BOOTSTRAPPING& DISSOLUTION SIMILARITY: Q&A
Jiri Hofmann
BioBridges 2019
Prague, September 26-27, 2019
SIMILARITY: CLAIM OF EQUIVALENCE
reference
test
Time (hours)
% r
ele
ased
Test Reference
Test A Test B
Biobatch Biobatch
Pilot Clinical
Strength 1 Strength 2
API 1 API 2
+Excipient −Excipient
Site 1 Site 2
…
Type
(M)ANOVA
f1(2)
Bayesian
MD
Bootstrap
MSD
T2 test
EDNE test
PCA
…
(M)ANOVA (multivariate) analysis of variance; f1(2) similarity factor; MD maximum
deviation; MSD multivariate statistical distance; EDNE Euclidean distance of the
nonstandardized expected (values); PCA principal component analysis
FIT FACTOR: MOORE & FLANNER
Sim
ilari
ty f
acto
r
Time (min)
% r
ele
as
ed
reference
test
Mean difference (%)
−10%
𝑓2 = 50 𝑙𝑜𝑔 1 +1
𝑛
𝑖=1
𝑛
𝑹𝒕 − 𝑻𝒕2
−0.5
× 100
Moore & Flanner (1996). Pharm Technol 20(6): 64-74
𝒇𝟐 = 𝟒𝟗. 𝟖𝟗𝟏𝟗𝟕
Time (min)
% r
ele
ased
reference
test
+3.7%
−0.4%
−6.0%
−10.7%
−13.7%
−15.4%
𝒇𝟐 = 𝟓𝟎. 𝟏𝟎
9.95%
similar
dissimilar
−10%−10%
−10%−10%
SIMILARITY FACTOR: WHAT‘S WRONG WITH YOU?
Time (min)
% r
ele
ased
𝒇𝟐 (𝟏−𝟔) = 𝟓𝟏. 𝟖𝟓
Time (min)
reference
test
Time (min)
Shape & time correlationVariabilityNumber of points
1
3
2
45 6
𝒇𝟐 (𝟏−𝟒) = 𝟒𝟕. 𝟖𝟗
3
5
2
64
1
𝒇𝟐 (𝟏−𝟔) = 𝟓𝟏. 𝟖𝟓
85%
𝑓2 = 50 𝑙𝑜𝑔 1 +1
𝑛
𝑖=1
𝑛
𝑹𝒕 − 𝑻𝒕2
−0.5
× 100
(…) impossible to evaluate false
positive or false negative (…). (…) too
liberal in concluding similarity.
Liu et al. (1997). Drug Info J. 31: 1255-1271
(…) f2 (…) series of monotone (…)
transformations of the Euclidean
distance, (…) procedure where neither
the type I error rate nor power can be
controlled and which really hurts a
statistician‘s soul.
Hoffelder (2019). Biom J. 61(5): 1120-1137
STATISTICIANS: VIEW ON SIMILARITY FACTOR
Cartesian coordinate system
𝒅(𝒑, 𝒒)
𝑨
𝑩2
3
𝒑𝟏 𝒒𝟏
𝒑𝟐
𝒒𝟐
x
y
EMA: REFLECTIONS
[6.3] The f2 (…) unfavorable statistical
properties (…), no (…) quantification
of the risk to false positively conclude
on similar dissolution is possible.
[Answer:] (…) is based only upon the
(…) numerical value for f2 (point
estimate ≥ 𝟓𝟎). (…) uncertainty related
to the f2 sampling distribution is not
accounted for.
23 March 2017
EMA/CHMP/138502/2017
Committee for Human Medicinal Products (CHMP)
Reflection paper on statisticalmethodology for the comparativeassessment of quality attributes indrug development.
26 July 2018
EMA/810713/2017
Human Medicines Research and Development Support
Question and answer on the adequacyof the Mahalanobis distance to assessthe comparability of drug dissolutionprofiles
The aim (…) to a larger extent to emphasise the importanceof confidence intervals to quantify the uncertainty aroundthe point estimate of the chosen metric (…).
