732284_en-US

liliconcom

C81 - Determining optimum production settings in bioreactor production using Functional Data Explorer – Functional Input to Non-functional output

Typical user personas/situation slides

Job Title/Roles	Manufacturing, Product, Quality, Process Engineer	Upstream Scientist, R+D, Tech Transfer, Scale up
Industry	Pharma, Chemical, Biotech
Critical Business Issue	Find cause of repeated failures in production	Improve productivity of R+D process by X%	Improve productivity of R+D process by X%
Problems/Reasons	Can't find root cause to production failures, process steps are changed over time and don’t have ability to look at it in one ‘big picture’, have lots of historical data and no tool to analyze. Have sensor readings over time that may give more detail, but too much detail to look at with graphing.	New process with no prior subject matter knowledge, attempting to run DoE’s on process factors but the sensor readings are hard to interpret. Haven’t got a good understanding of how the process looks whilst running and how that relates to production.	Batch process gives minimal improvements in new process, need to move to fed-batch. No idea of suitable feeding parameters.
Specific Capabilities	Identify the ideal and non-ideal sensor readings that need to be maintained to get high levels of production.	Identify the changes in sensor readings that are associated with productivity changes	Identify the best feeding parameters that maximise productivity
Business Value	Reputation improvement, supply chain restored, reduced down-time during investigations, avoiding auditor shut-downs. Meeting deadlines/targets.	Novel skillset for customers, faster time innovation, higher productivity levels, easier scale up and proof-of-concept	Faster time to project completion. Improved productivity. Easier transfer to scale up.
User Value	Reduced 'all hands on deck' panics, quicker resolution to get back to main role, improved credibility, improved morale. More confident in approach. Short learning curve.	Better process knowledge, better perception. Novel skillset.	Better knowledge for scale up. Improved perception by management.

 

Terms to look out for: scale up, bioreactor, microbial, mammalian, ambr, ambr250, ambr15, small-scale, process development, upstream.

Context

The work here is a continuation of this demo on using an FDOE for bioreactor production development, taking the methods applied by Clariant in this Discovery talk. In this case I’m recontextualising the process here to focus on how the vast amounts of sensor data can be gathered and used to predict the final product level. This can apply to prospects in two contexts:

• Manufacturing level – production processes will naturally vary in their sensor readings and functional control of their parameters (i.e. how well they maintain pH, stirring, temperature), which will lead to changes in the final production level – prospects can gather all of this information to investigate and identify critical process parameter settings over time or indicators of poor/good performance
• Experimental level – the application of DoE (changing the production parameters/media conditions) will naturally cause a change in the way that the bug acts over time, changing its respiration (dissolved oxygen), number of cells (biomass) and by-product production (pH) readings – knowing how these parameters change adds more process understanding on top of the information they are gaining from a standard DoE.  Bioreactors always gather sensor readings, they just aren’t often in detail.

The key message/question to challenge users is ‘You have all these sensor readings being gathered whilst you run production/experiments, why don’t you use them?’.

JMP has great potential to leverage the FDE analysis for better process understanding – currently there are no other commercial softwares capable of providing the level of insight possible with FDE. Current approaches take heavy understanding of statistics and coding, which your average user (engineer/scientist) is not going to have/want to have. This is the kind of tool I would have given my left leg for back when working on bioreactor process development and would have significantly shortened our project times!

 

 

Last Thing First

 

This video is currently being processed. Please try again in a few minutes.

(view in My Videos)

 

The final dashboard shows how you can take changes in the sensor readings and see how they relate to the final production. The right profiler gives us an understanding of how productivity is related to certain shape components for each sensor reading. Increasing the value of DO FPC1 leads to increase in productivity, which we can see on the left corresponds to a bigger dip in the dissolved oxygen from 50h onwards.

This dip is a great way to explain how this knowledge improves process understanding – the dip shows that the process is demanding more oxygen after 50h in order to be more productive. This can be used as a marker of quality (i.e. if it happens, we’re going to be productive) or to indicate that process improvements can be made (i.e. should they add even more agitation/aeration as the cells want more oxygen at this time).

You can then use the desirability functions to show how the sensor readings should appear in high-production and low-production settings. Whoever is running this process can use these as the ‘poster child’ sensor readings and go back to production and tell them to look out for these profiles.

Demo overview (the ‘Do It’ bit)

1. Fit the functional model of time vs. sensor readings (biomass, pH, DO). ID with the ‘batch’ column.
   1. Include the final product in the supplementary column – this is important for it to be carried over into the summary table later

 

2. Fit the appropriate model to each sensor reading to best fit the data (pH and biomass = P-spline, DO = wavelets).
   1. The DO sensor reading requires cleanup with the Savitzky Golay filter – this is a good opportunity to ask customers if they have noisy data they struggle with
   2. For demo purposes (to make a cleaner profile) reduce the number of shape components to 4 each – more than this and we risk making the tool look too complex for new users.

 

 

1. It’s important to give some rudimentary description of what the shape functions and FPC scores mean to the customer for the next steps to make sense – changing the FPC1 amount on pH gives a nice way of showing it adding more/less of a ‘bump’ at ~20h.

 

 

3. Save summaries – This extracts the shape functions and FPC scores (alongside the final productivity levels) for each sensor.

 

4. Fit Model – fit the FPC scores as the inputs to the process. In this case I’ve used GenReg to fit the FPC terms and selected a model that gives a good RSq and doesn’t select every model term – this is important to a) highlight that not every FPC score is important to production and b) avoid having the final profiler be too busy.
   1. Once complete, save the prediction formula for the FPC vs. production level.

 

5. Create the profilers of A) FPC scores to Productivity and FPC scores to Sensor Readings)
   1. – Keep the two profilers separate – this stops there from being FPC scores that are out of the bounds of the linear productivity model that can lead to negative production values (breaking the illusion in the demo).
   2. To help illustrate, I’ve colour coded the sensor readings to the FPC scores.

 

 

 

 

Case Studies