Applications of Multivariate Data Analysis

Presented by Petter Mree & Jonas Elfving

Copyright 2013 OSIsoft, LLC.

Novartis values from MVDA for PAT & QbD

Increase of process understanding
! ! ! ! Identification of influential process parameters Identification of correlation pattern among the process parameters Generation of process signatures Relationship between process parameters and quality attributes

Increase of process control

! ! ! ! Enabling on-line early fault detection Support for time resolved design space verification (real time quality assurance) Predicting quality attributes based on process data Excellent tool for root cause, trending analysis and visualization

Business Challenge
! -! -! Reduction of dimentionality ! Conformity check Real time release testing Trend analysis Root cause analysis

Solution
! ! ! !

Results and Benefits

Ability to be proactive rather than reactive to variation or poor quality. Saving batches Reducing OOS Helping to optimize the process
Copyright 2013 OSIsoft, LLC.

Topics
! Introduction: MVDA in the context of pharmaceutical production ! Case studies for MVDA I
! Process monitoring of a granulation process in pharmaceutical production

! Case studies for MVDA II

! Statistical Process Control Biopharmaceutical Production for optimization

! Short real-time demonstration

Using PI Server, PI Batch and PI Event Frame & PI Interface for SIMCA-online

Copyright 2013 OSIsoft, LLC.

Umetrics
! ! ! ! ! ! ! ! ! Part of ~1Billion conglomerate The market leader in software for multivariate analysis (MVDA) & Design of Experiments (DOE) 25+ years in the market Off line analysis tools On-Line process monitoring and fault detection 700+ companies, 7,000+ users Pharmaceutical, Biotech, Chemical, Food, Semiconductors and more Worldwide Presence with MKS Close collaboration with universities in USA, Sweden, UK and Canada
Copyright 2013 OSIsoft, LLC.

 Copyright 2013 OSIsoft, LLC.

Building a capable process

DOE
Control Plan
DOE Analysis Design Space MVPC Multivariate Process Control

QRA:
Quality Risk Assessment

MVDA

QFD
Quality Function Deployment

! !

DOE is a knowledge building tool for process development MVDA is used both for process understanding and process monitoring
Copyright 2013 OSIsoft, LLC.

Our Customers Goals in Pharma

! The goal in Pharma production is to help take advantage of data present in the development labs, and the production environments all the way from API to the final product. = ROI

Copyright 2013 OSIsoft, LLC.

The Need for Multivariate I

The information is found in the correlation pattern - not in the individual variables!
Copyright 2013 OSIsoft, LLC.

The Need for Multivariate I

Multivariate Control Limits

Copyright 2013 OSIsoft, LLC.

The Need for Multivariate II

! ! Data explosion, more process measurements than ever before, reduce false alarms Spectrometers
! NIR, FTIR, RAMAN, UV, LLSD ! MS, GC, HPLC ! Acoustic, Video ! P, T, Flow, pH ! pO2 pCO2

Process Sensors

! !

Require MVDA methods to visualise and extract reliable information from raw data MVDA handles noise, missing data, correlation and visualize in graphs

Copyright 2013 OSIsoft, LLC.

10

This control chart is familiar to you ?

SMI= x1*Novartis + x2*Roche + x3*Merck + x3*FB....

Copyright 2013 OSIsoft, LLC.

11

So this control chart is easy to understand....

+3 Std.Dev t[1] (Avg) -3 Std.Dev tPS[1] (Batch S0058_A_854826)

PO_WST3433_EXJADE_Drying_V01.M3:3 Predicted Scores [comp. 1]

tPS[1]

-1

-2

-3 0 10 20 30 40 50 60 70 80 90 100 110 120 130 $Time (normalized)

SIMCA-P+ 11 - 01.08.2009 14:42:24

t1= x1*Temperature + x2*Pressure + x3*Agitation speed ....

Copyright 2013 OSIsoft, LLC.

12

MSPC
Observation Level ! Example of a drying step
PO_WST3433_EXJADE_Drying_V01.M3:3 Predicted Scores [comp. 1]
+3 Std.Dev t[1] (Avg) -3 Std.Dev tPS[1] (Batch S0058_A_854826)

Control limits

tPS[1]

-1

-2

Average (signature) of all batches

0 10 20 30 40 50 60 70 80 90 100 110 120 130 $Time (normalized)
SIMCA-P+ 11 - 01.08.2009 14:42:24

-3

New batch assessed by the model

Copyright 2013 OSIsoft, LLC.

