Moving from conventional to rapid microbiology
update Merck June 2016
Rapid Detection of Micro-organisms by Direct Detection: A
Solution for Time Critical In-process Monitoring
Geert Verdonk
Pieta IJzerman-Boon
Edwin van den Heuvel
June 2016
Why Rapid Methods at Merck?
Application & Benefits
► In process controls (PAT concept):
• Disaster prevention
• Process monitoring
► Perfect tool for root cause investigations:
• Corrective actions on ‘real-time’ data
► Improving lab efficiency
• Automation
► Reduction of cycle time for product release
Cost avoidance
Reduction of waste
Improving efficiency
2
Rapid Method technologies
Basic principles
1. Growth based technologies

BacT/ALERT: CO2 colorimetry

AKuScreen: ATP bioluminescence

Milliflex Rapid: ATP bioluminescence
2. Direct measurement

Solid phase flow cytometry

Flow cytometry
3. Detection specific cell components

Q-PCR: Nucleic acid detection

MicroSEQ: Nucleic acid detection

Fatty acid analysis

Antibodies
3
Rapid Methods implementation strategy
at Merck
1.
2.
Focus on production process

Select critical control point

Investigate requirement regarding analysis time
Select suitable techniques

Dedicated to a process & sample type
3.
Pilot testing RMM
4.
Selection of RMM system
5.
Qualification & Validation process
(in close cooperation with statistics)
6.
Parallel testing: ‘real’ samples
7.
RMM Implementation in routine analysis
4
Biotechnology Production process
5
Problem
Columns can get contaminated
► Problem: current sanitization is aggressive for the column and
does not allow a long life cycle ($) , however it is needed to
prevent microbiological growth in the column.
6
Solution
Less stringent Sanitization?
► Solution: Use a light program with 0.1 M Sodium Hydroxide and
use the heavy sanitization based on an alert/action limit strategy
7
Needed
Rapid bioburden in combination with a
decision tree for sanitization buffers
► An on line/at line bioburden screening method that allows to
make decisions how to sanitize the purification columns
• If the rapid bioburden shows that you are in control:
continue with standard sanitization buffers allowing a
longer life cycle of the columns
• If the rapid bioburden shows alert/action levels:
switch to a more stringent sanitization buffer
8
Selection of technology
► A same day detection leads you to a non growth based method:
Direct detection of microbiological cells
► Platforms available: Solid phase and fluid phase detection
technology : multiple suppliers are available
► Specific requirements:
• At line detection: simple to operate
• Time to results for one sample: 1 hour
• Capability to measure between 0-100 cfu/10 mL. Higher volumes nice to have
• Low throughput allowed because of niche application
9
Selection led to the
MuScan technology
Filtration
Simple sample
preparation
Staining
Large volume filtration
possible on “perfect”
filters
10
Read Out
Viability staining followed
by high resolution image
analysis
Strategy
► Proof of concept studies: can we use the technology for the
requested application
► Prevalidation phase
► Validation using the PDA and Pharmacopeia guidance
11
FFU versus CFU
► CFU = Colony Forming Units
► FFU = Fluorescence Forming Units
► When trending FFUs in
monitoring a production
process: Should you see
the same trend compared
to trending CFUs?
► The actual comparisons of
numbers in fact does not
matter for this specific
purpose…...
12
Signal >>>>
Hypothesis for column control
FFU
► Situation in control
CFU
Time >>>>
FFU
Signal >>>>
► Situation needs
extra sanitization
CFU
1.
Define baseline for
technology
2.
Define alert limits and
action limits
3.
Can we compare the
technologies?
Time >>>>
13
Determination bioburden criterion
Perfect test method
► Suppose we test one sample with a perfect test method
► Then the variation in the observed counts comes from the
sampling, not from the method
► These counts typically follow a Poisson distribution with a certain
mean and variance both equal to 𝜆
► The parameter 𝜆 represents the average density or number of
microorganisms per test sample
14
Determination bioburden criterion
Perfect test method
► Poisson distributions for different values of mean parameter 𝜆
0
15
Determination bioburden criterion
Perfect test method
► The observed count may be 0 even if 𝜆 is not
P(Y=0) ≤ 5%
for 𝜆 ≥ 3.0
► So even with a perfect test and the most stringent criterion (y=0),
we must accept that we sometimes fail to detect a contamination
16
Determination bioburden criterion
► We should have a criterion (say y ≥ AL) such that for large
densities we decide to sanitize with high probability, and for
‘small densities’ we sanitize with small probability
• Not knowing what ‘small’ enough is, let’s require that for blanks (𝜆=0) this
𝛼-error probability is ≤5% (AL=alert limit) or ≤1% (AL=action limit)
1.0
0.8
0.6
0.8
AL
0.4
0.6
P(Y≥AL) ≤ 5 or 1%
High probability that
buffers with ‘small
densities’ will be
found acceptable.
High probability that
contaminated
buffers will be found
unacceptable.
0.4
0.2
