DISCUSSION PAPER
Bioequivalence Requirements: Highly Variable Drugs and Highly Variable
Drug Products: Issues and Options
PLEASE NOTE: THESE PAPERS ARE INTENDED ONLY
FOR DISCUSSION AT THE WORKSHOP AND EAC-BB MEETING ON JUNE 26-27, 2003
RESPECTIVELY. THEY SHOULD NOT BE INTERPRETED AS HPFB POLICY. UNTIL SUCH
TIME AS A NEW OR UPDATED POLICY IS FINALIZED AND PUBLISHED, THE THERAPEUTIC
PRODUCTS DIRECTORATE'S (TPD) PRACTICE WITH RESPECT TO BIOEQUIVALENCE REQUIREMENTS
REMAINS UNCHANGED.
PLEASE NOTE THAT WHILE THE EAC-BB MAKES RECOMMENDATIONS
TO THE DIRECTOR GENERAL, TPD, DECISION-MAKING RESPONSIBILITY REMAINS WITH
THE TPD.
Outline of discussion paper
1. Brief overview of sources of variability in bioequivalence studies
2. Brief overview of the Two One-Sided Test
3. Highly variable drugs and highly variable drug products: Problems
4. Chlorpromazine, the archetypal highly variable drug
5. An example of a highly variable drug product
6. An example of a (statistical) subject by formulation interaction
7. A case of a non-linear highly variable drug
8. Possible methods of dealing with highly variable drugs or drug products
9. Conclusions
1. Sources of variability
Highly variable drugs (HVDs) have been defined as drugs in which the
withinsubject variability (WSV) in pharmacokinetics estimated from the
ANOVA-CV equals or exceeds 30% (1). In traditional bioequivalence study
designs based on 2-periods, the factors in the ANVOVA model are: Formulation,
Period, Sequence and Subjects nested within Sequence. These factors account
for all the identifiable sources of between subject variability. Within-subject
variability is contained in the Residual Variance (also called the 'Error
Mean Square or Error Term'). The residual variance is made up of several
components: (i) WSV in absorption, distribution, metabolism and excretion
(ADME) combined with a component of analytical variability, (ii) within-formulation
variability (WFV), (iii) the subject by formulation interaction (S*F)
and (iv) unexplained, random variability. The components of the residual
variance cannot be subdivided further in a 2-period design. The hope is
that the two products in a bioequivalence study are of good pharmaceutical
quality so there is little within formulation variability and there is
negligible subject by formulation interaction. An advantage of replicate
designs, in which the test and reference formulations are each administered
twice is that the subject by formulation interaction can be 'teased out'
of the residual variance and it is possible to estimate within subject
variabilities associated with the test (Swt) and reference (Swr) formulations.
For example, common replicate designs with 2- sequences are (3-period)
TRT vs RTR and (4-period) TRTR vs RTRT.
2. The Two One-Sided Test
The modern concept of bioequivalence is based on a survey of physicians
carried out by Westlake in the 1970s which concluded that a 20% difference
in dose between two formulations would have no clinical significance for
most drugs. Hence bioequivalence limits were set at 80 - 120%. Plasma
concentration dependent measures such as Cmax or AUC are not normally
distributed; they are log normal, and hence bioequivalence limits became
80 - 125% (or ± 0.225 on the natural log scale). In traditional average
bioequivalence based on the Two One-Sided Test (2), the 90% confidence
interval around the geometric mean ratio of the test and reference formulations
is therefore required to fall within bioequivalence limits of 80 - 125%.
The width of the 90% confidence interval depends on the number of subjects
in the study and the magnitude of the residual variance. The ANOVA-CV
is simply the square root of the residual variance multiplied by 100.
3. Highly variable drugs and drug products: Problems
The problem with highly variable drugs is shown in Figure 1 which illustrates
the results of two bioequivalence studies on formulations of drugs A and
B. There are the same number of subjects in each study and the GMR is
the same in both. As far as the Two One-Sided Test is concerned, the only
difference between the two studies is the magnitude of the ANOVA-CV. Drug
A has a low within-subject variability (ANOVACV 15%) and the 90% confidence
interval falls comfortably within the bioequivalence limits of 80 - 125%.
Drug B is highly variable, however, with an ANOVA-CV of 35%. The study
fails because the lower bound of the 90% confidence interval falls below
the lower bioequivalence limit (80%). In other words, the study on drug
B was underpowered, the simple remedy for which would be to repeat the
study with a greater number of subjects. Highly variable drugs are usually
safe drugs with flat dose response curves and application of the present
preset bioequivalence limits of 80-125% amounts to imposition of unwarranted
tougher bioequivalence requirements than for lower variability drugs.