13

Statistical Process Control

BATCH CONTROL CHART
Variable, Batch: 617, Phase: 16
-15.4122 * 543071TT--607 - 15.4122 -2.84936 * 543071TT--600 - 2.84936 0.00281207 * 543071PT--644 + 0.00767695 -1.10938 * 543071TT--602 - 1.10938 0.00826856 * 543071-Acetone_43x10e13
1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 8400 8500 8600 8700 8800 8900 9000 9100 9200 9300 Num 9400 9500 9600

average of all batches

2 1 0 -1
9700 9800 9900 10000 10100

Model Data Ciclo - Oct 2010 v5.M2:16 Scores [comp. 1] (Aligned)

control limits ( 3! from avg.)

+3 Std.Dev t[1] (Avg) -3 Std.Dev t[1] (Aligned): 635

-2 -3

SIMCA-P+ 11 - 09.03.2011 19:17:38

t[1]

-4 -5 -6 -7 -8 -9
Model Data Ciclo - Oct 2010 v5 - batch -100 level (scores).M2 0 (PCA-X) t[Comp. 1]/t[Comp. 2]

Batch Process Signature

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700 $Time (smoothed)
SIMCA-P+ 11 - 14.03.2011 17:53:19

20

10 614 615 625 633 624 612 616 639 635 634 618 631 620 626 641 628 638 627

632 621

630

t[2]

-10

637

619 636 622

623

617

-20 -80 -70 -60 -50 -40 -30 -20 -10 0 t[1] R2X[1] = 0.873728 R2X[2] = 0.0564179 Ellipse: Hotelling T2 (0.95) 10 20 30 40 50 60 70 80

SIMCA-P+ 11 - 10.03.2011 19:27:42

Copyright 2013 OSIsoft, LLC.

14

 Copyright 2013 OSIsoft, LLC.

15

Work and Data flow

For Method Development Final Model Batch Level Observation Level
Recorded Process Parameter during granulation
14

Reduction of Dimensionality

ObsID(Obs ID ($PhaseID)) Mixer Power rate of change precss variable 0.01 * Mixer torqute process variable 0.1 * Mixer speed process variable 0.1 * Product temperature process variable Mixer power process variavle (electrical)

Aims: - Creation of batch signature - Identify correlation patterns

13

All Process Parameters

Individual Probes Individual Probes

12

11

10

320

330

340

350

360

370 Num

380

390

400

410

420

!
Copyright 2013 OSIsoft, LLC.

16

Work and Data flow

For Routine Use in Production SIMCA-online
40 20 0 -20 S0007_B_85 S0008_A_85 S0007_A_85 S0040_A_85 S0034_B_85 S0011_B_85 S0012_B_85 S0041_A_85 S0032_B_85 S0044_B_85 S0037_A_85 S0047_B_85 S0033_A_85 S0027_B_85 S0033-B_85 S0047_A_85 S0045_B_85 S0034_A_85 S0042_A_85 S0037_B_85 S0041_B_85 S0035_B_85 S0045_A_85 S0009_B_85 S0035_A_85 S0007_B_85 S0042_B_85 S0032_A_85 S0028_A_85 S0048_B_85 S0031_B_85 S0027_A_85 S0026_B_85 S0048_A_85 S0046_A_85 S0043_A_85 S0030_TEIL S0011_A_85 S0012_A_85 S0039_B_85 S0029-A_85 S0040_B_85 S0031_A_85 S0046_B_85 S0030-A_85 S0014_B_85 S0029-B_85 S0007_A_85 S0006_B_85 S0014_A_85 S0044_A_85 S0039_A_85 S0006_B_85 S0021_A_85 S0025_B_85 S0021-B_85 S0010_A_85 S0024-A_85 S0019_B_85 S0019_A_85 S0018_B_85 S0018_A_85 S0006_A_85 S0010_B_85 S0009_A_85 S0026_A_85 S0023-B_85 S0025_A_85 S0023-A_85 S0008-B_85 S0006_A_85 S0022_B_85 S0022_A_85

PO_WST10332_EXJADE_GRAN_Steintraining - batch level (scores).M1 (PCA-X), All Batches, PS-Complementory tPS[Comp. 1]/tPS[Comp. 2]

Batch Level

-40
tPS[2]

-60 -80 -100 -120 -140 -160 -180 -20 -10 0 10 20 30 40 50 tPS[1] R2X[1] = 0.402627 R2X[2] = 0.341738 Ellipse: Hotelling T2PS (0.95) 60 70 80 90 100 110 S0028-B_85 120

