674171_en-US

liliconcom

Purification Process modeling with several BU operations

Hi,

developing a process for drug substance production is done in several steps within BUs, e.g. there are different purification steps

Picture: each color is a BU where a DOE has been run and a model has been found which explains the process. Inputs are different in each BU, the experiments have been conducted independently, so output material in BU1 is not the same used in BU2 for further processing.

This is my customer a challenge: They want to combine all that data/models to get a final model for the whole process. 

As the several models are from independent DOEs each, and the inputs are as well not necessarily the same in each unit operation, I wonder if you have a good way on how to deal with such a situation.

• Just combining will lead to several missing values,
• then it is not a continuous process that the material created in one BU will be used in the experiment for the next BU, it has new material which should act like the other material but not for sure.

If that would be a continuous process I even thought about SEM to set-up the process and embed variables and knowledge gained from the several steps into one big model.

The other idea I'd have is simulation and using validation techniques with null factors and assuming we have a continuous setup and see the results and then test with confirmation runs. 

Any ideas or experiences? 

Thx,

Martin

Re: Purification Process modeling with several BU operations

Hi @martindemel ,

I have definitely heard about this requirement before. It was from Medimmune (now part of AstraZeneca) and it was nearly ten years ago.

I never found a useful solution.

SEM was not in JMP (Pro) then, but I had thought that it might be the solution. Now, I don't think it is.

I think that the only way might be to do the whole thing as a very big designed experiment.

Sorry I don't have a better answer,
Phil

Re: Purification Process modeling with several BU operations

Also the unit experiments do not appear to be connected in any way, e.g. the measured outputs of the prior step do not appear as a covariate or a measured input of the next step. If they had treated the output of the prior process step as a covariate for the next process step then at least we could explore how the output achieved from the prior step impacts the model fitted to the following step. 

Re: Purification Process modeling with several BU operations

Right, @malcolm_moore1 . My memory from the Medimmune requirement was that "amount of impurity A" might be a response from the 1st step, and a factor in the 2nd step ...just as an example.

Re: Purification Process modeling with several BU operations

It helps if you can chart out the expected relationships.   In the graph below, the X's are the controllable process factors, M's are the input material properties (or other covariates, the Y's are the measure output, and the f's are prediction model for the Y's, based on the X's and M's.  

I have done this type of work before for a 2 or three step process, but it is tedious and in JMP requires setting up a rather complicated collection of formulas.  An alternative is to setup the relationships and models for the process steps in an excel workbook and use the Excel Profiler to do visualization and Monte Carlo Simulation.  

Re: Purification Process modeling with several BU operations

The Link Profilers feature links both inputs and outputs. Could you create a large-ish dashboard consisting of each of the profilers, with each of the inputs/outputs consistently named? In the example below/attached, changing X2 in either profiler updates it in the other profiler, while changing any input on the left causes OutputA to update on the right (because it was updated as an output on the left).

 

Re: Purification Process modeling with several BU operations

Malcolms 100% on the mark. Since the experiments with the different unit operations weren't part of a structured experiment, then OLS methods are going to be a little rough.
They might have some luck using a tree-based data mining method, but its likely they also don't have very much data either. 
In generally the customer would be better off optimizing/characterizing each unit operation separately. In practical terms, they also ought to have a scale-down model that they could use for larger scale expeiments too.

 

 

Re: Purification Process modeling with several BU operations

Thank you all for somehow confirming that it is not a good idea to try to bring together disjunct experiments, and what might help them to get closer where they want to be. 

@Byron_JMP they have optimized/characterized each unit with a model they are happy with. They now want to bring that all together to gain more insight on how each step influences the other steps.

However, I told them already that without a consistent reference or ID to bring the data together (like the output of one is fed in the next ...) you may get something but you have no assurance that this is something resonable (very unlikely reasonable). 

 I like the graphical approach which brought me to the idea with SEM, however this again might be more a qualitative knowledge gained they could build on, and may do a final DOE over the different steps

@Jed, Like the idea, but it requires again that there are at least some common factors throughout the profilers, I cannot promise this is the case here. 

Thanks again, I didn't want to miss a way I have not thought of.  

Re: Purification Process modeling with several BU operations

Very interesting problem! As I understand it, there are already DOEs that unveiled the causal effects in the data. Is this customer constantly collecting data on their whole process? That is, observational data? If so, it is worth exploring the use of SEM as a machinery for doing causal inference with their observational data. Here's a quote from Bollen and Pearl (2013) on this topic:

"SEM is an inference engine that takes in two inputs, qualitative causal assumptions and empirical data, and derives two logical consequences of these inputs: quantitative causal conclusions and statistical measures of fit for the testable implications of the assumptions."

In the context that @martindemel has posed, the separate DOEs have already served the purpose of clarifying what the "qualitative causal assumptions" are.  If the customer has observational data for the whole process, then the information from those DOEs can be translated into a testable SEM that can derive "quantitative causal conclusions," such that the customer can explore how changes at any time in the process affects downstream variables.

It sounds like the existing data from the DOEs cannot be linked--if it could, then those data could also be used in SEM.

HTH,

~Laura

Re: Purification Process modeling with several BU operations

This publication might be of interest for that topic. Lifecycle DOE, an approach to augment designs in each unit step instead of having independent DoE per unit.  https://www.mdpi.com/2306-5354/11/11/1089