BILS 2015 TU Wien exputec BioVT Pharmaceutical Engineering
"Quicker Ways to a Scalable Process Based on Sound-Science Based Methodologies"
Christoph Herwig
3. 18/04/16 Ch. Herwig
Status quo of bioprocess design
The Scaling Tasks
Investigate! Redo! Hope!Stomach decision !
Process
Development
Piloting ManufacturingScreening
Scale-up Scale-up Scale-up
Productivity
Waste
Process Development Time
Revenue
Period
3
4. 18/04/16 Ch. Herwig
Challenges that can be solved by
faster process
development and
higher titers
prevent and
identify scale-up
effects
prevent failed
batches
4
5. 18/04/16 Ch. Herwig
How can these challenges be
addressed?
โขโฏ Clear and proven workflows for data
analysis
โขโฏ Tools to process data and ensure data
quality
โขโฏ Efficient use of chemometric and
mechanistic data science tools
โขโฏ10
6. 18/04/16 Ch. Herwig
Presentation Workflow
โขโฏ Goal: Quicker Ways to a Scalable Process Based on Sound-Science
Based Methodologies
โขโฏ Processing Strategies to Understand the Production Platform and to
Allow Transferability from Scale to Scale and from Product to
Product โ
6
Conclusions
Model
Building
Tutorial
Linking
MVDA
results to
metabolic
models
Information
extraction
and
Statistical
analysis
Big Data
Processing
9. 18/04/16 Ch. Herwig
Data import, data alignment and
contextualization
โขโฏ USP & DSP automatic data import
โขโฏ Excel spreadsheet, text documents, LIMS,
PIMS
โขโฏ Data contextualization
โขโฏ measurement units
โขโฏ campaign, phase definition
โขโฏ operators
โขโฏ Data survey & overview
โขโฏ overview plots
โขโฏ data density plots
ร โฏ All data in one format that can easily be analyzed and explored
Exputec Software
Excel
LIMS
PIMS
9
10. 18/04/16 Ch. Herwig
Principal Component Analysis:
Raw Data
โขโฏ Principal component analysis on individual runs to quantify
variations and detect relationships among variables.
10
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Energieeintrag (HK)
Druck
S3 Menge
Dosierung 2
Summe Substrat
Bio Wรคrmeleistung
S2
Product
S1
Nebenprodukt
Component 1
Component2
B169012
1 2 3
0
10
20
30
40
50
60
70
80
90
100
Principal Component
VarianceExplained(%)
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
12. 18/04/16 Ch. Herwig
Information extraction โ
Completion of data sets
โขโฏ Tools to derive meaningful descriptors
โขโฏ Control quality tools
โขโฏ Descriptors that describe the control quality of
process parameters, such as pH, pO2
โขโฏ RMSE, outlier detection
โขโฏ Bioprocess suite deriving scalable descriptors
โขโฏ ยต, qs, CER, OUR,
โขโฏ Yields
โขโฏ kinetic constants
โขโฏ Calculation of missing entities via combination of
dvariables and sound science first principles
ร โฏ Automatic extraction of meaninful descriptors for different
process phases, USP & DSP processes
0 5 10 15 20 25
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
feed time (h)
myreccalc(1/h)
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Yxsreccalc(c-mol/-cmol)
12
14. 18/04/16 Ch. Herwig
Gathered data to statistical
information
14
โขโฏ Raw data from DoE experiments transferred into information
โขโฏ Influence of pH, pCO2 and pO2 on specific rates and yields
VCD
Glucose
SPECIFICRATEYIELD
15. 18/04/16 Ch. Herwig
Hypothesis Generation using
MVIA Results
โขโฏ Processing of data (concentrations, flows) into
specific rates and yield coefficients (ยต, qs)
โขโฏ Tools
โขโฏ Principal Component Analysis
โขโฏ MLR
โขโฏ PCR
โขโฏ Factor Analysis
โขโฏ Multi-way methods
โขโฏ โฆ
High Performer Low Performer
ร โฏIdentify and understand trends and correlations
ร โฏIdentify interactions across unit operations
18. 18/04/16 Ch. Herwig
Segregation of Biomass
-Lysis-
18
Bio ma ss
l y sis
ph y sio l o g y
m a mm a l ia
0 500 1000 1500 2000 2500
-10
-8
-6
-4
-2
0
2
rdLDH(mU/(mL*h))
Initial LDH (mU/mL)
0 2000 4000 6000 8000 10000
-70
-60
-50
-40
-30
-20
-10
0
10
rdDNA(ng/(ml*h))
Initial DNA (ng/mL)
a) b)
0
10000
20000
30000
40000
50000
rpLDH(ยตU(mL*h))
rmeasLDH
rcorrLDH
d)
0 50 100 150 200 250
time (h)
0
50000
100000
150000
200000
250000
rpDNA(pg/(mL*h))
rmeasDNA
rcorrDNA
c)
0 50 100 150 200 250
time (h)
Degradation rates of a) DNA and b) LDH in fermentation supernatants with respect to initially detected extracellular concentrations of DNA or
LDH. Volumetric rates of c) DNA release into culture supernatants and d) LDH release into culture supernatants from direct measurements of
the respective marker (rmeas) and corrected for marker degradation (rcorr).
