How to cut timelines and boost efficiency in bioprocess development: Unleashing the power of hybrid modeling and process intensification

As the name suggests, hybrid modelling is a combination of mechanistic and data-driven models and is more flexible in learning mechanistic links from data. Mechanistic models, also known as “white box models”, are based on the first principles, able to explain the underlying biological and chemical phenomena involved. Usually none of the models employed in biopharmaceutical manufacturing are purely based on first principles as reaction mechanism and binding mechanisms are not known a priori from first principles. All these so called “mechanistic models” are based empirical equations, such as the Monod kinetics, to assess the reaction mechanism. Data-driven models are based on the available, experimental data generated from the process. 

A typical hybrid model workflow comprises of two components:

1. Estimation of specific reaction rates using empirical information i.e., data-driven approach 
2. These estimated values are fed into the mechanistic models for further analysis/processing as shown in Figure 

Thus, hybrid models allow the combinatorial effect of mechanistic and black box models to gain better insights into the overall process thereby leading to robust and highly predictive models. 

Process intensification

All our systems consist of two types of variables 

• Process parameters that are controllable over time (e.g. temperature, feed, elution gradient, etc.) 
• States: anything that is a result of the impact of the process parameters (VCD, viability, titer, metabolites, elution profile of impurities, etc.)

Given the initial conditions (e.g. initial VCD, metabolites etc.) and controllable process parameters over time, we can predict the states. 

If we want to vary process parameters over time, e.g., feed profiles, we need to introduce phases in classical linear regression (e.g., Design of Experiments (DoE) design). This is also called unfolding of a time series as shown in Figure 4. In classical DoE setting we need one more run than effects in the model. Hence, if we want to incorporate the effect of all phases in the model, we need at minimum n+1 number of runs, where n is the number of phases. If we assume independence of the phases, i.e., no memory effect from one phase to another that impacts on the mechanistic relation of the next phase, we only need one run to calibrate a dynamic model. 

Since we now know the individual benefits of dynamic modelling and process intensification, it’s necessary to club these methodologies and leverage the potential to achieve the larger goal of developing a process with a significant drop in developmental timelines and making the drugs available to patients as fast as possible.   

To set up this architecture, below are the steps to be followed for the general workflow as outlined in Figure 4.

1. Data collection and preprocessing:
   - Overarching dataset: Collect data on process parameters and states to be modelled.
   - Data cleaning: Ensure the data is accurate and relevant and check for potential outliers.
   
   
2. Implementing data-driven techniques:
   - Model training & Validation: make use of various algorithms like OLS, ANN, etc. to build a good fit predictive model that describes the reaction rates as a function of states and process parameters.
   
   
3. Developing mechanistic models:
   - Parameter estimation: Employ a previously trained data-driven model from the previous step to estimate the parameters like specific growth rate, product formation rate, etc. for the mechanistic model.
   - Model integration & validation: Integrate these estimated parameters into the mechanistic model and validate the model for better prediction and accuracy.
   
   
4. Integration & optimization with process intensification:
   - Initial hybrid model: Run this hybrid model using experimental data and check for model predictions and accuracy.
   - Process intensification: optimize the hybrid model, specifically focusing on parameter combinations which may improve the model’s overall predictability.
   
   
5. Iterative refinement:
   - Feedback loop: Execute an iterative procedure in which the forecasts of the hybrid model direct the subsequent experiments.
   - Continuous optimization: the hybrid model is based on the results, gradually augmenting precision and effectiveness.

To shorten the development times of processes in early stages, the application of dynamic modeling through mechanistic or hybrid modeling is a promising solution. By harnessing the capabilities of process intensification to streamline experimentation and combining it with the intuitions derived from hybrid models, researchers can unlock unparalleled efficiency in bioprocessing.

This collaboration not only significantly reduces developmental timelines, but also ensures the development of robust, scalable, and optimized processes that yield maximum output.

In a world where time is of utmost importance, the integration of these state-of-the-art methodologies is not merely an option; it is imperative. As the bioprocessing industry continues to evolve, embracing these innovations will be the key to maintaining a competitive edge and ushering in a new era of rapid and efficient bio-process product development.

If you have any questions or need further assistance, we are here to help. 

Use our contact form, or get in touch with one of our experts at: info.pasx-savvy@koerber.com

We look forward to hearing from you.