Advanced process control in biopharmaceutical production

Process control of bioreactors seeks to influence the individual complex intracellular reactions of billions of cells by controlling their extracellular environment.

Process models form the basis for most of the modern or advanced control strategies. The discussion here will focus on model-based control for mammalian cell processes.

The following figure illustrates a nonexhaustive categorization of process models. They can be delineated into qualitative, mathematical and statistical models. Mathematical models can be further subdivided into mechanistic and black box models. For control applications, generally, quantitative models are required.

As such, the most useful model types for the control of bioprocesses are first-principle models, neural-network-based models and PCA- or PLS-based multivariate statistical models.

The first-principle models discussed here refer to engineering, rather than biological, first principles. They combine mass balance, rate and yield equations to describe the dynamic profiles of biomass, nutrients and metabolites.

When these models move from treating the cell as a single unit to more detailed levels, considering for example, amino acids, nucleotides, proteins and lipid pools, the complexity, time for development and required computational power increase greatly.

To replicate the relationship between substrate and metabolite concentrations, cell growth and product formation, it is not necessary to use expensive theoretical models. Instead, parameters within the series of equations can be optimized in order to fit the equations to experimental data.

Typically, these semi-empirical models have some extrapolation capability and so first-principle models of this nature are most useful for optimization and control applications. Some experimental data is required but it is usually of the order of three to five batches depending on the number of parameters and range of operating conditions within the model.

Neural networks involve a type of black box modeling, where little or no understanding of the underlying mechanisms of the process is required. The neural network is programmed to predict an per-output
or outputs based on a set of inputs by training it with a set of experimental data and/or deterministic models.

The neural network learns by processing data without rules. At its core it is a nonlinear regression. A series of data inputs are applied to a number of functions or nodes, which form a layer within the neural network. The inputs are weighted, summed and then processed by a transfer or threshold function before being output. This output may be the final result or simply data to be passed on to further layers within the neural network. During training, the weights applied to each of the inputs are adjusted in order to minimize the difference between the predicted and actual results.

The development of a robust neural network model requires a much greater number of batches than first-principle models, typically of the order of 20 to 30.

The quality of the experimental data is also very important. It is necessary to identify all inputs that affect the output to be predicted, and also to avoid correlated inputs. The training data set must contain enough variability to span the entire operating region and the data must be preprocessed effectively to remove outliers and poor-quality data as well as to scale and normalize the inputs.

However, in the case of bioprocesses where there is a high degree of process complexity, the low level of process understanding needed can be a considerable advantage, particularly if there is access to historical datasets.

Traditional manufacturing industries have used univariate statistical process control (SPC) charts to monitor and improve their processes for some time.

In bioprocesses, many variables are recorded and can potentially have an effect on the output of the system. Multivariate statistical methods have been developed to allow similar analysis of these large datasets. Detection of abnormal conditions and root cause analysis can be conducted by applying Principal Component Analysis (PCA) and Partial Least Squares or Projection on Latent Structures (PLS) methods.

Generally speaking, PCA is most often used to detect batch abnormalities, while PLS can be used to predict output parameters such as product quality or batch endpoint based on input variables and the current process states.

Statistical models are data driven and, therefore, it is important that the set of good reference batches captures the variation in the process. Too little variation results in false alarms and too much variation causes the model to be insensitive. A set of ten or more batches is necessary to build a multivariate statistical model.

Once the effort to build a good process model has been expended, it is desirable to exploit that model to maximize the benefit. Model-based control strategies represent one potential application. There are numerous forms of model-based control including model predictive control (MPC).