P (Y < AL , i.e. accept buffer
sanitization)
extra
withoutProb.
of Acceptance
(Pa)
𝛼=P(Y≥AL) ≤ 5%
Blank buffer 𝜆 = 0
1.0
0.2
0.0
β=P(Y<AL) ≤ 5% 0.0
0
2
0
4
6
17 Defective
Percent
17
8
→ Average density 𝜆
10
Determination bioburden criterion
Experiments
► In order to follow our strategy for determining alert/action limits,
we need to characterize three things
• Distribution of observed counts for blank samples (baseline signal)
– Variability within and between runs
Experiment
blank samples
– Does it depend on the buffer or increase with sample volume?
• The distribution of counts along the density-response relation
Experiment
– Does the method add variability to the Poisson variability of the spike?
spiked samples
– Is variability of MuScan larger than that of the compendial method?
– Does the method variability depend on the spike level?
Dilution
• The shape (e.g. intercept, slope) of the density-response relation
experiment
– Is the count proportional to the spike level for both methods?
– Can we distinguish contaminated samples from blank samples?
18
Experiment blank samples
Design
► 3 different buffers (WFI, 10 mM NaH2PO4, 100 mM acetic acid)
► 2 volumes (10, 100 mL)
► 9 days
► 4 samples per buffer, volume, and day
► 1 stock solution for each buffer was filled out over 500 mL bottles
Each day, 1 bottle was used
19
Experiment blank samples
Model
► 𝑌𝑖𝑗𝑘𝑙 count for buffer 𝑖, volume 𝑗, run 𝑘, replicate 𝑙
► 𝑌𝑖𝑗𝑘𝑙 ~ 𝑃𝑜𝑖𝑠𝑠𝑜𝑛(𝜆𝑖𝑗𝑘 ) with mean 𝜆𝑖𝑗𝑘
• Instead of a Poisson distribution, data may follow a negative binomial
distribution, which allows more variability
► The mean count may depend on fixed effects for buffer,
volume or a combination of both, and possibly an additional
random effect of run, either or not depending on buffer
log 𝜆𝑖𝑗𝑘 = 𝜇 + buffer𝑖 + volume𝑗 + (buf*vol)𝑖𝑗 + run𝑘
with run𝑘 ~ 𝑁(0, 𝜎 2 ) or run𝑘 ~ 𝑁(0, 𝜎𝑖2 )
20
Experiment blank samples
Results
► MuScan produces raw data and processed data
► Plots show raw data, including fluorescent staining artefacts:
21
Experiment blank samples
Results
► Best model was negative binomial model with a single run effect
► Count was dependent on buffer, but not on volume
► We calculated the baseline level and potential alert and action
limits above which it is unlikely (probability<5% or <1%) that
samples come from the predicted distribution of blank samples
Observed or predicted %
NaH2PO4, raw data
𝜆 = 1.26 baseline
P(Y≥6) ≤ 5% alert
P(Y≥10) ≤1% action
NaH2PO4, processed data
𝜆 = 0.62 baseline
P(Y≥4) ≤ 5% alert
P(Y≥5) ≤1% action
Experiment spiked samples
Design
NaH2PO4
Acetic acid
PBS
6 blank
samples
100 mL
6 blank
samples
100 mL
2 blank
samples
100 mL
BioBall®
± 550 CFU
20 μL
Test with MuScan
Similar experiment as previous
(different buffers, 9 days)
but only one sample volume,
and compare with compendial
Test with compendial
Rehydration
fluid 1.1 mL
2 blank
samples
Spiked samples
~10 CFU/100 mL
Repeat this 3x for acetic acid, 3x for NaH2PO4, 1x for PBS
Experiment spiked samples
Model
► Similar model as for blank samples, but now with method
and buffer instead of buffer and volume
► Run variability may incorporate correlation between methods
• Correlation may be expected since run variability is caused by
variability in the spike level, and both methods tested samples from the
same solution
24
Experiment spiked samples
Results
NaH2PO4, E. coli, MuScan raw data
NaH2PO4, E. coli, processed data
𝜆 = 8.30
Observed or predicted %
𝜆 = 8.96
Recall limits
5% alert, 1% action
5%, 1%
59%
12%
≤5%
≤5%
25
Experiment spiked samples
Results
NaH2PO4 buffer, E. coli, compendial
Observed or predicted %
𝜆 = 9.15
Assume
AL=1
≤5%
26
Experiment spiked samples
Results
► For MuScan technology there is less room than for the
compendial method to set an alert limit that distinguishes blanks
from samples with ~10 CFU in one 100 mL sample
► However, we can move the distribution of the spiked samples
away from the blank sample distribution, by increasing the
sample volume, e.g. testing 200 mL
27
Conclusions
► Using advanced statistical methods, we can characterize the
performance of the MuScan technology
► Based on this, we can determine a sensible monitoring strategy
using the MuScan technology (FFUs), independent of the
compendial method (CFUs)
28
Determination bioburden criterion
Baseline distribution for a
perfect test method
𝛼=0
𝛼≤5%
β≤5%
29
Determination bioburden criterion
Perfect test method
Observed count y
15
12
9
6
The line y=𝜆
models averages,
but there is variation
1. AL=1:
𝛼-error=P(Y≥AL)=0 ≤ 5% for 𝜆 = 0
β−error=P(Y<AL) ≤ 5% for 𝜆 ≥ 3.0
2. AL=4:
β−error=P(Y<AL) ≤ 5% for 𝜆 ≥ 8
2.
3
AL=4
1.
0
AL=1
0
3
6
𝜆 9
12
15
Average density 𝜆
Determination bioburden criterion
► If method is not perfect, there may be
• False positives: Intercept larger than 0
Observed count y
15
12
9
6
• False negatives: Slope smaller than 1
• Additional variability between runs
► These things all make it more difficult
to distinguish 𝜆>0 from baseline
► Note that also the compendial test is
3
not perfect
0
2.
0
3
6
9
12
15
Average density 𝜆