M. von Stosch - Vaccine Technology Summit 2023 Page 1
34359738368

Faster process development
with hybrid modeling and transfer learning
Dr. Moritz von Stosch
21 March 2023
Vaccine Technology Summit 2023

M. von Stosch - Vaccine Technology Summit 2023 Page 3
Who is DataHow?
2017
Incorporation
45
Team Members
3
Locations
3
Customer Regions
• 4 Co-Founders
• 35yrs R&D experience
• 20yrs Industry experience
• 30yrs Machine Learning
• 9 EU Countries
• 40% Eng.
• 30% ML
• 30% IT
• HQ in Zürich (CH)
• Subsidiary in Lisbon (PT)
• Subsidiary in Milano (IT)
• 60% Europe
• 30% USA
• 10% Asia

M. von Stosch - Vaccine Technology Summit 2023 Page 4
Our Partners
> Hundreds of industrial
process data sets
> 150 trained users
on DataHowLab
Technology Partners Academic Partners
Customers
Customers

Page 5
Technologies we believe will change the way process are developed.
M. von Stosch - Vaccine Technology Summit 2023
What?
Automatic transfer of knowledge between
products & scales
Where?
DataHowLab + SpectraHow
Why?
• Leverages all available data
• Further reduces data requirements
• Improves design of experiments
• Makes Quality by Design economically
attractive
What?
Combine process engineering knowledge
with machine-learning
Where?
DataHowLab backend
Why?
• Reduces data requirements
• Extrapolates to multiple products
and/or process features
• Increases process understanding
Active Learning
What?
Direct learning and continuous model
refinement from process data
Where?
DataHowLab + SpectraHow
Why?
• Automatically learns from new data
• No need of prior knowledge
• Adapts to new knowledge and
improves optimal decision making
Hybrid Models Transfer Learning

Why we believe process development &
manufacturing will change

M. von Stosch - Vaccine Technology Summit 2023 Page 7
Opportunities shaping the future of CMC development &
manufacturing
Data:
- Data-infrastructure
- Data standards
- Price of Data generation
decreases
Machine-learning:
- Preventive Healthcare
- Drug Discovery
- Clinical Trials
- CMC development &
manufacturing
Automation & High-
throughput:
- Increase of high-
throughput process
equipment & analytics
will shift the bottleneck
to design creation and
data analysis
Analytics:
- Increasing availability of
online analytics will
increase the amount of
data
- Increasing availability of
off-line analytics (MCMS)
will increase information
about molecule

M. von Stosch - Vaccine Technology Summit 2023 Page 8
Threads to CMC development & manufacturing
AI powered Drug
Discovery & Clinical
Trials:
- Increase in molecules in
the pipeline shifting
bottleneck to CMC
Preventive
Healthcare:
- Reducing margins to 13%
or completely eliminating
them
Biosimilars & Patent
Cliff:
- Will reduce margins for
blockbuster drugs,
requiring
Cell Therapies &
Continuous
Processing:
- For these fields modeling
no longer is an option but
a necessity

M. von Stosch - Vaccine Technology Summit 2023 Page 9
Trends shaping the future of CMC development & manufacturing
Data:
- Data-infrastructure
- Data standards
- Price of Data generation
decreases
Machine-learning:
- Preventive Healthcare
- Drug Discovery
- Clinical Trials
- CMC development &
manufacturing
Automation & High-
throughput:
- Increase of high-
throughput process
equipment & analytics
will shift the bottleneck
to design creation and
data analysis
Analytics:
- Increasing availability of
online analytics will
increase the amount of
data
- Increasing availability of
off-line analytics (MCMS)
will increase information
about molecule
AI powered Drug
Discovery & Clinical
Trials:
- Increase in molecules in
the pipeline shifting
bottleneck to CMC
Preventive
Healthcare:
- Reducing margins to 13%
or completely eliminating
them
Biosimilars & Patent
Cliff:
- Will reduce margins for
blockbuster drugs,
requiring cost effective
manufacturing.
Cell Therapies &
Continuous
Processing:
- For these fields modeling
no longer is an option but
a necessity

M. von Stosch - Vaccine Technology Summit 2023 Page 10
Develop Processes Faster Every Time without compromising yield
#
of
experiments
per
development
cycle
101
102
100
103
x
x
x
x
x
x
x
x
x
Drug candidates
x
x
x x
x
x x x
x
x
Transfer Learning
across Processes & Scales

How could the future look like?

