The purpose of this webinar is to highlight GSK's approach to:
- create a simple, mechanistically descriptive model
- verify its utility with clarity of objectives, and
- communicate understanding via creative but aligned metrics
... for a challenging chemical reaction.
1. Development of an
ammonolysis reaction
kinetics model for improved
process understanding and
parametric flexibility
Nick Falco
Product Development
GlaxoSmithKline R&D
2. Purpose and scope
Purpose of this webinar is to highlight our approach to
create a simple, mechanistically descriptive model
verify its utility with clarity of objectives, and
communicate understanding via creative but aligned metrics
for a challenging chemical reaction
Development of an ammonolysis reaction kinetics model
2
Welcome!
10 Sept 2014
3. Purpose and scope
Introduction – Chemistry, equipment, and limitations
Methodology – Modelling objectives and approach
Model fitting – Internal data, assumptions, and the model
Model verification – External data and selection of criteria
Impact on the process – So what does it all mean?
Development of an ammonolysis reaction kinetics model
3
Overview of the presentation
10 Sept 2014
5. Introduction - chemistry
• Intermediate stage in synthesis of an active pharmaceutical ingredient (API)
– Solution chemistry in acetonitrile and aqueous ammonium hydroxide
– Reaction is typically run at > 80°C
• Compound A forms reacts with ammonium hydroxide in solution to generate
both desired product (B) and over-addition, di-amino impurity (C)
• Downstream system has limited purging capability of both impurities of focus
– Compound A – un-reacted starting material
– Compound C – di-amino impurity from over-addition
What are the constraints?
Development of an ammonolysis reaction
kinetics model
5
Ammonolysis reaction presented an opportunity for kinetics modelling
10 Sept 2014
6. Introduction - Equipment limitations, scientific challenges
•High pressure - at elevated temperatures necessary for reaction progress, system is self-pressurizing via ammonia vapor
•Complex sampling protocols at elevated temperatures and pressures
•Early iterations of process utilized pressure reactors
–Allowing higher temperatures and concentrations of ammonium hydroxide
•Process designed to run at > 85°C for maximum speed while maintaining processability in general use reactors (ie pressure limited)
•Manufacturing plant imposed pressure limitation
–Restricted temperature and/or NH4OH concentration
–Maximum operating temp ~85-95°C
–Initial response – go as fast as we can ~ 85-95°C
Development of an ammonolysis reaction kinetics model
6
Elevated system pressure resulted in cumbersome sampling and barriers to gaining knowledge
10 Sept 2014
NH3 (v)
N2 (v)
A (s)
A (soln)
A B (soln)
B C (soln)
NH3 (soln)
Setting conditions in the absence of relative reaction kinetics data...
...what could go wrong
8. Methodology - objectives
Chemistry team designed a ‘robustness’ Design of Experiments (DoE) for the purpose of identifying statistically significant parameters.
But, an opportunity for a kinetics model, built on the same data set…
Objectives of a mechanism-based model:
•Describe the competing reaction kinetics
–Identify relative kinetics profile for primary and secondary reactions
•Simplify the manufacturing process
– Predictive capability for ‘plant-floor’ chemist
– Reduce sampling load
•Improve process understanding
Development of an ammonolysis reaction kinetics model
8
‘Free’ data from statistical DoE provided a way forward for mechanistic understanding
10 Sept 2014
9. Methodology - approach
Start simple for model build
–Relatively complex system, start with fewer components for fitting and simulation
– Assess fit, if not suitable then build further
Set reaction completion criteria based on purging capability downstream (ie how much can the subsequent stages remove)
–un-reacted compound A
–over-reacted compound C
Identify modelling outputs
–Describe – Reaction kinetics equations
–Simplify – Prediction of reaction endpoint
–Improve – Brilliantly demonstrate knowledge
Development of an ammonolysis reaction kinetics model
9
A ‘keep it simple’ approach for maximum applicability
10 Sept 2014
0.00
0.50
1.00
1.50
2.00
2.50
3.00
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
Reaction Temperature (°C)
Time (hr)
Start here
11. Model fitting – internal data and model scope
•Experimental DoE was initially designed for statistical output (ie parametric criticality)
–Laboratory robustness
–Supersaturated partial factorial design
