I came across this article (sciencedirect.com/science/article/pii/S0260877425000020) and found it really interesting. I was wondering if anyone here has experience with something similar, that is, building a digital model of the fermentative behavior of a strain using both online bioreactor data and offline sampling.
For my PhD, I’m working on predictive models, currently correlating OD data from the Pioreactors with the concentration of the metabolite of interest, which typically varies with pH and temperature. The idea of creating a virtual environment that can be continuously fed with experimental data sounds amazing to me!
The paper is behind a paywall, but the pioreactor UI/System is technically a digital twin in my opinion. The term is usually used in large industry to tell you the states of any actuators, readings of sensors, etc.
The genetic algorithm they mention is one of many optimization algorithms used to solve multivariable optimization problem. It has nothing to do with genetics btw it is just that the algorithm mimics evolution by altering input variable values similar to how organisms mutate in evolution.
I am somewhat skeptical on the prospect of running an optimization script (or “AI”) directly on a digital twin to try to optimize output. For example have it alter variables like PH and temperature through their relevant actuators during the batch growth cycle. This is because these biological systems are quite complex and in a sense are optimizing themselves by literally mutating on a cellular level depending on the reactor conditions.
What do you think their setup from the paper has that the Pioreactor does not have?
Thanks for your reply. That is so cool. I’m quite new to this topic, so my question was more about whether it would be possible to build a simpler, more minimalistic model that I could train by inputting data such as the carbon source at a certain concentration, the composition of the culture medium (nitrogen, vitamins, minerals), temperature, and a specific strain. The idea would be for the model to simulate a fermentation process aimed at producing, for example, a target of x g/L of lactic acid. Ideally, all the information would come from a black-box approach. I’d first run many fermentations with the strain of interest, systematically varying the parameters, record all the results, and then use those data to train and test the model, assuming in a simplified way that no lab-directed evolution or spontaneous mutations occur.
I get what you’re trying to do. Sounds like your objective function is a target lactic acid concentration which I assume you have a way to test for in the lab. I would suggest taking a crack at the taguchi array optimization method to test a wide range of fermentation parameters. This video is a cool explanation on how its used:
While running these experiments you might come across patterns that you can model mathematically that would make a simple model. Or you can combine all the information to try to make a predictive model.
Make sure you test across all parameters in your control that you could suspect to have an effect on the objective function.