José C. Menezes

The Preeminence of Multivariate Data Analysis as a Statistical Data Analysis Technique in Pharmaceutical R&D and Manufacturing

Abstract The Pharmaceutical Industry has large amounts of data available to it. Much of this data is available from analytical instruments which have the capability and precision to dig deeper into processes, and provide unprecedented understanding of these processes. The process of turning this information into knowledge could conceivably be carried out by a number of routes, including statistical data analysis. Although a number of statistical data analysis techniques could carry out this transformation, one technique; multivariate data analysis (MVA) has become the primary technique in this arena. Transparency, reproducibility and an ability to conform to regulatory directives have all contributed to this current situation, and this chapter explains how.

Monitoring mAb cultivations with in-situ raman spectroscopy: The influence of spectral selectivity on calibration models and industrial use as reliable PAT tool

Raman spectroscopy is a suitable monitoring technique for CHO cultivations. However, a thorough discussion of peaks, bands, and region assignments to key metabolites and culture attributes, and the interpretability of produced calibrations is scarce. That understanding is vital for the long‐term predictive ability of monitoring models, and to facilitate lifecycle management that comply with regulatory guidelines. Several fed‐batch lab‐scale mAb mammalian cultivations were carried out, with in situ Raman spectroscopy used for process state estimation and attribute monitoring. The goal was to evaluate its use as a process analytical technology (PAT) tool to detect residual glucose and lactate levels, understand their dynamics and interconversion, and eventually estimate key performance culture and product quality attributes. Glucose and lactate models were optimized up to 0.31 g L−1 with 3 Latent Variables (LVs) and 0.19 g L−1 (2 LVs) accuracy, respectively. Glutamine and product titer models, were not specific and accurate enough, even though indirect calibrations were obtained with a RMSEP of 0.12 g L−1 (4 LVs) and 0.29 g L−1 (5 LVs), respectively. A critical discussion and details about the extensive work done in calibration development and optimization are provided. Namely, considering a risk‐based selection of variability sources impacting sample spectra, executing designed experiments with spiked cultivations, and using advanced chemometric procedures for variable selection and model cross validation. A strategy is presented to evaluation Raman spectroscopy as a reliable PAT technology fit‐for industrial use.

Subspace identification methods for a fast dynamic model structure screening

Abstract Modelling multiple-input multiple-output petrochemical industrial dynamic systems is a complex task. Empirical models, based on linear state-space dynamic models often provide a sufficient degree of approximation in a statistically efficient way (i.e. with a small number of parameters). The use of subspace identification methods (SIM) proved to be an useful tool to estimate state-space model parameters since there is no need to specify the model structure prior to the model estimation task. However it is necessary to estimate the models order and to select the proper inputs for each state-space model. In this article, it is presented a method based on the combination of bootstrapping and subspace identification techniques in order to quickly test many model alternatives in a very efficient way. The proposed method is an approximated approach that can be used to pre-select viable model alternatives (supported by the observed input-output data).