–Ten experiments
–Reaction stopped when complete (A < completion criteria)
•Five ‘Large-scale mimic, Extended time’ experiments intended to generate significant quantity of Compound C (over-addition impurity)
• Model scope limited by process parameter ranges established in the DoE
–Reaction temperature, Concentration of aqueous NH4OH, Solvent volumes of aq NH4OH and acetonitrile
Development of an ammonolysis reaction kinetics model
11
Piggybacking on a statistical approach
10 Sept 2014
Experiment
NH4OH solution (% w/w NH3)
Temp (°C)
MeCN (volumes)
NH4OH (volumes)
Designation
1
18
75
4
12
Robustness
2
18
75
2
8
Robustness
3
18
93
2
8
Robustness
4
18
100
2
12
Robustness
5
18
100
2
8
Robustness
6
29
75
2
12
Robustness
7
29
75
4
8
Robustness
8
29
93
4
8
Robustness
9
29
100
2
12
Robustness
10
29
100
4
8
Robustness
11
18
80
2.4
9.6
Large-scale mimic, extended time
12
19
87
2.4
9.6
Large-scale mimic
13
20
80
2.4
9.7
Large-scale mimic, extended time
14
20
87
2.4
9.6
Large-scale mimic, extended time
15
22
80
2.4
9.6
Large-scale mimic, extended time
Parameter
Concentration of ammonia in water
Volumes of ammonium hydroxide
Volumes of acetonitrile
Temperature of reaction
Time of reaction
12. Model fitting – assumptions
•No impact of dissolution of Compound A on reaction kinetics
–Despite forming a gel-like heterogeneous mixture during heat- up, the mixture dissolved as contents reached reaction temperature.
•Heat-up profiles were measured for several experiments and applied across the rest
–Linear heat-up rate was assumed
–Heat-up times were relatively short (approx 1-1.5 hours) compared to overall reaction times
–Incorporation of the heat-up ramp allowed for slower, fixed ramps to be used when scaling-up to the Pilot Plant reactor (which was consistently slower to heat up)
•Regarding ammonia dissolution, reaction was assumed not to be mass transfer limited (ie solution ammonia was available for reaction)
Development of an ammonolysis reaction kinetics model
12
Starting simple, with solution chemistry
10 Sept 2014
NH3 (v)
N2 (v)
A (s)
A (soln)
A B (soln)
B C (soln)
NH3 (soln)
13. Model fitting – internal data
•Reaction HPLC samples were taken as the reaction neared completion and at extended times
– Full impurity profile was collected, but modelling was focused on only Compounds A, B, and C
– Reaction deemed complete when A < completion criteria
– Reaction deemed past point of no return when C > excess criteria
•This led to modelling rationale of ‘reaction window’ was used to signify reaction success
Development of an ammonolysis reaction kinetics model
13
An example data set starts to for our understanding of the ‘reaction window’
10 Sept 2014
Sample number
Overall reaction time
(hr)
Time upon reaching rxn temp
(hr)
Compound C
(% area)
Compound A
(% area)
1
9.7
7.0
0.15
1.87
2
11.7
9.0
0.19
0.56
3
12.7
10.0
0.21
0.33
4
13.7
11.0
0.24
0.18
5
14.7
12.0
0.25
0.11
6
29.9
27.2
0.55
0.00
7
30.7
28.0
0.58
0.00
8
31.7
29.0
0.59
0.00
9
33.2
30.5
0.63
0.00
10
36.7
34.0
0.69
0.00
Time of ‘completion’
Time of ‘failure’
Reaction window
0.00
0.50
1.00
1.50
2.00
2.50
3.00
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
Reaction Temperature (°C)
Time (hr)
14. Model fitting – description of the model
• Reactions in solution were described in the Dynochem
– Rate constants and activation energy were fit for both primary and secondary reactions
Development of an ammonolysis reaction
kinetics model
14
Second order reactions in solution resulted in a good fit
10 Sept 2014
AB
1 1
exp 0
RT RT
r k E
ref
a
r = reaction rate
k0 = rate constant at reference temperature
Ea = activation energy
R = gas constant
Tref = reference temperature
T = reaction temperature
Phase Solution liquid (set volume)
A g plot
NH3 g plot
B g plot
C g plot
acetonitrile g plot
water g plot
temperature C plot
Reaction
s in
Solution
k> 1 E-y L/mol.s at 75 C Ea> kJ/mol * A + NH3 > B
k> 4 E-x L/mol.s at 75 C Ea> kJ/mol * B + NH3 > C
Variable
s
Conversion % plot
C % plot
A % plot
Calculat
e
Conversion:= solution.B / solution.A.Y0 * MWB/MWA
C := (solution.C*C.Mw/1000) / (solution.A.Y0*A.Mw/1000) |
B:= solution.A / solution.A.Y0
Finished
% A (predicted) % C (experimental) C
% C (experimental) % A (predicted)
Time of
completion
Time of
failure
15. Model fitting – the resulting model
• Describe the competing reaction kinetics
– Identify relative kinetics profile for primary and secondary reactions
• Upon fitting the kinetics model was described as the following set of
mechanism-based reaction equations
Next up: how well does it fit?