1
2
3
4
5
6
7
DISSOLUTION SIMULATION: 𝑦 = 100 × 1 − 𝑒−𝑘𝑡
Population 𝒇𝟐
Sim
ilari
ty (
%)
Time (min)
% r
ele
ased
reference
test
Population 𝒇𝟐 = 𝟒𝟎
𝒌 = 𝟎. 𝟎𝟔𝟎𝟎
𝒌 = 𝟎. 𝟎𝟑𝟓𝟔
Empirical power
𝒇𝟐 = 𝟑𝟖
𝒇𝟐 = 𝟒𝟐
1
N
N=10‘000
Samples
𝒇𝟐 > 𝟓𝟎N
𝜶 = 𝟓%
Time (min)
% r
ele
ased
𝒌
ෝ𝜶 > 𝟓𝟎%
CV
(%
)
Time (min)
𝒇𝟐 = 𝟒𝟎
1
6
𝝈𝟐1
6
Mean variability
12
12
% r
ele
ased
12
12
EMA Q&A: BOOTSTRAP
[Answer:] (…) bootstrap methodology
could be used to derive confidence
intervals for f2 (…), (…) the preferred
method over f2 and MD. EMA/810713/2017
boot·strap /ˡbu:tstræp/ noun IDM pull /
drag yourself up by your (own)
bootstraps (informal) to improve your
situation yourself, without help from
other people
Oxford Advanced Learner‘s Dictionary of Current English
(2000). Oxford University PressRaspe (1785): Baron Munchausen's
Narrative of his Marvellous Travels
and Campaigns in Russia
3.08 4.92
BOOTSTRAPPING: PRINCIPLEUnknown population of cannon balls
5
4
1
Original data 𝒙 N=12
N=12
N=12N=12
7
6
2 32
6
43
1
7
6
5
41
3
2 𝑥∗2
𝑥∗𝐵2
6
6
3
4
7
7
ҧ𝑥 = 4.00
ҧ𝑥∗1 = 3.17
ҧ𝑥∗2 = 3.92 ҧ𝑥∗𝐵 = 4.00
𝑥∗1
Fre
qu
en
cy
Bootstrap distribution
𝒔𝒆𝑩 = 𝟎. 𝟓𝟔
sample
sample
sample
90%CI
B=10‘000
𝒔𝒆 =𝒔𝒅𝒙
𝑵= 𝟎. 𝟓𝟗
𝒔𝒆𝑩 bootstrap standard error
𝒔𝒆 standard error of the mean
5% 95%
3.173.92...
4.00
bootstrap replicates
bootstrap sample
BOOTSTRAPPING DISSOLUTION: CI
90% bootstrap CI(1)
Type Lower Upper
Normal 60.2217 66.9493
Percentile 59.8591 66.5973
Basic 60.5737 67.3119
BC 60.6481 67.4051
BCa 60.7555 67.5536
Bootstrap-t(2) 60.5844 67.8099(1)R (v3.6.1); B0=10‘000; (2)B1=1‘000; BC
bias-corrected; BCa bias-corrected and
accelerated.
𝒇𝟐∗𝑩 = 𝟔𝟏
𝒇𝟐∗𝟏 = 𝟔𝟒
𝒇𝟐∗𝟐 = 𝟔𝟎
1
2
B
𝒇𝟐∗𝒊
sample
sample
reference
test
Time (hours)
N=12
N=12
Similarity (...) be declared if the
CI for f2 (…) entirely above 50.EMA/810713/2017
12
12
12
12
12
12
𝒇𝟐𝒙 = 𝟔𝟒
BOOTSTRAPPING DISSOLUTION: ALL EASY?
90% bootstrap CI(1) (data I)
Type Lower Upper
Normal 75.3720 105.6605
Percentile 62.7877 092.9877
Basic 88.0448 118.2448
BC 87.4104 099.7090
BCa 87.4081 099.7060
Bootstrap-t(2) 87.9071 123.3391(1)R (v3.6.1); B0=10‘000; (2)B1=1‘000
90% bootstrap CI(1) (data II)
Type Lower Upper
Normal 49.0571 65.8814
Percentile 50.5090 67.4505
Basic 47.4880 64.4295
BC 50.4077 67.3106
BCa 49.9249 66.4951
Bootstrap-t(2) 47.5351 66.7653(1)R (v3.6.1); B0=10‘000; (2)B1=1‘000
BOOTSTRAPPING DISSOLUTION: ALPHA & POWER
Population 𝒇𝟐
Sim
ilari
ty (
%)
𝒇𝟐 point estimate
5% percentile
Low %CV in dissolution High %CV in dissolution
Population 𝒇𝟐
𝒇𝟐 point estimate
5% percentile
N=10‘000
B=1‘000
CV
(%
)
CV
(%
)
Time (min) Time (min)
𝒇𝟐 = 𝟏𝟎𝟎 𝒇𝟐 = 𝟏𝟎𝟎
1
6
1
6
ෝ𝜶 > 𝟓% ෝ𝜶 > 𝟓%
Sim
ilari
ty (
%)
[Major objection] Biowaiver (…) may not
be sufficiently supported (…). (…) 90%
confidence interval of the f2 similarity
factor calculated by bootstrapping
(…). The output should include at
least the number of bootstraps, the
5%, 50% (median) and 95% percentiles.
