One recipe for all
Updated: Jun 30, 2021
Who doesn’t like rice? My favorite is basmati rice and I have the perfect recipe for it: 1 cup of basmati and 1 ½ cups of cold water, boil for 12 min, let it rest for 10 min and done! My friends, however, prefer rice with peas which adds some uncertainty to my perfect recipe: The ratio between rice and peas can vary. Frozen or fresh peas influence the cooking time differently. Maybe I need to add the peas later to avoid undercooked rice or overcooked peas.
We could try different recipes, right? But now let’s image we cook for 100 people and one grain of rice is worth 10 $. Now we are getting closer to the challenges that the biopharmaceutical industry is facing. We need to get the recipe right as early as possible to obtain a method to tell us how long we need to boil the peas and rice considering all the variables.
Now what has this got to do with ultrafiltration?
In bioprocessing we deal with a similar challenge. We know how to process the product under ideal conditions, but we also need to know how variations influence the outcome.
Let’s say we want to concentrate our protein of interest using ultrafiltration. We have chosen the membrane cut-off, membrane size, buffer composition, pressure and flowrate to concentrate the protein quickly and without losses. In many cases, however, other proteins (like HCPs) are also present, and they influence the process duration, membrane fouling and selectivity. Different levels for HCPs due to upstream process variations in combination with a large number of possible operational parameters add another layer of complexity.
What influences the ultrafiltration process?
Qualitatively, the influences of process parameters are well understood. Ultrafiltration processes with high protein concentrations require more time than low concentrations and we can counteract this prolonged process time by increasing the pressure and flowrate. The mechanism how proteins influence the process time can be different.
In general, the target (protein of interest or other API) increases the concentration gradient on top of the membrane but does not enter the pores if chosen correctly. This effect becomes visible when the protein of interest is highly concentrated - >50 g/L. Process impurities, on the other hand, often cause membrane fouling (blocking of the pores). In our experiments we simulated membrane fouling in a two-component system and < 1 g/L of impurities were sufficient to increase process duration drastically.
Additionally, we observed different influences on the process when only processing the target component, the simulated impurity or both.
How can we account for the variations?
When the outcome of a process depends on many variables and the mechanism is not sufficiently understood, machine learning methods in combination with a well-chosen design space are promising solutions to describe the process. If sufficient data is available, they can very accurately predict the correct process behavior for many different conditions. However, biopharmaceutical experiments are expensive and therefore the sole application of pure machine learning approaches are limited. Contrary, mechanistic models are often the best approximations of the true process behavior, but the underlying assumptions of their structure do not cover the true process behavior in many more complex cases, hence they also become inaccurate. In this case, hybrid models - a combination of machine learning and mechanistic or user knowledge – are preferred. Hybrid models enable for powerful predictions and fast adaptations to different products while reducing experimental effort based on process knowledge.
So, whether you want to quantify the influence of different impurities during ultrafiltration or want to get your rice with peas, carrots or tuna done at first try, hybrid modeling is a promising candidate.
For more information on hybrid modeling in filtration check out our open-source publication.