A highly variable drug product (HVDP) is a formulation of poor pharmaceutical
quality in which the drug itself is not highly variable, but there is
a big component of within formulation (tablet to tablet, capsule to capsule,
patch to patch) variability (WFV). HVDPs pose a problem because they are
cannot be detected in traditional 2-treatment, 2- period, 2-sequence cross-over
design studies. Replicate designs, however, facilitate their detection
because the within-subject variabilities of the test and reference formulations
can be estimated separately. When they are very different, the probable
explanation is that one of the formulations is a HVDP. Now if the brand
is a HVDP, then a better quality test product should not be penalized
by forcing it to meet the variability of the poor quality reference product.
4. Chlorpromazine: The archetypal highly variable drug.
Table 1. summarizes the ANOVA-CVs from three different types of studies
on chlorpromazine conducted at different times by three different analysts
by three completely different analytical methods as outlined in the table.
Study-1 was a 3-period bioequivalence study carried out in 37 subjects
in which the test formulation was administered once and the reference
formulation was administered twice. Study-2 was a 3-period pharmacokinetic
study (3) in which an oral solution was administered. Solution data gives
the best estimate of true pharmacokinetic within-subject variability since,
in the absence of a formulation, there is no subject by formulation interaction,
and there is no component of within formulation variability included in
the estimate. Study-3 was a 2-period interaction study in which chlorpromazine
was administered with and without quinidine. The use of solution data
in study-2 shows unequivocally that chlorpromazine behaves as a highly
variable drug in terms of both Cmax and AUC. There was remarkable consistency
of the ANOVA-CVs across the three studies (Table 1) despite the differences
in study design, analytical method and analytical personnel.
Table 2 summarizes the results of the bioequivalence study on two formulations
of chlorpromazine (unpublished data) which is a good illustration of the
kind of problems that beset bioequivalence studies on highly variable
drugs. Despite the large number of subjects (n=37), all comparisons of
Cmax failed US-FDA conditions which require a 90% confidence interval
around this measure to be within 80-125%. The most interesting point here
is that a reference to reference comparison also failed decisively. The
reason is illustrated in Figure 2 which shows stick plots of the parametric
values of ln Cmax (left panel) and ln AUClast (right panel) for each of
the 37 subjects after the two administrations of the reference formulation.
The results for six subjects are shown in Figure 2a and highlighted in
Figure 2b from which the remaining 31 subjects have been deleted. Deletion
of data from these subjects produced a substantial reduction in the ANOVA-CVs
of both Cmax and AUClast. Since chlorpromazine is a genuine highly variable
drug, it is likely that any high quality formulation of it would respond
in a similar manner to the reference formulation in our study. At present
the only recourse to the bioequivalence dilemma is to increase substantially
the number of subjects in the study.
5. An example of a highly variable drug product
After a traditional 2-treatment, 2-period, 2-sequence cross-over bioequivalence
study on two formulations of the beta-blocker nadolol failed, it was decided
to investigate the sources of variability by conducting a 4-period replicate
study. Thus a 2- treatment, 4-period, 4-sequence cross-over study was
conducted in 22 healthy volunteers. The results (Table 3a) indicated a
failure of both AUClast and Cmax in terms of the 90% confidence intervals
failing to fall within preset bioequivalence limits of 80 - 125%, although
Cmax passed the Health Canada requirement for the GMR to fall within these
bioequivalence limits. Examination of the two administrations of the test
formulation (Table 3b) demonstrated that the drug itself was not highly
variable in terms of either measure and the GMR for both was 97%, close
to the ideal. In contrast, however, the reference to reference comparison
(Table 3c) showed both Cmax and AUClast very highly variable, the GMRs
of both measures was 87%, and the reference formulation failed when tested
against itself. The reason for the problem is illustrated graphically
in Figure 3a which shows stick plots of the two values of ln Cmax after
administration of the test (left panel) and reference (right panel) formulations.
Three subjects with the greatest differences in Cmax after the two administrations
of the reference formulation were highlighted (Figure 3a, right panel),
whereas the same subjects had very much smaller differences after the
test formulation (left panel). Figure 3b is the same as Figure 3a with
all but the three most variable subjects deleted. These three subjects
made the greatest contribution to the formulation variability (WFV) of
the reference which behaved as a highly variable drug product. This phenomenon
cannot be attributed to subject by formulation interaction because the
subject means for each formulation were relatively close, despite the
wider spread with the reference formulation. In fact, the subject by formulation
interaction term was negligible for both measures (0% for Cmax and 8%
for AUClast).
This study raises some interesting questions that came to light solely
because of the nature of the replicate design. After the earlier 2-period
bioequivalence study in which within formulation variation and subject
by formulation interaction term were inseparable components of the residual
variance, it was tempting to attribute failure to subject by formulation
interaction. The four period design, however, showed clearly failure was
attributable to a very different problem. One could say that the 'right'
answer was that the test formulation was not bioequivalent with the reference
formulation simply because their variances were so different. Should this
mean then, that a superior quality reference product can never reach the
market?