Development of an ammonolysis reaction
kinetics model
15
Description of the reaction kinetics is a good start
10 Sept 2014
1 3 A NH
1 1
1 10 exp
[ ]
mol RT RT
kJ
y
mol s
L
dt
d A
r
ref
x
2 3 B NH
1 1
4 10 exp
[ ]
mol RT RT
kJ
z
mol s
L
dt
d C
r
ref
a
Primary reaction: A + NH3 → B
Tref = 75 °C
Competing reaction: B+ NH3 → C
Tref = 75 °C
17. Model verification – external data
•A separate exercise was performed using additional or external data sets to verify or validate the model.
–Eight additional lab-scale runs were performed within the range explored in the Internal Data DoE
– Pilot plant runs were monitored as the team scaled the process up.
–Pilot plant runs all > 85°C
Development of an ammonolysis reaction kinetics model
17
Validating the data at lab and pilot plant scales
10 Sept 2014
Reference
NH4OH soln (% w/w NH3)
Temp (°C)
MeCN (L/kg)
NH4OH (L/kg)
Laboratory 1
18
93
3
12
Laboratory 2
18
93
2.4
9.6
Laboratory 3
22
75
3
12
Laboratory 4
22
85
3
12
Laboratory 5
22
85
2.4
9.6
Laboratory 6
22
85
2.4
9.6
Laboratory 7
26
85
3
12
Laboratory 8
29
75
3
12
Pilot Plant 1
19
87
2.4
9.6
Pilot Plant 2
19
87
2.4
9.6
Pilot Plant 3
19
87
2.4
9.6
Pilot Plant 4
19
87
2.4
9.6
Pilot Plant 5
19
87
2.4
9.6
Pilot Plant 6
20
87
2.4
9.6
Pilot Plant 7
20
87
2.4
9.6
Pilot Plant 8
20
87
2.4
9.6
Pilot Plant 9
20
87
2.4
9.6
Pilot Plant 10
20
87
2.4
9.6
Pilot Plant 11
20
87
2.4
9.6
Pilot Plant 12
20
87
2.4
9.6
Pilot Plant 13
20
87
2.4
9.6
Pilot Plant 14
20
87
2.4
9.6
Recall the team’s initial response to pressure limitation...
...high temperature, go fast!
18. Model verification - discussion
• Measured data – HPLC reaction samples
– Dynochem model fit the model to HPLC area percent
• Initial verification directly utilized these sample results
• Data visualized using predicted vs actual plots
Development of an ammonolysis reaction
kinetics model
18
First, using area percent to determine model fit
10 Sept 2014
R² = 0.9505
0
10
20
30
40
50
60
70
80
90
100
0 10 20 30 40 50 60 70 80 90 100
Predicted %area GSK2127769A
Experimental %area GSK2127769A
y=x
% area GSK2127769A - Fitting
%area GSK2127769A - Verification
Linear trendline (all points)
R² = 0.9407
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2
Predicted %area GSK2485959A
Experimental %area GSK2485959A
y=x
%area GSK2485959A - Fitting
%area GSK2485959A - Verification
Linear trendline (all points)
Compound C
Compound C
Compound C
Compound C
Compound C
Compound A
Compound A
Compound A
Compound A
Compound Compound A Predicted vs Actual C Predicted vs Actual
Low-level % area for
consumption of A
High level %
area for
generation of C
Model fits in focal areas... but is this representative of the end goal?
