Latent variable modeling approaches to assist the implementation of quality-by-design paradigms in pharmaceutical development and manufacturing

With the introduction of the Quality-by-Design (QbD) initiative, the American Food and Drug Administration and the other pharmaceutical regulatory Agencies aimed to change the traditional approaches to pharmaceutical development and manufacturing. Pharmaceutical companies have been encouraged to use systematic and science-based tools for the design and control of their processes, in order to demonstrate a full understanding of the driving forces acting on them. From an engineering perspective, this initiative can be seen as the need to apply modeling tools in pharmaceutical development and manufacturing activities. The aim of this Dissertation is to show how statistical modeling, and in particular latent variable models (LVMs), can be used to assist the practical implementation of QbD paradigms to streamline and accelerate product and process design activities in pharmaceutical industries, and to provide a better understanding and control of pharmaceutical manufacturing processes. Three main research areas are explored, wherein LVMs can be applied to support the practical implementation of the QbD paradigms: process understanding, product and process design, and process monitoring and control. General methodologies are proposed to guide the use of LVMs in different applications, and their effectiveness is demonstrated by applying them to industrial, laboratory and simulated case studies. With respect to process understanding, a general methodology for the use of LVMs is proposed to aid the development of continuous manufacturing systems. The methodology is tested on an industrial process for the continuous manufacturing of tablets. It is shown how LVMs can model jointly data referred to different raw materials and different units in the production line, allowing to understand which are the most important driving forces in each unit and which are the most critical units in the line. Results demonstrate how raw materials and process parameters impact on the intermediate and final product quality, enabling to identify paths along which the process moves depending on its settings. This provides a tool to assist quality risk assessment activities and to develop the control strategy for the process. In the area of product and process design, a general framework is proposed for the use of LVM inversion to support the development of new products and processes. The objective of model inversion is to estimate the best set of inputs (e.g., raw material properties, process parameters) that ensure a desired set of outputs (e.g., product quality attributes). Since the inversion of an LVM may have infinite solutions, generating the so-called null space, an optimization framework allowing to assign the most suitable objectives and constraints is used to select the optimal solution. The effectiveness of the framework is demonstrated in an industrial particle engineering problem to design the raw material properties that are needed to produce granules with desired characteristics from a high-shear wet granulation process. Results show how the framework can be used to design experiments for new products design. The analogy between the null space and the Agencies’ definition of design space is also demonstrated and a strategy to estimate the uncertainties in the design and in the null space determination is provided. The proposed framework for LVM inversion is also applied to assist the design of the formulation for a new product, namely the selection of the best excipient type and amount to mix with a given active pharmaceutical ingredient (API) to obtain a blend of desired properties. The optimization framework is extended to include constraints on the material selection, the API dose or the final tablet weight. A user-friendly interface is developed to aid formulators in providing the constraints and objectives of the problem. Experiments performed industrially on the formulation designed in-silico confirm that model predictions are in good agreement with the experimental values. LVM inversion is shown to be useful also to address product transfer problems, namely the problem of transferring the manufacturing of a product from a source plant, wherein most of the experimentation has been carried out, to a target plant which may differ for size, lay-out or involved units. An experimental process for pharmaceutical nanoparticles production is used as a test bed. An LVM built on different plant data is inverted to estimate the most suitable process conditions in a target plant to produce nanoparticles of desired mean size. Experiments designed on the basis of the proposed LVM inversion procedure demonstrate that the desired nanoparticles sizes are obtained, within experimental uncertainty. Furthermore, the null space concept is validated experimentally. Finally, with respect to the process monitoring and control area, the problem of transferring monitoring models between different plants is studied. The objective is to monitor a process in a target plant where the production is being started (e.g., a production plant) by exploiting the data available from a source plant (e.g., a pilot plant). A general framework is proposed to use LVMs to solve this problem. Several scenarios are identified on the basis of the available information, of the source of data and on the type of variables to include in the model. Data from the different plants are related through subsets of variables (common variables) measured in both plants, or through plant-independent variables obtained from conservation balances (e.g., dimensionless numbers). The framework is applied to define the process monitoring model for an industrial large-scale spray-drying process, using data available from a pilot-scale process. The effectiveness of the transfer is evaluated in terms of monitoring performances in the detection of a real fault occurring in the target process. The proposed methodologies are then extended to batch systems, considering a simulated penicillin fermentation process. In both cases, results demonstrate that the transfer of knowledge from the source plant enables better monitoring performances than considering only the data available from the target plant.