6. An example of a (statistical) subject by formulation interaction
A bioequivalence study on two percutaneous patch formulations of nitroglycerin
was carried out in 37 subjects in a 2-treatment, 4-period, 4-sequence
cross over design. Serial blood samples were collected over 12 hours after
which the patch was removed to facilitate measurement of the elimination
phase. Table 4a shows the test formulation met bioequivalent requirements
in Cmax and AUClast, despite high variability in both measures. This was
a consequence of the study being adequately powered with a total of 148
observations in the data set. The test to test (Table 4b) and the reference
to reference (Table 4c) comparison also met bioequivalence requirements.
Examination of the variabilities associated with the test and reference
formulations (Tables 4b and 4c) showed them to be roughly comparable,
but there were large subject by formulation interactions associated with
both Cmax (28%) and AUClast (21%). These are depicted in Figures 4a and
4b in which the four measurements for each individual are plotted, with
the two test values slightly to the left of the two reference values.
In figure 4a, eight subjects contributing to the subject by formulation
interaction in Cmax are highlighted, while in Figure 4b, five subjects
contributing to the interaction in AUClast are highlighted. Thus it would
appear that a subset of the population may respond differently to the
two formulations, although the observation was not confirmed by repetition
of the replicate study in the same individuals, or in a second sample
of the population.
Two important points arise from this study. One is the lack of sensitivity
of average bioequivalence to a substantial subject by formulation interaction.
Individual bioequivalence (not shown) is very sensitive to the interaction
and failed the study. The second point concerns the clinical significance
of a subject by formulation interaction with any given drug, which is
a difficult question to study prospectively. Problems in judging the clinical
significance of subject by formulation interaction in general may have
contributed to the demise of individual bioequivalence at US-FDA.
7. A case of a non-linear highly variable drug
Two bioequivalence studies were carried out on two formulations of propafenone
which is a non-linear, highly variable drug. Both studies were traditional
2-treatment, 2- period, two sequence cross-over studies. In the first
study, 74 healthy subjects were dosed after an overnight fast, and in
the second, 25 healthy subjects were dosed after a standardized high fat
breakfast. The first study was successful in that both measures met bioequivalence
requirements despite high within-subject variability (Table 5a) The Fed
study was not powered sufficiently for the 90% confidence interval around
ln Cmax to fall within bioequivalence limits of 80-125% (Table 5b). The
question remains, however, should it be necessary to subject 74 healthy
volunteers to two doses of propafenone in order to test two formulations
of the drug for bioequivalence?
8. Possible methods of dealing with highly variable drugs and drug products
a) No confidence interval for Cmax of HVDs/HVDPs
Cmax is often the most variable of the two measures, partly because it
is a single point determinant which is dependent upon an adequate blood
sampling schedule around tmax. A simple method of dealing with highly
variable drugs would be to treat them as 'uncomplicated drugs' and not
require a 90% confidence interval around Cmax. Each of the four examples
discussed in this paper were also highly variable in terms of AUClast
but its variability was less than Cmax (in test versus reference comparisons)
which means a smaller number of subjects (observations) would have been
required to achieve adequate statistical power.
b) Arbitrarily broaden the bioequivalence limits for HVDs
Since highly variable drugs are generally safe drugs with shallow dose
response curves, it is reasonable to tolerate greater than 20% differences
between test and reference formulations. The EU-CPMP guidelines, for example,
permit a sponsor prospectively to justify broadening the bioequivalence
limits from 80 - 125% to, say, 75 - 133%.
c) Broaden the bioequivalence limits according to the within-subject
variability of the reference formulation
The bioequivalence limits can be scaled to the within-subject variability
of the reference formulation by the use of the residual variance in a
two-period design or the within-subject variance associated with the reference
formulation in a replicate design. The fundamental concept is shown in
Equation 1. which implies the 90% confidence interval around the difference
between the log transformed means of the test and reference formulation
must fit between bioequivalence limits of qABE which is set by the drug
regulatory body. The commonly accepted bioequivalence limits are set at
80 - 125% (0.8 - 1.25) which is ±0.225 on the natural log scale. In scaling,
the bioequivalence
-0.223
( µt - µr)
0.223 ![blank](/web/20061214032302im_/http://www.hc-sc.gc.ca/dhp-mps/medeff/advisories-avis/prof/images/blank.gif) ![blank](/web/20061214032302im_/http://www.hc-sc.gc.ca/dhp-mps/medeff/advisories-avis/prof/images/blank.gif) ![blank](/web/20061214032302im_/http://www.hc-sc.gc.ca/dhp-mps/medeff/advisories-avis/prof/images/blank.gif) (
µt - µr)2
0.2232 Equation
1.
limits are divided by the within-subject standard deviation at which
the limits are to be permitted to be broadened ( WO).