19. 18/04/16 Ch. Herwig
Segregation of Biomass
-Lysis-
19
Bio ma ss
l y sis
ph y sio l o g y
m a mm a l ia
b)
0 25 50 75 100 125 150 175
0.00
2.50E6
5.00E6
7.50E6
1.00E7
1.25E7
1.50E7
1.75E7
Cells/mL
time (h)
VCC
TCC
LCC
d)
0 50 100 150 200 250
0.00
2.50E6
5.00E6
7.50E6
1.00E7
1.25E7
1.50E7
Cells/mL
time (h)
VCC
TCC
LCC
c)
0 50 100 150 200 250
0.0
5.0E6
1.0E7
1.5E7
2.0E7
2.5E7
Cells/mL
time (h)
VCC
TCC
LCC
a)
0 50 100 150 200 250
0
2
4
6
8
10
12
Glc(g/L)
time (h)
0
1
2
3
Gln,Lac(g/L)
Glucose
Lactate
Glutamine
4
pH, 1
pH, 2 pH, 3
pH, 1
20. 18/04/16 Ch. Herwig
Segregation of Biomass
-Lysis-
20
Bio ma ss
l y sislysis โ results
Lysis has a influence on the physiological interpretation!
lysis should be considered
ph y sio l o g y
m a mm a l ia
b)
0 50 100 150 200 250
-0.02
0.00
0.02
0.04
ยต(1/h)
time (h)
ยตVCC
ยตLCC
21. 18/04/16 Ch. Herwig
LINK THE HINT TO
MECHANISTICS
Linking Cluster Analysis to Metabolic Models
21
22. 18/04/16 Ch. Herwig
3 4 5 6 7 8 9
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Process Time [h]
Variables
Data Analysis to understand lactate production
and consumption in Cell Culture
Prediction of qLac days by a
PLS-R model
Prediction itself not very useful,
since it is very time consuming
to measure all 66 variables
However, the weights of the
PLS-R model can be used to
find mechanistic links
11 Batches from a DoE
66 variables (Specific rates, MFA-results etc.)
8 Samples each
-2 0 2 4 6 8 10
-2
-1
0
1
2
3
4
5
6
7
8
Actual
Predicted
1 2 3 4
0
10
20
30
40
50
60
70
80
90
100
Number of PLS components
PercentVarianceExplainediny
22
23. 18/04/16 Ch. Herwig
0 2 4 6 8 10
0
1
2
3
4
Process Time [d]
0 2 4 6 8 10
0
1
2
3
4
Process Time [d]
0 2 4 6 8 10
0
1
2
3
4
Process Time [d]
Cluster 1
Cluster 2
Cluster 3
Cluster 4
PLS-R variable importance (VIP)
for qLac
Cluster detection (k-means cluster analysis) based
on PLS-R VIP to detect main correlations with
response (e.g.: qLac) ร identify mechanistics
VIP > 1 variable is significant
VIP < 1 variable is insignificant
Phase detection by k-means cluster analysis (blues
lines)
Average VIPs
23
24. 18/04/16 Ch. Herwig
Use simplified metabolic flux
models
โขโฏ Only the Central Carbon
Metabolism is considered
โขโฏ Redox and energy metabolism
โขโฏ Amino acid metabolism
โขโฏ But which measurements are now
really omportantn in which phase?