SIMCA-P+ 11 - 08.02.2009 17:15:01

Score ContribPS(S0007_B_854825:15.7895 - Avg:15.7895) , Weight=p1

Observation Level

PO_WST10332_EXJADE_GRAN_Steintraining.M2:7, PS-Complementory Score Contrib PS(S0007_B_854825:15.7895 - Avg:15.7895), Weight=p[1]

1
Missing

-1

Product Temperature

Mixer Torque

Mixer Power Rate

Chopper Speed

Mixer Speed

-2
Flow liquid feed

Aims: -! Conformity check -! Real time release testing - Trend analysis - Root cause analysis

Liquid feed pump speed

Mix Power (calculated)

Var ID (Primary)

PO_WST10332_EXJADE_GRAN_Steintraining.M2:7 Predicted Liquid feed pump speed

+3 Std.Dev XVar(Liquid feed pump speed) (Aligned) (Avg) -3 Std.Dev XVarPS(Liquid feed pump speed) (Batch S0007_B_854825)

SIMCA-P+ 11 - 08.02.2009 17:17:00

120

Identification of responsible Parameter(s)

100

80

60

40

20

-20

Total Liquid Added

Bowl Temperature

Bowl Pressure

Mixer power

-40 0 10 20 $Time (normalized) 30 40

SIMCA-P+ 11 - 08.02.2009 17:17:46

50

Investigation on process data

Increase of level of detail

Copyright 2013 OSIsoft, LLC.

17

MVDA applied to granulation

Example for a qualitative model used for MSPC

! High Shear Granulation (Production Scale) ! Four phases

! ! ! ! ! ! ! ! Dry Mixing Wet mixing Water addition Granulation Power consumption and torque Product temperatures Granulator and chopper speed Pump properties and flow parameters
Copyright 2013 OSIsoft, LLC.

! Variables

18

Key consideration for method development

! Which observations should be included ?
! Sufficient number of batches to cover natural variability ! DoE Data for special cause variations ! Exclusion of anomalous, unsteady, discontinuous data (spikes)

! Which variables should be included ?

! Exclusion of variables with no impact and low reliability ! Weighting and transforming of variables ! How many scores should be considered

! Data alignment and synchronization

! Definition of start/stop conditions and phases ! Merger of variables with different acquisition rates ! Normalization of time based maturity variables vs. absolute time
Copyright 2013 OSIsoft, LLC.

19

Granulation Process
Data Setup Input: Initial conditions X(t)
Variable initial conditions :
!Amount of granulation liquid !Granulation time
1.! 2.! 3.! 4.! 5.! 6.! 7.! 8.! 9.! Power consumption Power consumption rate Torque Product Temperature Mixer Speed Chopper Speed Water addition rate Flow Liquid Speed Liquid Speed Pump

Output: Quality Attributes

Dry Mixing Water Addition Kneading Granulation process

time

IPC results: !LOD !PSD

X0

Y
Copyright 2013 OSIsoft, LLC.

20

Example Granulation
Results of DoE investigation
6 4

S0015_B
2

S0016-A S0015-A S0016-B

S0011-B S0011-A

S0012-B S0012-A

t[2]

S0009A_ S0014_A S0014_B S0013_B S0013-A S0010A_S0009B_ S0010-B

-2

-4

-6 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 t[1]

R2X[1] = 0.494965 R2X[2] = 0.151559 Ellipse: Hotelling T2 (0.95)
SIMCA-P+ 11 - 03.08.2008 17:12:10

10

11

12

"! Different phases during granulation are monitored

"! Process variability are reflected by the red lines "! Clustering of DoE batches can be visualised "! Common cause vs. special cause variation
Copyright 2013 OSIsoft, LLC.

21

Loadings
Identification of process parameters contributing to process variability
Loadings of Granulation Models
0.5 0.4 0.3 0.2 0.1

Mixer Power Rate

Mixer Power Liquid Added Mixer Torque Power (calcula Bowl Bowl Temperature Pressure Product Temperature

p[2]

-0.0 -0.1 -0.2 -0.3 -0.4 -0.5 -0.1

Flow Liquid Feed Liquid Pump Speed Chopper Speed Mixer Speed
0.0 0.1 0.2 p[1]
R2X[1] = 0.401621 R2X[2] = 0.318985
SIMCA-P+ 11 - 04.01.2010 18:07:36

0.3

0.4

0.5

0.6

"! Which parameters are most influential? "! How do the variables correlate to each other?
Copyright 2013 OSIsoft, LLC.

