Why a Formulation Bridge Needs More Than a Mean Ratio
A new formulation may be intended to deliver the same active drug at the same nominal dose. Yet formulation can change the rate or extent of absorption, so the established dose may not produce the same exposure distribution. A useful bridge therefore asks two related questions: how do the formulations compare in pharmacokinetic measures such as AUC and Cmax, and do any exposure differences meaningfully change predicted clinical benefit or risk?
A geometric mean ratio (GMR) summarizes the ratio of population geometric means for a pharmacokinetic measure under the new and established formulations. Its confidence interval describes uncertainty in that population-level ratio. It is not an interval containing a specified proportion of individual patients’ exposure ratios. To predict the fraction of patients reaching an exposure target, a model of individual exposure variability is also needed.
A common starting point is a randomized crossover study in which each participant receives both formulations, with adequate washout and comparable dose, food conditions, sampling, and adherence. Crossover comparisons can reduce between-person variability, but the design does not remove the need to assess period effects, carryover, missing data, and whether the studied participants and conditions are relevant to the intended use. A common conventional bioequivalence criterion is that the 90% confidence interval for the GMR of AUC and Cmax falls within 0.80 to 1.25. That criterion is a population-mean comparison, not a guarantee of identical individual exposures or clinical outcomes; the appropriate criteria depend on the product and regulatory context.
Worked Example: Interpreting Relative Exposure
Worked Example: Interpreting Relative Exposure
Suppose a hypothetical randomized crossover study compares a new oral formulation with an established formulation at the same dose. The estimated AUC GMR is 0.88, with a 90% confidence interval of 0.84 to 0.92. The Cmax GMR is 1.06, with a 90% confidence interval of 1.00 to 1.12. Assume the crossover conduct, washout, and sampling are adequate for this illustration.
For AUC, compare the confidence limits with the prespecified conventional bounds: \(0.80 \leq 0.84\) and \(0.92 \leq 1.25\). Thus the interval is wholly within 0.80 to 1.25. On the log scale, the same comparison is \(\ln(0.80)=-0.2231\), \(\ln(0.84)=-0.1744\), \(\ln(0.92)=-0.08338\), and \(\ln(1.25)=0.2231\); both confidence limits are inside the bounds.
The AUC point estimate indicates an estimated 12% lower population geometric mean with the new formulation, since \(1-0.88=0.12\). The Cmax estimate indicates a 6% higher population geometric mean, since \(1.06-1=0.06\); its interval also lies within the conventional bounds. Under the stated criterion, these results support comparable population-average exposure.
They do not establish that every patient has comparable exposure, nor do they by themselves answer whether the established dose should be retained. That decision also depends on the location and steepness of the exposure–response relationships, patient-level variability, and the consequences of falling below an efficacy target or exceeding a safety threshold.
From Population Variability to Target Attainment
For a practical illustration, suppose individual steady-state AUC values are approximately lognormally distributed. This is an assumption about the distribution of exposure in the target population, not a consequence of observing a GMR. Let the natural logarithm of exposure be normally distributed with mean \(\mu_{\log}\) and standard deviation \(\sigma_{\log}\). The median exposure is then \(\exp(\mu_{\log})\). If the coefficient of variation (CV) is known, the log-scale standard deviation can be derived from the lognormal variance.
This relation follows because for a lognormal random variable, \(E(X)=\exp(\mu_{\log}+\sigma_{\log}^2/2)\) and \(\operatorname{Var}(X)=[\exp(\sigma_{\log}^2)-1]\exp(2\mu_{\log}+\sigma_{\log}^2)\). Dividing the variance by the squared mean cancels the terms involving \(\mu_{\log}\), leaving \(CV^2=\exp(\sigma_{\log}^2)-1\). If \(M\) is the median exposure, then \(P(X\geq T)=1-\Phi(\ln(T/M)/\sigma_{\log})\), where \(T\) is the target and \(\Phi\) is the standard normal cumulative distribution function.
Worked Example: Comparing Efficacy-Target Attainment
Suppose a hypothetical exposure–response analysis supports an AUC target of 700 mg·h/L for a prespecified efficacy criterion. At the established dose, the median AUC is estimated as 1,000 mg·h/L; at the same nominal dose with the new formulation, it is 880 mg·h/L. Assume a CV of 35% under both formulations and a lognormal exposure distribution. For this decision exercise, the development team has prespecified that at least 80% target attainment is desirable.
First, convert the CV to the log-scale standard deviation:
For the established formulation, the standardized threshold is
Therefore, the predicted proportion at or above the target is \(1-\Phi(-1.049)=\Phi(1.049)\approx0.853\), or about 85.3%.
For the new formulation at the same dose,
The predicted target attainment is \(1-\Phi(-0.6732)=\Phi(0.6732)\approx0.750\), or about 75.0%. Under the stated 80% decision criterion, the established dose meets the target while the new formulation at the same nominal dose does not. This is a model-based prediction conditional on the assumed median, CV, lognormal distribution, and exposure target; it is not a direct observation that exactly those proportions of future patients will respond.
Linking an Exposure Difference to a Dose Decision
An exposure difference matters clinically only to the extent that it changes outcomes. The exposure–response model should therefore be evaluated over the exposure distributions predicted for both formulations, not only at their geometric means. If benefit rises sharply near a target, a modest downward shift may materially reduce target attainment. If benefit has already plateaued, the same shift may have little consequence. Similarly, safety predictions should focus on the exposure range associated with a relevant adverse effect, rather than treating any pharmacokinetic difference as automatically important.
