Advanced modeling of industrial-scale fermentation process for antibiotic production

Atli Freyr Magnusson

Research output: Book/ReportPh.D. thesis

445 Downloads (Pure)

Abstract

Biotechnology plays an integral role in the modern economy and is responsible for producing various bulk and specialty chemicals critical to the function of society. The bioprocess involves using microorganisms as a miniature factory to convert substrates into valuable molecules. Antibiotics are antimicrobial substances used to fight bacterial infections and are commonly produced via fermentation. As global antimicrobial resistance becomes a larger and larger threat to humanity, there has been an increased interest in the potential of old-generation antibiotics to address the current need for new antibiotics. Faced with increasing demand, we need to modernize the current production methods. The chemical and biochemical industry is transitioning through the fourth industrial revolution, or Industry 4.0, which is the joining of technologies that blur the lines between the physical, digital, and biological worlds.

Digital twins based on process models are a crucial technology for industries in the changing competitive landscape following the change to Industry 4.0. Digital Twins are a digital representation of a real-world physical product, system, or process. This technology’s foundation and primary enabler is a mathematical model that accurately captures the relevant physical and biological phenomena. However, modeling fermentation systems is challenging due to biological complexities. Furthermore, the pharmaceutical industry has always had a strong emphasis on quality. Legal regulations specify the required purity of Active Pharmaceutical Ingredients and place limits on potentially harmful or efficacyreductive impurities. These impurities may be byproducts of the bioprocess itself and may have very similar chemical and physical properties, making them impossible to separate in the purification process. So far, bioprocess modeling has focused on the ability to predict the productivity of fermentation with little to no focus on product quality.

The objective of this project was to accelerate model development implementations by developing state-of-the-art mathematical models that can analyze and simulate the Fusidic Acid fermentation process. Due to its status as an antibiotic, there is a growing need to improve the yields of Fusidic Acid at the industrial site hosted by the LEO Pharma A/S. However, any attempts to push productivity can lead to increased byproduct formation, which needs to be tightly controlled, especially in a pharmaceutical setting. The developed models are designed with the end goal in mind to predict the harvest of Fusidic Acid and the formation of a particular byproduct that is extremely difficult to separate. These models can then be applied to gain a deeper process understanding, test new process conditions, and be used as soft sensors. Data is needed to build the models, and one of the essential variables used in virtually all fermentation models is the concentration of cells in the medium. More specifically, the concentration of living or viable biomass.

This thesis describes how two models were built to predict concentrations of the main product and the byproduct accurately. One of the models is a purely statistical model that uses batch data collected in real-time from a currently active production and creates an Input-Output correlation. The novelty of the method is that it is the only tool that can directly model batches of various duration without a complicated preprocessing step known as batch trajectory synchronization. It was also the only chemometric model that could predict both batch productivity and quality, whereas traditional methods could only predict productivity.

The second model is a hybrid model, which combines the available scientific knowledge with machine learning in a synergistic way. Biological systems are highly complex, and it can take years of research to gather the available knowledge required to capture all the relevant process phenomena accurately. On the other hand, data-driven models do not require extensive knowledge but infer patterns and knowledge from data. Before developing the first principles model, an experimental procedure was designed to collect the relevant data. A new linear calibration methodology was discovered that allows the conversion of dielectric spectroscopy data to viable biomass concentration, which gives quick and reliable estimates of viable biomass. This information was then used to calibrate a mechanistic model that could accurately and reliably describe biological growth and main product concentrations. With the mechanistic approach, the predictive qualities improved significantly, with an average prediction error of 6.6%. Furthermore, the investigation into model uncertainties revealed that the model structure had limited uncertainty propagation when simulating the process, indicating that the model is a good representative of the physical system. Neural Networks were then directly integrated into the mechanistic model to find the hidden patterns that relate the current batch culture conditions to changes in byproduct concentrations. The final hybrid model is a kinetic description of the process bound by conservation laws that can describe the growth and consumption profiles of biomass, substrates, main product, and byproducts of the Fusidic Acid fermentation. Byproduct accumulation in the fermenter could be accurately predicted with an average prediction error of 22.8% throughout a full fermentation period. The primary indicator of batch quality is the concentration of byproducts, and this model can determine whether a batch meets quality criteria. Furthermore, simulations can adjust process conditions and discover substrate control methods that completely eliminate the presence of byproducts during harvest or find alternative scenarios that increase the batch outputs of the main product while still maintaining the quality criteria.
Original languageEnglish
Place of PublicationKgs. Lyngby
PublisherTechnical University of Denmark
Number of pages135
Publication statusPublished - 2023

Fingerprint

Dive into the research topics of 'Advanced modeling of industrial-scale fermentation process for antibiotic production'. Together they form a unique fingerprint.

Cite this