19. Back to the objectives
Development of an ammonolysis reaction kinetics model
19
A look back at what we really want
10 Sept 2014
Objectives of the mechanism-based model:
Describe the competing reaction kinetics
Identify relative kinetics profile for primary and secondary reactions
Simplify the manufacturing process
Predictive capability for ‘plant-floor’ chemist
Reduce sampling load
Improve process understanding
Need to get to
Reaction endpoint time
‘Failure’ time
for meaningful conclusions
20. Model verification - assumptions
We want to be able to extract reaction endpoint time and failure time
•Sampling of reaction was assumed to be adequately frequent to calculate a reaction completion time
•As pilot plant runs were intended to deliver high quality material, it was not feasible to extend these runs past the point of ‘failure’
(ie Compound C > excess C)
–A linear trend of Compound C area percent was used to determine time of failure
Development of an ammonolysis reaction kinetics model
20
Enabling a discussion of reaction endpoint
10 Sept 2014
% A (predicted)
% C (experimental)
C
% C (experimental)
% A (predicted)
Adequate sampling: This gap is small
In absence of these data points, linear trend was extrapolated
21. Model verification - discussion
•Output data was revised to ‘Time to completion’ and ‘Time to failure’.
•Obtained both experimental and model-predicted values
–Experimental Actual- or extrapolated times to Completion/Failure
–Model Dynochem-simulated time to Completion/Failure
Development of an ammonolysis reaction kinetics model
21
Revised metrics for a more approachable model
10 Sept 2014
Reference
NH4OH solution
(% w/w NH3)
Time to completion (hr)
Time to failure (hr)
Temp (°C)
Experimental
Model
Experimental
Model
Fitting data set
1
18
75
19.5
20.1
70.0†
67.1
2
18
75
21.5
19.0
60.3†
63.6
3
18
93
5.0
4.4
8.5†
9.7
4
18
100
3.0
2.3
4.8†
4.5
5
18
100
3.5
3.5
4.6†
6.8
6
29
75
11.0
11.3
40.2†
37.8
7
29
75
na
na
na
na
8
29
93
4.0
3.3
5.6†
7.4
9
29
100
1.0
1.3
2.9†
2.8
10
29
100
2.3
1.8
2.8†
3.7
11
18
80
13.0
11.9
39
36.2
12
19
87
7.0
6.0
16.2†
16.3
13
20
80
11.0
10.7
34
32.7
14
20
87
8.0
5.7
16
15.3
15
22
80
9.2
9.7
30
29.8
Verification data set
1
18
93
na
na
na
na
2
18
93
5.0
4.3
9.1†
9.7
3
22
75
na
na
na
na
4
22
85
7.0
6.6
20.6†
17.8
5
22
85
7.0
6.7
17.9†
17.9
6
22
85
7.0
6.9
17.1†
18.0
7
26
85
5.0
5.6
14.6†
15.2
8
29
75
10.0
12.0
39.3†
40.3
% A (predicted)
% C (experimental)
C
% C (experimental)
% A (predicted)
Time to completion
< criteria % a/a A
Time to failure
> criteria % a/a C
22. Model verification - discussion
• Statistical analysis using time to
completion & failure
– XY Predicted vs Actual
– Root Mean Square Error
– Mean Absolute Error
• Fitting and Verification data sets show
good model fit
Development of an ammonolysis reaction
kinetics model
22
Applying a statistical approach to tangible outputs
10 Sept 2014
R² = 0.965
0.0
5.0
10.0
15.0
20.0
25.0
0.0 5.0 10.0 15.0 20.0 25.0
Model predicted time to reaction completion (hr)
Experimental time to reaction completion (hr)
Reaction endpoint (hr) - Fitting
Reaction endpoint (hr) - Verification
y = x
Linear trendline (all points)
‘Time to completion’
Reaction endpoint Predicted vs Actual
Time to completion Time to failure
Fitting data External data Fitting data External data
Number of runs N 14 6 14 6
Mean hr 8.5 6.8 24.0 19.8
Sum of squared residuals SSR hr 16.5 4.9 48.8 10.1
Root mean square error RMSE hr 1.1 0.9 1.9 1.3
% Root mean square error %RMSE % 12.8 13.3 7.8 6.5
