MILLIONS OF PATIENTS HAVE BENEFITTED FROM NEXT GENERATION DRUGS to recover from cancer, autoimmune disorders, and many other diseases. These drugs are produced using biomanufacturing technologies. The biomanufacturing industry is growing rapidly and becoming one of the key drivers of personalised medicine and life sciences. As such, the global biomanufacturing market is projected to reach $388 billion in 2024.
Despite its success, biomanufacturing is a challenging business. It is cost intensive with high risks of failure. In addition, biomanufacturing methods use live systems (e.g., bacteria, virus, insect cells, etc.) during the production process. This enables highly complex and unique active ingredients to be generated compared with conventional drugs. However, the use of live systems leads to unique production challenges related to predictability, stability, and batch-to-batch variability.
the project used techniques such as machine learning and simulation-optimisation to predict and control biological systems
To date, several industries have benefited from Operational Research (O.R.) methodologies to improve operational efficiency, and reduce costs and lead times. However, the applications of O.R. methodologies to the biomanufacturing industry are still immature. One of the main reasons is that the competitive advantage in biomanufacturing used to be mainly driven by the scientific advances related to the underlying biological and chemical dynamics. However, with the growing market competition, the industry encounters an increasing need for a data-driven, O.R.-based approach to improve business practices.
A multi-disciplinary team of researchers from Eindhoven University of Technology (TU/e) and Merck Sharp & Dohme Animal Health (MSD AH) have been collaborating over three years to improve biomanufacturing efficiency. The team consists of Dr. Tugce Martagan, Prof. Dr. Ivo Adan, and Ph.D. candidate Yesim Koca from TU/e, and Oscar Repping, Bram van Ravenstein and Marc Baaijens from MSD AH. The collaboration resulted in a portfolio of optimisation models and decision support tools to reduce biomanufacturing costs and lead times. The project combines operations research and life sciences. This is one of the first examples showing how operational research can improve biomanufacturing practice.
Overall, the project used techniques such as machine learning and simulation-optimisation to predict and control biological systems. A variety of optimisation models and decision support tools were developed to translate the underlying biological dynamics into business metrics, such as lead times and costs. Although the project was conducted and implemented by MSD AH, its true impact extends to other biomanufacturing companies, including human health applications. This is mainly because the project addresses common industry challenges related to predictability, batch-to-batch variability and planning under uncertainty. In general, the project consists of three parts: reducing bioreactor changeovers, increasing fermentation yield, and creating better production plans.
HOW TO REDUCE CHANGEOVER TIMES IN BIOMANUFACTURING?
A typical biomanufacturing process can be broadly classified into two main steps, fermentation and purification. Fermentation is often carried out in bioreactors (e.g., highly controlled stainless steel vessels for a cell culture to grow). After fermentation, the batch proceeds with a series of purification operations (e.g., centrifugation, filtration, chromatography, etc.) to achieve good quality standards. The main focus of the first project is fermentation. During fermentation, the cell culture follows a specific growth pattern: The fermentation starts with the lag phase with no cell growth. The cells start to slowly grow in the acceleration phase, and continue with the exponential growth phase where the cell growth speed reaches its maximum. Then, the cell growth slows down through the deceleration and stationary phase, followed by the death phase. The cells produce the final product (e.g., proteins, antibodies) as they grow. Typically, the batch culture is harvested during the deceleration or stationary phase.
After the batch is harvested, the bioreactor needs to be cleaned and sterilised to be prepared for a new batch. These changeovers are costly and time consuming. For example, it might take one full work day to clean and sterilise a bioreactor. Therefore, there is a significant business case for reducing changeover times in biomanufacturing.
To reduce the bioreactor changeover times, MSD AH developed a new replenishment technique called bleed-feed. With this technique, some predefined fraction of the culture is extracted (to be sent downstream for further processing) and a special fresh medium is added to the remaining culture. The remaining cell culture acts as a new seed for the next bioreactor run. This means that the technique allows the changeover activities for the subsequent batch to be skipped. However, the technique can be performed during the exponential cell growth phase only. Otherwise, the technique does not work, and the culture needs to be harvested in full. In this setting, identifying the best bleed-feed time is challenging because the duration of each cell growth phase is stochastic. This means that we do not know the exact moment when the exponential growth phase stops. If the bleed-feed technique is carried out too early, then we might not achieve the maximum yield from that batch. In contrast, if it is conducted too late, then the batch needs to be harvested and the bioreactor needs to be set up for the next batch.
The proposed configuration has enabled an increase in the bioreactor yield by 50% while reducing the yield variability by 25%
To address this problem, a stochastic optimisation model was developed using the theory of Markov Decision Processes. The objective of the model is to identify the optimal bleed-feed time that maximizes the expected profit. The optimisation model captures the complex dynamics and uncertainties related to the underlying biological dynamics. Then, the model links the underlying biological dynamics with business metrics, such as cost and throughput. Prior to the optimisation model, the bleed-feed decisions were made based on domain knowledge. The optimisation model provided a data-driven, quantitative approach for decision-making, which improved process efficiency.