24
25. 18/04/16 Ch. Herwig
Glc GPI ald GAPDH pgk eno pepkin g6Pdh prodh
1
2
3.1
4.1
5.2
6.2
7.3
8.3
Variables
ProcessTime[d]
0
0.5
1
1.5
2
Ala Arg Asp Asn Glu Gln Gly His Ile Leu Lys Phe Pro Trp Tyr Val mu qp Formate NH4 Lip Anap PPP_nox gpt asp-got aspg gls leu cat ile cat shmt phe deg serc3 mehf sink
1
2
3.1
4.1
5.2
6.2
7.3
8.3
Variables
ProcessTime[d]
Met Ser Thr Cit Pyr Suc pdh cisac icdh akgdh succdh fum maldh val cat thr deg Resp
1
2
3.1
4.1
5.2
6.2
7.3
8.3
Variables
ProcessTime[d]
0
0.5
1
1.5
2
arg cat glud his deg lys deg met deg Trp deg Tyr cat NH3 sink
1
2
3.1
4.1
5.2
6.2
7.3
8.3
Variables
ProcessTime[d]
0
0.5
1
1.5
2
Clusters variable importance(VIP) for
Lactate production/consumption
Cluster 1: Important variables for qLac for all time points; mainly related to glycolysis ร Overflow / Lactate
Cluster 3: Important variables for qLac at the early phase (plus late phase); mainly related to TCA cyle activity
Cluster 4: Important variables for qLac at the late time points; probably related to nutrient limitation / stationary phase
Red: important to predict qLac; a value > 1 means the variable is significant
Blue: not important to predict qLac; a value < 1 means the variable is insignificant
Cluster 2: Less important variables for qLac
25
26. 18/04/16 Ch. Herwig
Mechanistics insights by target
oriented use of statistical tools
โขโฏ Highly target-oriented analysis of large data sets by combination of
different data driven methods
โขโฏ Do not get lost in the woods!
โขโฏ Automated clustering of process
variables in to physiologically
meaningful groups with regard
to the response variable
(e.g.: qLac)
โขโฏ Mechanistic insight can be
acquired from data driven
methods if the tools are
applied appropriately
26
28. 18/04/16 Ch. Herwig
The Modelling Work Flow
Structure
Identification
Parameter
Estimation
Model Validation
Model
Refinement
Model Use
YES
NO
Experimental
Data
Priori
Knowledge
28
29. 18/04/16 Ch. Herwig
Mechanistic Model for
Glucose Uptake Kinetics
29
ยงโฏ Propose hypothesis:
โ ๐โ๐บ๐๐โ= ๐( ๐บ๐๐)
ยงโฏ Formulate equation:
โ ๐โ๐บ๐๐โ=โ ๐โ๐ฎ๐๐, ๐๐๐โโโ[ ๐บ๐๐]/
[๐บ๐๐]+โ ๐ฒโ๐ฎ๐๐โโ
ยงโฏ Estimate parameters
individually by minimizing cost
function โโโ(โ
๐ฆโ๐, ๐๐๐๐ ๐ข๐๐๐โโโ
๐ฆโ๐, ๐๐๐๐๐โ)ยฒโ
ยงโฏ Compare with literature
30. 18/04/16 Ch. Herwig
Toolset for model calibration
โขโฏ Model calibration:
โขโฏ finding suitable estimates for model
parameters
โขโฏ Sensitivity and identifiability analysis:
โขโฏ Link to your measurements and
processing platform
30
31. 18/04/16 Ch. Herwig
Mechanistic optimization
โขโฏ Use a kinetic model for model-based
optimization
โขโฏ Development of mechanistic process
model
โขโฏ In-silico optimization using optimization
algorithms
โขโฏ Identification of optimal operating
conditions
โขโฏ Control Implementation
ร higher productivity using less experiments
ร Knowledge on mechanistic relationships can
be transferred to the next product
31
36. 18/04/16 Ch. Herwig
Take Home
Workflow
36
โขโฏ Bioprocess development and manufacturing challenges can be
solved by process data science
โขโฏ Proven workflows and a tailored toolset for bioprocess data
analysis and process optimization is necessary
Tranfer
knowledge to
other
process /
product / site
Optimize the
process by
mechanistic
models
Link your hint
to
mechanistic
hypotheses
Get a hint
using
information
extraction and
statiscial tools
Get a grab on
your data by
tailored
workflows
37. 18/04/16 Ch. Herwig
Investigate! Redo! Hope!Stomach decision !