A dose adjustment can be considered as a candidate when exposure is approximately dose-proportional over the relevant range. If the new formulation’s AUC is 0.88 times that of the established formulation at the same dose, an initial proportional adjustment to match median AUC is \(1/0.88\), or about 1.136 times the dose. This calculation is not itself a recommendation: it assumes dose proportionality, and matching AUC does not ensure matching Cmax or all clinically relevant exposure features. The adjusted candidate must be checked against both efficacy and safety predictions.
Worked Example: Checking a Candidate Dose Adjustment Against Safety
Continue the hypothetical case. Suppose the established formulation has a median Cmax of 79.4 mg/L, and the new formulation’s Cmax GMR is 1.06. Assume Cmax is lognormally distributed with a CV of 30% for both formulations. For this illustration, an exposure–safety analysis identifies 120 mg/L as a threshold of concern, and the team’s prespecified decision rule is that no more than 25% of patients should exceed it.
The log-scale standard deviation is
At the established dose, the median Cmax is 79.4 mg/L. The standardized value for the 120 mg/L threshold is \(\ln(120/79.4)/0.29356\approx1.407\), so the predicted exceedance probability is \(1-\Phi(1.407)\approx0.080\), or about 8.0%.
At the same nominal dose, the new formulation’s median Cmax is \(79.4\times1.06=84.164\) mg/L. The candidate AUC-matching dose multiplier from the prior example is \(1/0.88\approx1.136\). If Cmax scales proportionally with dose, the adjusted median Cmax is \(84.164\times1.136\approx95.64\) mg/L. Its standardized threshold is \(\ln(120/95.64)/0.29356\approx0.773\), giving an exceedance probability of \(1-\Phi(0.773)\approx0.220\), or about 22.0%.
Under these explicitly hypothetical rules, the candidate adjustment predicts AUC target attainment of about 85.3% and Cmax-threshold exceedance of about 22.0%; both meet the stated 80% and 25% criteria. This supports evaluating the adjusted dose further, not automatically adopting it. In particular, uncertainty in the GMRs, exposure variability, exposure–response parameters, and dose proportionality should be propagated before a decision is finalized.
Assumptions, Diagnostics, and Alternatives
A formulation bridge is only as reliable as its inputs. Check whether the PK analysis estimates the intended comparison, whether sampling captures relevant peaks and total exposure, and whether food, adherence, and study conditions match the intended use. Examine the exposure data for skewness, outliers, and formulation-specific variability. A lognormal model can be a useful approximation, but poor fit or subgroups with materially different exposure may make one median and one CV inadequate. If the formulation changes the shape of the concentration–time profile, equal AUC does not imply equal peak, trough, or time above a pharmacologically relevant concentration.
The central model assumptions in the numerical examples were: lognormal individual exposure; the stated CV applying to the target population and formulation; a meaningful threshold supplied by a relevant exposure–response analysis; and proportional scaling of exposure with dose for the adjustment calculation. Each requires justification from data or established pharmacology. If these assumptions fail, use a model that better represents the observed distribution and exposure–response relationship, and make predictions across the full exposure profile where needed.
The main alternatives answer different questions. A conventional bioequivalence analysis tests whether confidence intervals for population geometric mean ratios lie within prespecified bounds. It is useful for a standardized PK comparison, but does not directly quantify clinical target attainment. A direct clinical-outcome comparison can assess outcomes in patients, but may require a larger or longer study and may be less informative about why outcomes differ. A model-informed bridge combines PK and exposure–response evidence to project clinical consequences, but depends on model adequacy and uncertainty. These approaches can be complementary; none makes the other evidence irrelevant.
A common mistake is to interpret a GMR of 0.88 as meaning that every patient receives 12% less exposure, or that exactly 88% of patients reach an efficacy target. The GMR is a ratio of population geometric means. Individual exposures vary around those means, and target attainment depends on the full distribution and the target’s location. Another mistake is to apply a dose increase that matches AUC and assume safety is unchanged. The Cmax calculation above shows why each clinically relevant exposure measure and its response relationship must be checked separately.
Specify the formulations, dose, conditions, PK measures, study design, and prespecified interpretation criteria.
Estimate GMRs and confidence intervals, and distinguish uncertainty in population means from variability among individuals.
Use justified exposure distributions and exposure–response relationships to evaluate efficacy and safety for the established and candidate doses.
Retain or adjust the dose only if the relevant predictions meet prespecified criteria, and assess uncertainty and assumptions before implementation.
Check Your Understanding
Answer each question using the assumptions stated, then compare with the solutions.
- A lognormal exposure distribution has CV 20%. Calculate \(\sigma_{\log}\).
- A hypothetical formulation has median exposure 900 mg·h/L, CV 25%, and an efficacy target of 750 mg·h/L. Estimate the proportion at or above target.
- A GMR estimate is 0.78 with a 90% confidence interval of 0.74 to 0.82. Does this interval meet the conventional 0.80 to 1.25 criterion? What does that result establish?
- Explain why a population GMR confidence interval cannot be used directly as the range containing most individual patients’ exposure ratios.
Solutions. (1) \(\sigma_{\log}=\sqrt{\ln(1+0.20^2)}=\sqrt{\ln(1.04)}\approx0.1980\). (2) \(\sigma_{\log}=\sqrt{\ln(1+0.25^2)}=\sqrt{\ln(1.0625)}\approx0.2462\). Then \(z=\ln(750/900)/0.2462\approx-0.7405\), so target attainment is \(1-\Phi(-0.7405)=\Phi(0.7405)\approx0.770\), or about 77.0%. (3) No: the lower confidence limit, 0.74, is below 0.80. This fails the stated criterion; it does not by itself prove that the formulations differ clinically or determine the dose decision. (4) The confidence interval quantifies uncertainty in the estimated population-average ratio. Individual ratios have a distribution driven by between-person and within-person variability, which requires separate estimation.