HOW TO INCREASE FERMENTATION YIELD?
In this project, the main objective was to maximise the production yield obtained from fermentation. This was especially critical for the Boxmeer facility in the Netherlands, seen in the photograph at the start of the article, as one of their bioreactors consistently produced a lower amount of yield compared with others in the facility. Interestingly, the bioreactor with the lower output was new, using state-of-the-art technology. Further investigation showed that this new technology used a different bioreactor mixing mechanism, leading to a different type of airflow inside the bioreactor. This implied that some of the controllable input parameters needed to be adjusted for that specific bioprocess (specific parameters are not disclosed for confidentiality). However, there was no available information in the literature to help decision-makers define the best configuration for these controllable input parameters. This implied that several experiments needed to be conducted at industry-scale to collect data and optimise the controllable input parameters. However, industry-scale experiments involved very high risks due to process uncertainties, limited resources, and high operating costs. Therefore, a smart decision-making mechanism was needed to identify the best parameter configuration through minimum number of industry-scale experiments. Although this problem was motivated by a case study in the Boxmeer facility, it addresses a common industry problem in the context of optimal learning: industry- or laboratory-scale experiments need to be designed in a smart way to collect the required information in the least possible number of experiments. This is especially critical when the resources for conducting these experiments are limited because of budget restrictions or operational constraints.
To address this problem, predictive models were built, based on the theory of Bayesian design of experiments. These models used limited amount of industry data to predict how bioreactor yield would change as a function of the critical process parameters. Then, a stochastic optimisation model that belongs to the class of the optimal stopping problem was designed to control these critical process parameters. As a result, the model suggested an optimum process configuration based on the results of eight industry-scale experiments. The proposed configuration has been implemented for more than one year at MSD AH and enabled an increase in the bioreactor yield by 50% while reducing the yield variability by 25%. This also helped to improve the environmental sustainability of these processes through higher production outcomes per bioreactor run.
HOW TO DEAL WITH PLANNING AND SCHEDULING?
Biomanufacturing operations are performed by highly skilled scientists using specialized equipment. Capacity planning for these limited resources is critical for successful and timely completion of orders. Failing to satisfy delivery dates results in loss of credibility and reputation. In addition, biomanufacturers face unique challenges in operational scheduling. Each client order requires several tasks to be completed. The tasks and their durations differ between orders depending on the final product requirements and quality of the starting material. The use of live cells often introduces ‘no-wait’ constraints between steps. The engineered nature of these products adds uncertainty at each step and imposes simultaneous requirements on highly skilled labour resources and specialized equipment to guarantee the best outcome. Creating a good schedule that can quickly react to these dynamics is a challenge. Despite these challenges, there is a need for a specific rhythm in the production system.
The project created novel tools to complement life sciences with operational research
To address these planning and scheduling challenges, we first developed a simulation model of the biomanufacturing operations. The simulation model was developed with the Arena software, and contained 8000 connections representing 48 different products with their unique routings on 25 pieces of equipment and more than 50 process steps. The simulation model was validated using two years of historical production data on lead times, utilisations, bottlenecks, inventory levels, and throughout. Then, simulation-optimization was used to generate a portfolio of flexible production schedules for each week (namely, rhythm wheels). The optimisation module used the Tabu search algorithm to maximise the throughout. The tool was designed to enable MSD AH to quickly react to changes in their production system by dynamically adjusting their production schedules. In addition, the tool was also used to evaluate and justify capacity expansion decisions.
INVENT. IMPACT. INSPIRE
‘The project greatly serves towards our motto: Invent. Impact. Inspire.’ says Bram van Ravenstein, Associate Director at MSD AH. More specifically, it created novel tools to complement life sciences with operational research. The developed tools have been in-use at MSD for almost two years. As a result, the production outcomes of certain batches from the Boxmeer facility increased by 97% without investments on additional resources, such as equipment and workforce. Recently, several follow-up projects were initiated nationally and internationally. For example, the Boxmeer facility is currently collaborating with several other facilities to help them use the O.R. tools. The upper management has recently encouraged new initiatives for knowledge transfer to the human health department. As more companies embrace the applications of O.R. methodologies, the impact will be significant for the society – cheaper and faster access to life-saving treatments. As Oscar Repping, Executive Director at MSD AH states, ‘We [the industry] will benefit from O.R., as such, we will avoid investments, we will become more predictive, leading to cost reduction, leading to more capacity on our production lines, meaning that we can make this world a better place.’
the production outcomes of certain batches from the Boxmeer facility increased by 97% without investments on additional resources
Additional information
Notes on contributors
TUGCE MARTAGAN
Tugce Martagan is an Assistant Professor in the Department of Industrial Engineering at Eindhoven University of Technology. She received her Ph.D. in Industrial and System Engineering from the University of Wisconsin-Madison in 2015. Her research focusses on stochastic modelling and optimisation, especially in the context of the pharmaceutical industry and healthcare operations management.