Process
Development
Piloting ManufacturingScreening
Scale-up Scale-up Scale-up
Productivity
Waste
Process Development Time
Revenue
Period
Benefits through
shown Approach
37
Process
Development
Pilo?ng Manufacturing Screening
Scale-up
Increase produc1vity
by sustained op1miza1on, scalability and
elimina1on of fail batches
Cut Process Development Time by 50% Revenue Period
Inves?gate! Verify! Control & predict! Explore!
Scale-up Scale-up
39. 18/04/16 Ch. Herwig
Thank you
for your attention!
Univ.Prof. Dr. Christoph Herwig
Vienna University of Technology
Institute of Chemical Engineering
Research Division Biochemical Engineering
Gumpendorferstrasse 1a/ 166 - 4
A-1060 Wien
Austria
emailto: christoph.herwig@tuwien.ac.at
Tel (Office): +43 1 58801 166400
Tel (Mobile): +43 676 47 37 217
Fax: +43 1 58801 166980
URL : http://institute.tuwien.ac.at/chemical_engineering/bioprocess_engineering/EN/
https://www.facebook.com/BioVTatTUWien
39
42. 18/04/16 Ch. Herwig
Computational environment for real-
time implementation and control
โขโฏ Object oriented design of
different data storage and
processing components.
โขโฏ Classes are able to store
data, perform specific
functions, and communicate
with each other
42
Process Information
Management System
OPC server
Calculations:
feeding rates
offgas analysis
volume calculation
outlier detection
โฆ
Observer
Bioprocess object
yu(t0)
y
Model-predictive
controller & PID
u(t0)
xpred
u(t1)
y: measurements/outputs (e.g. offgas)
u: system inputs (e.g. feeding rates)
x: system states (e.g. concentrations)
43. 18/04/16 Ch. Herwig
Calculation of respiratory rates and
outlier detection
43
Bioprocess
Calculator offgas
Calculator outlier
Observer
FAIR,FO2,O2_offgas,CO2_offgas
OUR, CER, RQ
OUR, CER outliers removed
7.3565 7.3566 7.3566
x 10
5
0
0.1
0.2
0.3
0.4
Original signal
New signal 7.3565 7.3566 7.3566
x 10
5
0
0.5
1
G
Grubbs distance
7.3565 7.3566 7.3566
x 10
5
0
0.02
0.04
0.06
stdandard deviation
7.3565 7.3566 7.3566
x 10
5
0
0.1
0.2
0.3
0.4 New signal
48. Get There Faster.
Exputecโs Software Solutions by INCYGHT
ยงโฏ Ef๏ฌcient implementation of our solutions at your site using
Exputec INCYGHT software
Batch 1 Batch 2 Batch N
Campaign analysis
. . .
MVIA
(Multivariate
Information
Analysis) for
extraction of
mechanistic
information out of
procee data
M-DoE
(Mechanistic
Design of
Experiments) for
reduced number
of experiments.
Multivariate
statistical methods
for exploratory
data analysis:
hypothesis
generation and
testing
Proven workflows
for identification of
sources of
variability,
increasing process
robustness, and
optimization.
48
49. 18/04/16 Ch. Herwig
a) Tradi?onal feed pro๏ฌle design b) Physiological feed pro๏ฌle design
Accelerate!
Wechselberger et. al. 2012
Speed up by using
physiological information in DoEs!
รโฏ Careful selection of physiological factors for the DoE
significantly reduces number of experiments
49
50. 18/04/16 Ch. Herwig
รโฏ Dynamic experiments increase information & throughput
Accelerate!
a) Quick iden?๏ฌca?on of scalable
rela?onships
b) Dynamic feeding pro๏ฌles based
on speci๏ฌc substrate uptake rate
Zalai et. al. 2012
Speed up by using physiological
information & dynamics!