M. von Stosch - Vaccine Technology Summit 2023 Page 12
Process life-cycle activities & modeling opportunity
Process Conditions Screening
● How do the cells behave?
● Over-&-over doing the same # of runs,
checking the same conditions
● What if, you would already have an idea
from previous studies?
Process Conditions Optimization
● Where does the process perform best?
● Available knowledge informed
investigation of factors & their ranges.
● What if, you could reuse the knowledge
from previous phases?
Process Scale-up
● Does the process performance scale?
● Performing runs at larger scale to check
for shifts in rate-limiting steps.
● What if you would need less runs, due
to the model providing more insights
Process Characterization
● Can small changes throw process off?
● Derive process quality control &
monitoring strategy by variation study.
● What if experimental evidence was only
required where the model is not sure?
Process Performance Qualification
● Is process robustly delivering quality?
● Repeat runs often enough to conclude
on consistency of performance.
● What if you would require less runs, as
performance agrees with predictions?
Root-cause analysis
● What is source of process deviations?
● Performing runs to qualify assumption
on root-cause of observed deviations.
● What if you could study the result of
deviations with a process model?

M. von Stosch - Vaccine Technology Summit 2023 Page 13
Imagine, you start with a new candidate…..
New candidate
Scale
exploration
Calibration
Experiments
Transfer to
next scale
Filing
Root cause
analysis
Validation
experiments
Transfer to
manufacturing
Data Analysis,
Transfer Learning &
Process modeling
Which Cell line?
Which Parameters?
Which Media?
Learning
opportunity
Transfer into
digital twin

M. von Stosch - Vaccine Technology Summit 2023 Page 14
Collaboration around Digital Process Development Platform
Technician/
Operator
Process
Scientist
Data
Scientist
Manager
Uploads/
Analyzes
Data
Obtains
Visualizations
for Report
Obtains
Insight
Builds/Uses
Models
Builds/
Validates
Models
Reduces
Standard
Work
Obtains
Global View
Validates
Scenarios
Effort
Win
Digital
Process
Platform

Transfer Learning

Starting Point: Model-based process development for one molecule
Model-based process development FOR ONE MOLECULE
Customer
Experiment
Reduction
Potential
Project Type Description
A
(Big Pharma)
50 - 60 % Ph1 to Ph2
(screening + scale-up, biosimilar)
Screening and scale-up experiments could be designed more
effectively, using process knowledge and data inherent correlation.
B
(CDMO)
50 - 60 % Ph2
(workflow improvement study, optimization)
Multi-objective max, optimizing >5 process parameters & trajectories
(dynamic optimization).
Reduction of experiments, due to increased insight in process
dynamics.
C
(Big Pharma)
20 – 30 % Ph3
(characterisation + scale-up, workflow improvement)
More effective scaling up the process and assessing robustness, due
to process dynamics insights and transfer of knowledge across scales.
Utilizing DataHowLab standard workflow significant gains in efficiency are possible,
translating into faster process development at defined technical risk.
DataHow’s broad process optimization capabilities, process scope and flexibility allow it to tackle
a large variety of process challenges,
• Hybrid & Quality modeling
• Process Robustness Analysis
• Design space Identification
• Process Optimization
*All customers A to F are amongst the 20 largest big pharma and CDMO companies
M. von Stosch - Vaccine Technology Summit 2023 Page 16