Evidence that the statistical software
has been validated should be
provided as well.
BOOTSTRAP IN PRACTICE: DL
Percentiles (P)
P Shah et al. DDSolver(1) RE (%)
5% 53.01 53.53147 −0.97
95% 68.34 68.72088 −0.55
B=1‘000; (1)Zhang et al. (2010)
reference
batch 1
Shah et al., 1998
Time (hours)
% r
ele
ased
Shah et al. (1998). Pharm Res 15(6): 889-896;
Zhang et al. (2010). AAPS J. 12(3): 263-267
input
output
software
BOOTSTRAP IN PRACTICE: ROUND 1
Percentiles
05% 86.9786
50% 89.3210
95% 91.5900
B=50‘000 in DDSolver
Percentiles
05% 50.6544
50% 52.8409
95% 55.7312
B=50‘000 in DDSolver
biobatch
waiver
Time (min)
% r
ele
as
ed
AGJ pH 1.2 Phosphate buffer pH 4.5 Phosphate buffer pH 6.8
0.00068 mg/mL 0.0013 mg/mL 0.11 mg/mL
Percentiles
05% 97.9487
50% 98.9720
95% 99.5367
B=50‘000 in DDSolver
biobatch
waiver
biobatch
waiver
[Response to draft] (…) to construct
more versions of 90% confidence
interval (…) to see if results regarding
similarity are sufficiently robust.
BOOTSTRAP IN PRACTICE: ROUND 290% bootstrap CI (Bootf2BCA)(1)
Type Lower Upper
Percentile 50.6599 55.7327
Bootstrap-t(2) 50.7184 56.6114
Normal 50.3144 55.4415
BCa 50.9257 56.3030
Basic 50.1124 55.1852(1)Mendyk et al., 2013; (2)in-house R-code,
R (v3.6.1); B0=50‘000; B1=1‘000
biobatch (1 tablet)
waiver (4 tablets)
Time (min)
% r
ele
ased biobatch (1 tablet)
waiver (1 tablet)
% r
ele
ased
Mendyk et al. (2013). Diss Technol 20(1): 13-17
Phosphate buffer pH 6.8
Dose 𝑨 = 𝑩
+0.5% Tween 20
[Response to draft] The assessor still finds
unclear what is difference between
normal approximation method and
basic bootstrap (…). The Applicant is
asked to fill in this information (…).
[Response to final] (…) the Applicant is
asked to provide (…) software
validation for individual approaches to
see that obtained results can be
considered as relevant.
BOOTSTRAP IN PRACTICE: ROUND 3 & 4
reference
test
Time (hours)%
rele
ased
𝜃 − 𝑏𝑅 ∓ 𝑣𝑅Τ1 2 ∙ 𝑧 1−𝛼
2 𝜃 − 𝜃∗ 1−𝛼 , 2 𝜃 − 𝜃∗ 𝛼
Davison & Hinkley (1997). Cambridge Press
Normal approximation
Basic (backwards)
SUMMARY & QUESTIONS
● f2: no control of type I error Statistics vs. clinical relevance? Signalsof problems available?
● Bootstrap f2: confidence interval … but
harder to pass similarity criterion
Type I error control? Which CI?
Comparison to other methods?
Applicable generally (EMA Q&A)?
● Equivalence margin: average 10% Current margin already conservative?Maximum 10% (EMA Q&A) or average10%?
● Dissolution in non-sink conditions API or product performance?
> library(fortunes)
> fortunes::fortune(222)
Some people familiar with R describe
it as a supercharged version of
Microsoft's Excel spreadsheet
software.
-- Ashlee Vance (in his article "Data Analysts Captivated
by R's Power") The New York Times (January 2009)
> library(fortunes)
> fortunes::fortune(358)
The existence of a method is not a
sufficient reason to use that method.
-- Jari Oksanen (about relative advantages of several
multivariate analysis methods) R-SIG-Ecology
(November 2013)
QUOTES: LIBRARY(FORTUNES)
R Core Team (2019). https://www.R-project.org/