50
52. 18/04/16 Ch. Herwig
Segregation of Biomass
-Lysis-
52
Bio ma ss
l y sisCell Lysis
deIinition:
โloss of integrity of a cell (destruction of the cell membrane)โ
motivation:
lysed cells were once "produced" and can be included in the description of
the growth kinetics
lysed cells could serve as a nutrient source
an inIluence on the product quality can not be excluded
measurement methods:
classiIication over intracellular substances in the supernatant
mammalians:
โขโฏ LDH measurements in the supernatant
โขโฏ DNA measurements in the supernatant
microbials
โขโฏ C-balance
ph y sio l o g y
m a mm a l iaM ic r o bia l s
53. 18/04/16 Ch. Herwig
Quality by Design
Process Parameters
Temperature
Stirrer Speed
Dissolved Oxygen
pH
Air Flow
Pressure
Feedrate
Nutrient concentrations
Inducer concentration
Biomass concentration
Induction Time
Conductivity
Redox level
Strain
Expression cassette
โฆ
Product quality attributes
Enzyme activity
Titer
Purity
Stability
Batch-to-batch variability
Efficiency
Cost of product
Space-time-yield
Protein folding
Glycosylation pattern
Viability
Ease of further processing
(downstream)
Potential risks for end-user
โฆ
???
18/04/16 53
54. 18/04/16 Ch. Herwig
Process Parameters
Temperature
Stirrer Speed
Dissolved Oxygen
pH
Air Flow
Pressure
Feedrate
Nutrient concentrations
Inducer concentration
Biomass concentration
Induction Time
Conductivity
Redox level
Strain
Expression cassette
โฆ
Product quality attributes
Enzyme activity
Titer
Purity
Stability
Batch-to-batch variability
Efficiency
Cost of product
Space-time-yield
Protein folding
Glycosylation pattern
Viability
Ease of further processing
(downstream)
Potential risks for end-user
โฆ
Quality by Design
???
54
55. 18/04/16 Ch. Herwig
Current interpretation of QbD:
Cooking Recipe: DoE
Speci๏ฌc Ac?vity
[kU/gbiomass]
Induc?on Phase
Temperature
[ยฐC]
Induc?on Phase
Feeding
Exponent k
โDesign Spaceโ
Data
CPPs
CQA
CQA
55
CPPs
56. 18/04/16 Ch. Herwig
DOE: Insignificant factors-
all for nothing?
โขโฏ Factors turn out to be insignificant
โขโฏ Variance cannot be explained by the
original factors
โขโฏ Possible reasons:
โขโฏ 1) wrong factors were chosen for
investigation
โขโฏ 2) noise on experiment higher than effects
of factors
โขโฏ How to proceed?
ร Go beyond MLR analysis for the efficient
exploitation of DoEs!
-0,10
-0,05
-0,00
0,05
x_EFB
ยต_FB
maxspectitersupg/g
N=9 R2=0,604 RSD=0,002799
DF=6 Q2=-0,347 Conf. lev.=0,95
-0,10
0,00
0,10
x_EFB
maxspectiterpelletg/g
N=9 R2=0,0
DF=6 Q2=-1,
Factor
Transformation
MVDA
Knowledge
56
57. 18/04/16 Ch. Herwig
-0,10
-0,05
-0,00
0,05
x_EFB
ยต_FB
maxspectitersupg/g
N=9 R2=0,604 RSD=0,002799
DF=6 Q2=-0,347 Conf. lev.=0,95
-0,10
0,00
0,10