22

First experience in production

Preventive maintenance
PO_WST10332_EXJADE_GRAN_V1 - Batch Level scores: t[1]/t[2]
S0073_B_854826 S0071_A_854826 S0073_A_854826 S0071_B_854826 S0072_A_854826 S0072_B_854826 S0069_B_854826 30 S0074_A_854826 S0068_B_854826 S0067_B_854826 S0067_A_854826 S0068_A_854826

Last campaign This campaign Last batch after liquid feed tube change

20 S0066_B_854826 S0065_A_854826 S0064_B_854826 S0066_A_854826

t[2]

10 S0064_A_854826

S0080_A_854826

S0063_B_854826 S0084_A_854826 S0083_B_854826 0 S0075_B_854826 S0075_A_854826

-10

-20

-10 t[1]

10

20

Ellipse: TCrit (95%) = x ! /19.9118 ! + y ! /17.1288 ! = 1

SIMCA-Batch On-Line View 2.2 - 01.08.2009 16:58:57

Copyright 2013 OSIsoft, LLC.

23

Root cause analysis

Liquid feed pump speed trajectory during the latest 2 campaigns

Copyright 2013 OSIsoft, LLC.

24

Root Cause: Worn out liquid feed tube

By replacing the tube # batches were saved. $ XXXXXXX saved

Copyright 2013 OSIsoft, LLC.

25

Cell cultivation process

! Motivation: established process, not fully characterized, most of process understanding based on experience ! Modeling: > 80 DS batches, fully meeting release specifications but some variability in main fermentation yield is observed ! Defined reference (golden) batches, i.e. batches which provided the highest total amount of antibody during cell culture
26

Copyright 2013 OSIsoft, LLC.

Statistical Process Control for cell cultivation process

Batch: 41m, Phase: main -0.025 * XVar(VCD (cells/mL)) - 0.025 -0.233716 * XVar(Lactate (mg/L)) - 0.233716 -0.0526316 * XVar(Luft (U/L)) - 0.0526316
1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44

-0.38843 * XVar(Glucose (mg/L)) - 0.38843 -0.0251397 * XVar(Titer (mg/L)) - 0.0251397 -0.150825 * XVar(Spec Feed (mL/cells)) - 0.150825

All the variables for each batch are summarized into one quantity (carrier of the information) process signature

YVar(Day) (smoothed and shif ted)

SIMCA-P+ 12.0.1 - 2010-04-09 09:13:16 (UTC+1)

Perfusion Start
10 5
tPS[1]

REFERENCE (i.e. avg of golden batches)

3 sd from ref.

-2

t[2]

-4

0
-6

-5 -10 -20 -15 -10 -5 0 t[1] 5 10 15 20

-8

-10

-12 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Day 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
SIMCA-P+ 12.0.1 - 2011-03-18 12:30:07 (UTC+1)

42 43 44 45

Copyright 2013 OSIsoft, LLC.

27

Improve Process Understanding

Compared the other batches (non-golden) against the golden to establish which are the variables responsible for the observed differences
6

3 sd from ref.
2.0 1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 -2.0

tPS[1]

-2

-4

-6

-8

Score Contribution

D Medium Feed Exchange rate (1/24h)

Medium Feeding Rate (Kg/24h)

Glucose Conc (mg/L)

VCD (1E5 cells/mL)

Lactate Conc (mg/L)

NH4+ Conc (mg/L)

LDH Conc (U/L)

Viability (%)

Weight (kg)

-10

-12 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

Day

SIMCA-P+ 12.0.1 - 2011-03-18 12:22:40 (UTC+1)

Var ID (Primary)
SIMCA-P+ 12.0.1 - 2011-04-29 18:06:16 (UTC+1)

pH

Copyright 2013 OSIsoft, LLC.

28

MVDA learning
! Generated useful process knowledge
! Enhanced process understanding supported by data ! Improved process consistency

! Established key parameters for cell cultivation

! medium feeding rate ! inoculation cell density ! cells aeration

! No correlation between cell behavior and DS quality attributes

golden
10 5

non-g

tPS[2]

0 -5 -10 -20 -15 -10 -5 0 tPS[1] 5 10 15 20

R2X[1] = Ellipse:

0.430057 Hotelling

R2X[2] T2 (0.95)

0.145144

Copyright 2013 OSIsoft, LLC.

29

Petter Moree
petter.moree@umetrics.com Director, Global Product Manager Umetrics

Copyright 2013 OSIsoft, LLC.

30

Brought to you by

Copyright 2013 OSIsoft, LLC.