December 18, 2014

Mechanistic Modeling: Does it Have a Future in Process Development?


The application of mechanistic models for process development and optimization has been a topic of discussion in the pharmaceutical industry for many years. Mechanistic models are considered as an alternative to a standard DoE type empiric approach. Most often, the DoE approach is preferred based on a cost-benefit evaluation where the time consumption, complexity, and lack of know-how on how to deploy are found too extensive when using mechanistic modeling. Novo Nordisk uses mechanistic models for the development of selected downstream process steps. Following his keynote presentation to attendees at the recent HTPD Conference in Italy, Arne Staby, PhD, Fellow, Novo Nordisk, Denmark, spoke to Process Development Forum about why he prefers mechanistic models, when they are appropriate to use, and how regulatory agencies currently view them.

What is the fundamental difference between mechanistic models and models based on Design of Experiments (DoE)?
Industry today is predominantly using statistical modeling – if they use modeling at all. You can get statistical models out of a DoE approach. That’s fine. In most cases, this works quite well. You can predict what will happen pretty well. However, if you want to have some physical meaning of the different parameters in your model, a statistical model does not provide any physical meaning to the parameters. A mechanistic model is typically based on the laws of mathematics, physics, and physical chemistry. So, the parameters in these models actually have physical meaning.

What is the benefit of having the information about physical properties?
If you have a mechanistic model, you know that all the correct parameters are part of your model if you have made the right assumptions. You know that nothing peculiar can suddenly happen because you have it all there. All of the important parameters will be described by your mechanistic model. Thereby, you can be much more certain that if you make a change, the system will do what you predicted. Statistical models based on QbD sometimes fail. When you set up a statistical model based on DoE, you don’t really know if you have all the right parameters in play and whether or not you’ve taken everything into account. You just can’t be sure.

Novo Nordisk has been having discussions with authorities about using mechanistic modeling in regulatory submissions. Could you tell more about your experience?
A couple of years ago, CDER invited industry for partnerships for discussions on how to prepare an application using QbD. Then representatives of CBER approached us and asked for a presentation on the subject too. They made us an “unofficial partner” of a project where we could attend meetings and receive feedback. This led to discussions with other regulatory bodies as well, such as EMA and PMDA. An essential part of the discussions with the regulatory agencies was process understanding and at that time, we had started using mechanistic modeling. So, we wanted to show them a case where mechanistic modeling had been applied for process understanding, and we also included another step where we used DoE. We were a bit anxious as to how they would receive a mathematical expression as a design space instead of a set of parameter ranges, which you typically get from a DoE set-up. To our surprise, they accepted the approach, and discussions were merely on the experimental level required to verify the model.

What is the biggest myth or misunderstanding about mechanistic models in the pharma industry and how do you dispel that myth?
Many think mechanistic modeling is too cumbersome. You need a critical mass of people who can make it happen. If you have just one expert in your company who can do this, it will not fly. It does take quite a bit of teaching, but once people understand the concept and its benefits, buy-in will happen. In fact, our internal experts often perform fewer experiments to set up the models than people need to do to establish parameter ranges based on OFAT and DoE.

When should a mechanistic model approach be used over the more accepted DoE approach?
Mechanistic models are not ideal process understanding tools for all types of unit operations. In my view, mechanistic models should be used whenever possible. For some unit operations, such as ion-exchange chromatography, mechanistic models are easily employed for prediction. However, when it comes to reverse-phase chromatography or HIC, the mathematical description of hydrophobic interactions between surface and proteins are not optimal today making it difficult to use the modeling approach.They just aren’t good enough. So, in those cases, we turn to DoE or a much more limited mechanistic model. With respect to crystallization, we would use DoE. However, for reactions, we always use mechanistic models.

What do you predict the future will be for mechanistic modeling in bioprocessing?
There is quite a bit of interest from industry about mechanistic modeling. We have tried helping some of these companies and are generally willing to share our experience. I see many benefits for these models to document and optimize processes as well as for troubleshooting when something is going wrong.

Read More Q&As With Industry Leaders Here

Tags: DoE, QbD, mechanistic modeling