(701a) Application of Hybrid Modeling Approaches Towards Accelerated Process Characterization

The Quality by Design paradigm in pharmaceutical development calls for detailed process understanding that is typically achieved by intensive experimentation especially for poorly understood processes. In silico modeling is identified as a potential tool to minimize the cost, labor and time involved in experimentation leading to more efficient use of resources. This goal of reducing the process development cycle has led to intense research efforts in the area of model-based experimental design across the pharmaceutical industry and academia.

However, most of the published model-based experimental design frameworks for biopharmaceutical applications require significant amounts of experimental data as well as considerable technical expertise and/or proprietary software. For early-stage development where process understanding is poor and experimental data are fragmented and limited, it can be challenging to utilize and implement these modeling frameworks.

The case study presented here focuses on identifying experimental design spaces (parameters and ranges) for multivariate process characterization studies utilizing in silico models developed using data generated via an initial set of univariate process characterization experiments performed on a CHO perfusion process. The modeling framework is easily transferable to other bioreactor setups and fits into existing experimental process development workflows with minimal modifications. This framework enables smart recommendations for future data collection efforts, iterative model-assisted design of experiments, as well as effective process optimization and control.

In our framework, we utilize mass-balance based multivariate data analytics and gaussian processes to gain initial understanding into the process, calculate specific rates of metabolite consumption and waste generation, as well as assess batch-to-batch variability. We then incorporate information from these first analyses into a hybrid-model structure for prediction of titer and selected key critical quality attributes as a result of changes in process levers. The model also facilitates process optimization and provides insight into areas of the design space that would benefit most from further experimentation (i.e., areas of high variability or poor predictive performance), setting the stage for iterative model-assisted process characterization

With relatively limited experimental data, our framework provides critical insights into factors influencing the process as well as decent predictive capability on several critical quality attributes of interest. We are also able to analyze the impact of different feed schemes, and operating conditions on process performance. Most relevant for process characterization, we use the framework to identify parameter combinations to focus experimental efforts on.

This framework demonstrates the benefit of introducing modeling approaches to in-flight process development efforts with limited experimental data.