M. von Stosch - Vaccine Technology Summit 2023 Page 17
Can we use embedding technology to transfer learning between CHO cell-lines?
Irish
Scottish
Geordie
English
Speech
to Text

M. von Stosch - Vaccine Technology Summit 2023 Page 18
The embedding method outperform classical approaches in describing
the systems behaviour with few experiments from novel process.
0.4
0.8
1.2
N=2 N=4 N=12
Error
of
titer
for
a
New
Product
N=6 N=8 N=16
New Product only
One Hot (dummy)
Product Embedding
Knowledge transfer across cell lines using hybrid Gaussian process models with entity embedding vectors
Clemens Hutter, Moritz von Stosch, Mariano N. Cruz Bournazou, Alessandro Butté
https://doi.org/10.1002/bit.27907

M. von Stosch - Vaccine Technology Summit 2023 Page 19
Combining transfer learning with active learning for model-based process
development.
Model
I
m
p
r
o
v
e
d
T
a
r
g
e
t
?
ne Parallel
Experiments
Design 1
Design 2
Design ne
Design ne
experiments
Excitation
Design
(DoE, iDoE)
Model
I
m
p
r
o
v
e
d
T
a
r
g
e
t
?
ne Parallel
Experiments
Design 1
Design 2
Design ne
Design ne
experiments
Calibration
Design

Transfer Learning on a Practical Example

M. von Stosch - Vaccine Technology Summit 2023 Page 21
Too hard to model, too little variation in process parameters too
much in material attributes
Cell A: Bulk Pool
Cell A: Clone 1
Cell B
Production
medium A
Production
medium B
Feed 1
Feed 2
15
Runs
Test
8
Runs

M. von Stosch - Vaccine Technology Summit 2023 Page 22
Increasing variation in process parameters by adopting transfer
learning technology
15
Runs
Campaign 1
3 cell lines
2 production media
2 feed types
Product: Mab 1
15
Runs
17
Runs
+
Campaign 2
1 new cell line
3 new production media
4 new feed types
Product: complex ScFv fusion
Transfer Learning
Test
8
Runs
Test
8
Runs

M. von Stosch - Vaccine Technology Summit 2023 Page 23
Understanding sources of variation and exploiting them to
achieve greater process performance
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Ammomia Viability Glucose Lactate VCD Titer
Relative
RMSE
Model predictions on previously unseen data
One Product Only Transfer learning
15
Runs
17
Runs
+
Campaign 2
1 new cell line
3 new production media
4 new feed types
Product: complex ScFv fusion
Test
8
Runs

To date: Accelerate reduction potential when deployed across molecules
with transfer learning.
Customer
Experiment
Reduction
Potential
Project Type Description
D
(CDMO)
80%
Ph2
(platform optimization)
Knowledge transfer across mAb products allowed to identify optimal
conditions with only 10 experiments, even for CQAs
E
(CDMO)
70%
Ph1
(screening)
Knowledge transfer across cell lines allowed identifying optimal
operation region with less experiments for new cell line.
F
(Big Pharma)
60%
Ph2
(platform optimization
Knowledge transfer across mAb products allowed to identify optimal
feeding strategy.
Consistent use of DataHowLab provides the base to learn across molecules. This inherits
outstanding potential for acceleration
+ Knowledge transfer
• Hybrid & Quality modeling
• Process Robustness Analysis
• Design space Identification
• Process Optimization
Model-based process development ACROSS MOLECULES
M. von Stosch - Vaccine Technology Summit 2023 Page 24

Digital Twins for model-based process
development

M. von Stosch - Vaccine Technology Summit 2023 Page 26
Data
Historical Data
Knowledge Transfer
Design
Predictive Monitoring
What is a digital twin?