maxspectiterpelletg/g
Run DoE
MLR analysis on
original factors
Data Processing,
Analyze more
descriptors of the
process
MVDA
Exploratory analysis
Hypythesis generation:
New factors
Sort out co-linear
factors (e.g.variance
inflation factor)
MLR analysis on
transformed factors
Recycling of DoE Results
57
-1,5
-1,0
-0,5
0,0 ยต
maxspectitersupg/g
N=9 R2=0,510 RSD=0,002881
DF=7 Q2=0,313 Conf. lev.=0,95
-1,5
-1,0
-0,5
0,0
ยต
maxspectiterpelletg/g
N=9 R2=0,335 RSD=0,003226
DF=7 Q2=0,047 Conf. lev.=0,95
58. 18/04/16 Ch. Herwig
Linking metabolite production and
substrate uptake
58
Hypothesis:
overflow metabolism is
coupled to high glucose
consumption rate
a) Threshold
๐โ๐๐๐ โ๐๐๐โโ0.11 [โ ๐ ๐๐๐/โ
10โ9โ๐๐๐๐๐ โโโ]
b) Linear equation
๐๐๐ โ ๐โ๐๐๐โ< ๐โ๐๐๐ โ๐๐๐
โ ๐โ๐๐๐โ=โ0.25โ2.21โโ ๐โ๐๐๐โ
๐ ๐๐๐ธ = 0.028 [โ ๐ ๐๐๐/โ10โ9โ
โ ๐๐๐๐๐ โโโ]
59. 18/04/16 Ch. Herwig
Vom Labor in den Prozess
2x
@50%
Toolset Integration
Data- Knowledge
Management
Process
Understanding
59
60. 18/04/16 Ch. Herwig
Sensitivity of model outputs to
variations in one model parameter
40 datasets were simulated by applying ยฑ2% variations on a process value
60
0 100 200
0
10
20
30
40
cXLtot
Time [h]
0 100 200
0
0.5
1
cS1
Time [h]
0 100 200
0
5
10
cPEN
Time [h]
0 100 200
0
2
4
6
cPOX
Time [h]
0 100 200
-0.2
-0.15
-0.1
-0.05
0
OUR
Time [h]
0 100 200
0
1
2
3
4
5
cA0
Time [h]
0 100 200
0
10
20
30
40
cA1
Time [h]
0 100 200
0
2
4
6
8
10
cS2
Time [h]
0 100 200
0
0.05
0.1
0.15
0.2
CER
Time [h]
0 100 200
0
0.05
0.1
0.15
0.2
rX
Time [h]
0 100 200
-0.25
-0.2
-0.15
-0.1
-0.05
rS1
Time [h]
0 100 200
-0.3
-0.2
-0.1
0
0.1
0.2
rS2
Time [h]
61. 18/04/16 Ch. Herwig
Identifiability of parameter
subsets
โขโฏ The colinearity index measures the degree of linear
dependence of model parameters.
โขโฏ It equals unity if the columns are linearly dependent.
โขโฏ If it exceeds 10-15, then the corresponding parameter
subset is poorly identifiable.
61
0 1 2 3 4 5 6 7
0
5
10
15
20
25
30
35
40
45
50
minimumofcollinearityindex
size of a parameter sets
Parameter Unit
Optimized
Value
Literature
Value Reference
๐ ๐ณ๐๐
๐ฎ๐๐
๐๐๐๐/๐๐๐๐ โ1,11 โ1,10 (Lee et al., 2003)
๐ ๐ฎ๐๐,๐๐๐.๐๐๐ ๐๐ด ๐ฎ๐๐ ๐๐๐๐/(๐๐ โ ๐๐๐๐ โ โ) โ1,99 โ 10โ11
โ4 โ 10โ11 (Lee et al., 2003)
๐ฒ ๐ฎ๐๐ ๐๐ 8,75 2,25 (Aehle et al., 2012)
๐ ๐ณ๐๐,๐ช๐๐๐๐๐๐๐๐๐๐ ๐๐๐๐/(๐๐๐๐ โ โ) โ1,80 โ 10โ10
๐ ๐ฎ๐๐,๐๐๐ ๐๐๐๐/(๐๐๐๐ โ โ) โ3,37 โ 10โ10
โ1,8 โ 10โ10 (Aehle et al., 2012)
ยต ๐๐๐ โโ1 0,024 0,035 (Craven et al., 2013)
๐ฒยต,๐ฎ๐๐ ๐๐ 3,81 4,8 (Dhir et al., 2000)
ยต ๐ ๐๐๐๐,๐๐๐ โโ1 0,015 0,019 (Dhir et al., 2000)
๐ฒ๐,๐ฎ๐๐ ๐๐ 1,58
ยต ๐ ๐๐๐๐,๐๐๐๐๐ โโ1 0,0015 0,00266 (Borchers et al., 2013)
๐ฒ๐๐๐๐๐ โโ1 0,004 0,04 (Craven et al., 2013)
๐ ๐ต๐ฏ ๐
+
/๐ฎ๐๐ ๐๐๐๐/๐๐๐๐ โ0,54 โ0,68 (Craven et al., 2013)