The latter is to be set by a drug regulatory agency. The left hand side
of Equation 1 is divided by the within subject standard deviation of the
reference formulation ( WR)
Equation 2. Rearrangement of Equation 2 gives Equation 3 which broadens
the bioequivalence limits for highly variable drugs in a systematic way.
![equations](/web/20061214032302im_/http://www.hc-sc.gc.ca/dhp-mps/prodpharma/activit/sci-com/bio/images/hvd_ps_discussionpaper_p12_e.gif)
The method of Boddy and Co-workers (4) for widening the bioequivalence
limits for highly variable drugs/products is a slightly different approach
to the same concept. Here the bioequivalence limits are set as a fixed
multiple (k) of the standard deviation ( WR
or RES).
In other words, ( ABE
= k WR),
such that k represents the number of standard deviations by which the
means are allowed to vary. Thus, when k is set at (0.223/ WO)
the relationship becomes exactly the same as Equation 3.
Scaling the ABE metric amounts to the same thing as scaling the bioequivalence
limits, since the relationships shown in Equations 1-3 are used for both
methods. For the purposes of illustration, the point at which the bioequivalence
limits are permitted to be broadened by scaling ( WO)
was set at 0.20 or 0.25. The results are shown in Figure 5, and some specific
examples with Swt = 0.3, 0.5 or 0.7 are shown in Table 6. The value WO
= 0.20 was selected by US-FDA during the decade long debate on individual
bioequivalence, although for scaled ABE, we prefer the more conservative
WO
= 0.25.
9. Conclusions
a) Scaled versus unscaled average bioequivalence
Unscaled average bioequivalence is very sensitive to differences between
the means whereas scaled bioequivalence is much less sensitive. Scaling
can allow the geometric mean ratio to rise to unacceptably high levels
when the reference variance is very high, and it may be wise to place
a limit on the maximum GMR for bioequivalence. For example, the GMR could
be required to fall within 80 - 125%. If reference scaling is to be allowed,
TPD must set WO.
We suggest WO
= 0.25 since WO
= 0.20 appears to be too liberal. (However, WO
can be set at any reasonable value)
It is our view that average bioequivalence based on replicate designs
has merit because valuable evidence on the pharmaceutical quality is provided
in that disparities between the test and reference variances are made
apparent. Substantial subject by formulation interactions, if present,
are also detected, but TPD may prefer to ignore them.
b) Subject by formulation interaction
As illustrated by the example of the bioequivalence study on the two
patch formulations of nitroglycerin, average bioequivalence is not sensitive
to the presence of a substantial subject by formulation interaction whereas
individual bioequivalence is very sensitive. When a study is conducted
by replicate design, however, a substantial subject by formulation interaction
will be detected, even though average bioequivalence is not sensitive
to it. If an interaction is deemed to be important clinically, then a
regulatory agency can take any action, including failing the study because
of it.
c) Highly variable drug products.
As stated earlier, an advantage of replicate design is that data are
provided on the within-subject variabilities of the test and reference
formulations. If the test formulation is shown to be a highly variable
drug product, then it should be allowed to fail average bioequivalence
without resort to scaling. When an innovator formulation behaves as a
highly variable drug product, the problem is more difficult to resolve
because that was the formulation linked to pivotal clinical trials during
the therapeutic and toxicological evaluation of the drug substance. By
the time generic formulations appear, the innovator's highly variable
drug product will have been on the market for several years, presumably
without major untoward incident. One could surmise then that introduction
of a better quality, less variable generic product would pose no hazard
to the patient. Therefore we would recommend reference scaling be allowed
when the reference formulation is a highly variable drug product.
- H. H. Blume, and K. K. Midha. Bio-International '92, Conference on
Bioavailability, Bioequivalence and Pharmacokinetic Studies. Pharmaceutical
Research 10:1806-1811 (1993).
- D. J. Schuirmann. A comparison of the two one-sided tests procedure
and the power approach for assessing the equivalence of average bioavailability.
J. Pharmacokinet. Biopharm 15:657-680 (1987).
- P. K.-F. Yeung, J. W. Hubbard, E. D. Korchinski, and K. K. Midha.
Pharmacokinetics of chlorpromazine and key metabolites. European
Journal of Clinical Pharmacology 45:563-569 (1993).
- A. W. Boddy, F. C. Snikeris, R. O. Kringle, G. C. G. Wei, J. A. Oppermann,
and K. K. Midha. An approach for widening the bioequivalence acceptance
limits in the case of highly variable drugs. Pharmaceutical Research
12:1865-1868 (1995).
|