

The identity line and a fitted linear mapping answer different questions. The identity line asks whether dissolution fraction and input fraction agree directly, without calibration. A fitted line allows an intercept and slope to describe systematic scale differences, but can appear adequate while relying on an implausible offset or extrapolation. Rank comparisons are simpler and robust to some scale differences, but discard much of the pointwise information. A pooled line uses all paired observations to summarize a common relation; separate formulation fits can reveal heterogeneity, but several unrelated lines do not demonstrate one shared multiple-formulation IVIVC.

Any fitted relationship is limited by the formulations, dose basis, dissolution conditions, sampling range, and in vivo input estimates used to construct it. A good fit in the observed range should not be read as validation for unstudied formulations or as proof that dissolution causes the observed in vivo pattern. Nor should a high pooled \(R^2\) be interpreted as a probability that the model is correct.

Common mistake: “The formulation ranks match and the pooled correlation is near 1, so the point-to-point IVIVC is established.” Rank and correlation summarize ordering and overall association; they do not rule out curvature, formulation-specific deviations, time-pairing problems, or an unsuitable dose scale. The correction is to inspect paired profiles and the fitted relationship by formulation, check the identity-scale interpretation and prediction range, and treat the fit as evidence conditional on its assumptions—not as validation by itself.