M. von Stosch - Vaccine Technology Summit 2023 Page 27
Meet the challenger

M. von Stosch - Vaccine Technology Summit 2023 Page 28
Why the ambr 250 perfusion system is the perfect challenge
process complexity
advanced operation
high parallelization
long experimental
runs
challenges
Biological systems are much more complex than cars or
turbines
Perfusion processes require a complex control and
operation of many inputs in a highly sensitive process
In High Throughput a large number of experiments
must be operated at the same time
Perfusion experiments are significantly longer than fed-
batch experiments

M. von Stosch - Vaccine Technology Summit 2023 Page 29
Answering to process development needs with a digital twin
DataHowLab
Data Knowledge
Hybrid model
Digitalization and control platform
Online
data
Historical
data
Real
Process
DigitalTwins
exp. planning
model driven
validation
model based
operation
predictive
maintenance
automation
forecasting

M. von Stosch - Vaccine Technology Summit 2023
Page 30
Examples of what could be done with this insilico bioprocess development platform
mbDoE & Process
Optimization
Data Visualization
& Analysis
What-if & Design
Space Analysis
Online
Forecasting &
monitoring

Summary

M. von Stosch - Vaccine Technology Summit 2023 Page 32
The traditional approach & the self-learning digital bioprocess twin
Risk
assessment
Statistical
Design of
Experiments
Multivariate
Response Data
Modeling
Automatic High-
throughput
platform Documentation
Reutilized knowledge
Process quasi developed de novo
for every candidate.
Traditional
Quality
by
Design
(QbD)
approach
( )
( )
v
d v
d
d v l d
Glc
Glc Glc v
Gln
Gln Gln v
Lac
Glc Glc Lac v
Osm
Osm v
Osm
T v
dX
X
dt
dX
X X
dt
dC
q C X
dt
dC
q C X
dt
dC
q C q X
dt
dC
m X
dt Y
dT
q X
dt
µ µ
µ µ
µ
ì
ï = -
ï
ï
= -
ï
ï
ï
= -
ï
ï
ï
= -
í
ï
ï
¢
= -
ï
ï
ï æ ö
= - +
ç ÷
ï
è ø
ï
ï
=
ï
î
!"
!#
= % & " − ( & "
!)
!#
= −*!/# & % & " − ( & ) − )$%
!+
!#
= ,
! =
#
$
% =
0.4 ) *
0.5 + *
, = 0.043 & exp(0.6 & (# − 6.9))
÷
÷
÷
ø
ö
ç
ç
ç
è
æ
×
÷
÷
÷
ø
ö
ç
ç
ç
è
æ
×
×
=
÷
÷
÷
ø
ö
ç
ç
ç
è
æ
µ
Glu
Lac
v
v
X
X
Glu
X
Lac
X
Glu
Lac
dt
d
0
0
0
0
0
0
-!
2-"
-#
!"
!#
=
−1 −1
2 0
0 1
)
*!
*"
+
,#$
-
) "#$ − "
- = -!, -", -#
$
0%
0&
/
/
/
I
v
dt
X
dP
P
D
X
X
P
I
v
dt
dP
X
D
X
dt
dX
X
P
X
P
×
=
×
-
×
÷
ø
ö
ç
è
æ
×
+
×
=
×
-
×
=
µ
µ
Model guided
Experiment design
Physical Lab
experiments
Drug candidates
#
of
experiments
101
102
100
103
x x
x x
x
x x x
x
x
x
x
x
x
x
x
x
x
x
Progressively experiments are run
insilico on the Digital Bioprocess Twins.
Self-learning
digital
bioprocess
twins
QbD2.0

M. von Stosch - Vaccine Technology Summit 2023 Page 33
Examples of other model-based process development scenarios
1. Wholistic process
development
2.
drug candidates
#
of
experiments
101
102
100
103
x x
x x
x
x x x
x
x
x
x
x
x
x
x
x
x
x
Across candidate learning 3. Across Scale learning

Thank you
m.vonstosch@datahow.ch
www.datahow.ch