Mean absolute error MAE hr 0.83 0.68 1.5 0.97
% Mean absolute error %MAE % 9.7 10.0 6.4 4.9
R² = 0.992
0.0
10.0
20.0
30.0
40.0
50.0
60.0
70.0
0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0
Model predicted time to failure (hr)
Experimental (or linear predicted) time to failure (0.7% GSK2485959A) (hr)
Reaction failure point (hr) - Fitting
Reaction failure point (hr) - Verification
y = x
Linear trendline (all points)
‘Time to failure’
Reaction failure point Predicted vs Actual
24. Impact – Visualizing the knowledge space
Influence: Reducing the reaction temperature from >85°C to 80°C led to wider reaction success
window and manufacturing flexibility
Understanding: Establishing model fit in these terms demonstrated process understanding
Development of an ammonolysis reaction
kinetics model
24
Illustrating fit and Influencing Process Conditions
10 Sept 2014
0
10
20
30
40
50
60
70
°C 75 75 80 80 80 85 85 87 93 93
% w/w NH3 18 29 18 20 22 22 26 20 18 29
Reaction time (hr)
[upon reaching reaction temperature]
Model predicted time to > 0.7% GSK2485959A
Model predicted reaction endpoint
Experimental time to > 0.7% GSK2485959A
Experimental reaction endpoint
estimated data point for reaction failure based on linear formation rate of
measured data point for reaction failure
failuree
failure
Recall: Setting conditions in the absence of relative reaction kinetics data...
...what could go wrong
Statistical model
indicated that
Time
Temperature
Ammonia conc
were significant
25. Impact – Process recommendations
• Describe competing reaction kinetics
Reaction kinetics equations were used directly as a QbD
medium impact model, providing a relative kinetic profile
for desired and undesired reactions, as well as an ability to
simulate future conditions.
The model was important for establishing design space.
• Simplify the manufacturing process
Reaction sampling no longer required as part of
process.
Predictive capability via Dynochem simulation
improved adaptability to future parametric changes
• Improve process understanding
Verification of model fit and visualization of the
knowledge space influenced the final conditions of the
process. Reducing temperature allowed for a wider
window for reaction success.
Model sufficiently established knowledge over reaction
time, enabling processing flexibility.
Kinetics model enhanced the statistical model
Development of an ammonolysis reaction
kinetics model
25
Reflecting on the modelling objectives
10 Sept 2014
Primary reaction: A + NH3 → B
Tref = 75 °C
Competing reaction: B+ NH3 → C
Tref = 75 °C
0
10
20
30
40
50
60
70
°C 75 75 80 80 80 85 85 87 93 93
% w/w NH3 18 29 18 20 22 22 26 20 18 29
Reaction time (hr)
[upon reaching reaction temperature]
Model predicted time to > 0.7% GSK2485959A
Model predicted reaction endpoint
Experimental time to > 0.7% GSK2485959A
Experimental reaction endpoint
estimated data point for reaction failure based on linear formation rate of
measured data point for reaction failure
1 3 A NH
1 1
1 10 exp
[ ]
mol RT RT
kJ
y
mol s
L
dt
d A
r
ref
x
2 3 B NH
1 1
4 10 exp
[ ]
mol RT RT
kJ
z
mol s
L
dt
d C
r
ref
a
26. Thank You
Acknowledgments
James Wertman
Steve Goodman
Qiaogong Su
Joe Flisak
Christian Airiau
Bob Yule
Phil Dell’Orco
Rest of GSK R&D Team
28. Statistical definitions
Statistical definitions used in model verification
• Root mean square error (RMSE):
• Mean absolute error (MAE):
Development of an ammonolysis reaction
kinetics model
10 Sept 2014 28
N
Experimental Model
RMSE
2
N
Experimental Model
MAE