By Gareth Macdonald
Optimizing filtration in biopharmaceutical production is a headache for process engineers. So why not let an artificial brain take the stain?
That’s the idea put forward by Marcel Ottens, PhD, and his team in the department of biotechnology at Delft University of Technology in the Netherlands. They report that current filtration fine tuning methods are sub optimal in a new paper.
Ottens says “These [current methods] are mostly experimental in nature and require a lot of lab scale work, using Design of Experiments (DoE) and statistical approaches, and even then, one can never be sure if the ‘true’ optimum has been found in this vast design space,” Ottens says.
Mechanistic modeling–a mathematical method for describing the elements of a system and their actions–has the potential to streamline the optimization of downstream processes using computer simulations. The tricky part is that before building a model, all the physical-chemical parameters of each element in a system need to be obtained, which is a time-consuming process.
Artificial neural network
To address this, Ottens and colleagues, developed a hybrid optimization approach that uses a screening method called high throughput experimentation combined with an artificial neural network (ANN)—a machine learning method able to spot patterns in complex data.
“Mechanistic modelling might be time consuming and speed limiting if used in multi-parameter optimizations,” according to Ottens. “We use ANNs to describe the behavior of products and impurities, for example, host cell proteins, during a chromatographic purification run.
“These ANN models need to be trained, and we use our mechanistic models for that. This allows us to do the optimization of a purification sequence in a fraction of time otherwise needed for mechanistic modeling-based optimization.”
In the study, Ottens and colleagues cite resin optimization as an example of another application of the approach, explaining the ANN-based method identified the optimal capture material 6.4 times faster than a mechanistic model alone.
Such a time saving is likely to be of interest to drug companies seeking to streamline process development operations, notes Ottens, who adds that the approach can be relatively straightforward to implement with the right knowhow.
“One would need simulation software. This would indeed require some specialist knowledge, but we are open to collaborations with such companies to help implement these advanced process development approaches,” he says.