Artificial intelligence the key to process understanding

Traditionally, pharmaceutical manufacturing has been accomplished using batch processing with quality control testing conducted on samples of the finished product. This conventional approach has been successful in providing quality pharmaceuticals to the public for a number of years. Significant opportunities, however, exist for improving pharmaceutical development, manufacturing and quality assurance (QA) through innovation in product and process development, process analysis and process control.¹ Other sectors, for example, manufacturing and chemical, have already adopted new ways of thinking. The pharmaceutical industry, through a number of regulatory initiatives, is now beginning to develop these concepts.

Since 2004, FDA, recognizing the need to promote a greater scientific approach to pharmaceutical development, has instigated new initiatives such as Pharmaceutical cGMPs for the 21st Century: A Risk-Based Approach and Guidance for Industry Process Analytical Technology (PAT) — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance.^1,2 In these documents, there is an emphasis on using new scientific understanding and innovative technologies to provide a science-based approach to pharmaceutical development. More focused and scientifically justified strategies will enable the pharmaceutical manufacturer to build quality into products rather than assure quality through testing to poorly defined specifications. The principles of quality by design (QbD) should be applied to generate process understanding, which must be underpinned by using multivariant models describing processes and their associated causes of variability.

Key points

There are many tools recommended within the PAT framework that enable process understanding for scientific, risk-managed pharmaceutical development, manufacture and QA.¹ These include multivariant tools for experimental design, data acquisition and analysis, process analysers (in-/on-/at-line) for the real-time monitoring of product quality, process control systems and knowledge management software applications. Much of the innovative thinking with respect to PAT and QbD is, however, focused on the development of novel probes/sensors for real-time and in-line monitoring of manufacturing processes.

This philosophy must be coupled with a greater understanding of the risks and causes of variability in pharmaceutical dosage form manufacture to achieve desired product quality and optimal clinical performance.

Design of experiments

Traditionally, tools for multivariant data analysis, such as principal components analysis (PCA) and partial least squares projections to latent structures (PLS), have been used in conjunction with design of experiments (DoE) to investigate the underlying relationships in pharmaceutical processes.³ New knowledge engineering technologies based on artificial intelligence (AI) have, however, in recent times been developed that offer some advantages compared with conventional multivariant tools and can be used to compliment existing technology.^4,5

This article aims to highlight the value of using AI technology to understand complex multivariant systems such as pharmaceutical formulations and processes. Used appropriately, these tools enable the identification and evaluation of product and process variables that may be critical to product quality and performance, and allow the generation of predictive models that link changes in important/critical variables to changes in key product quality attributes — and ultimately clinical quality and performance.

The reality

"Perhaps we don't hear much about AI methods used within today's technologies because it's slightly unnerving when computers emulate human thinking. Yet we, and computers themselves, continue to improve the way AI works quietly in the background to optimize, reduce process costs, and improve timing and product quality. For some tough, nonlinear applications, AI may be the only solution."⁶

In recent years, AI has become important in a number of fields in helping to make better use of information, increasing efficiency and enhancing productivity. Many UK banks have used AI technologies to detect fraudulent behaviour by analysing transactions and alerting staff to suspicious activity. This has reduced fraud cases for the first time in nearly a decade by more than 5% to £402.4 million in 2004.⁷ In bioinformatics, AI has been used to manage, discover and interpret information/knowledge from biological sequences and structures. In the manufacturing industry, AI is used to achieve automotive control of production processes.

The concept

AI is defined as the ability of a machine to learn and "think" for itself from experience and perform tasks normally attributed to human intelligence; for example, problem solving, reasoning and process understanding. AI technologies are widely used in situations where tasks are multidimensional, and where relationships are nonlinear and extremely complex; for example, the underlying relationships between formulation ingredients, process conditions and drug product quality.

A number of AI technologies have been developed and are available as commercial software applications, including neural networks, neurofuzzy logic and genetic algorithms (GAs). Artificial neural networks (ANNs) are computer-based programs designed to mimic the process of the human brain's cognitive learning. ANNs work by building up a network of interconnecting processing units or nodes, which are the artificial equivalents of biological neurones. The ANN is able to discover knowledge (e.g., cause–effect relationships) that is hidden in experimental data and generate predictive models linking important factors to key measurable outputs; for example the effect of drug particle size and wet granulation conditions on tablet dissolution.

Fuzzy logic is a powerful problem-solving technique based on the mathematical theory of fuzzy sets first defined in 1960s by Zadeh.⁸ It has applications in control and decision making, and derives its power from being able to draw conclusions and generate responses based on vague, ambiguous, incomplete and imprecise information. Neurofuzzy logic is a hybrid technology combining the adaptive learning capability of neural networks with the interpretation power of fuzzy logic. This provides a powerful tool for generating interpretable rules from complex and nonlinear data.

Genetic algorithms are adaptive heuristic search algorithms, which are loosely based on the principles of genetic variation and the Darwinian concept of natural selection. Together with neural networks, GAs provide an excellent means of optimizing multivariant systems by identifying and evolving solutions until the desired combination of properties (formulation ingredients or process parameters) giving optimum performance (e.g., disintegration, dissolution or uniformity of content) is found.

In science and engineering, AI technologies have been widely used for knowledge discovery and knowledge engineering. Knowledge discovery (sometimes referred to as data mining) is a process of automatically extracting hidden knowledge from experimental data. As a data-driven process, it requires no explicit prior knowledge and discovers relationships, trends and patterns based solely upon available data. Neural networks, neurofuzzy logic and GAs have been employed in this field and are known to provide some notable benefits to the end-user.

When using AI tools, no assumptions or hypotheses regarding the relationships in data are required to generate models. Strict statistical experimental design is also unnecessary, providing that the data adequately covers experimental space.

The tools can cope with missing and historical data, and can use data from several sources providing the format and measurements are equivalent, and that the major causes of variation are captured and quantified. AI methods are very good at dealing with complex nonlinear relationships hidden in multivariant data, and have the ability to generate intelligible and actionable rules. Genetic programming, which is a more recent development in AI, has been shown to generate mathematical relationships linking the major causes of variation with key quality indicators or outputs.⁹

AI application

AI technologies have been applied in the pharmaceutical field since the early 1990s. In particular, the advantages of neural networks compared with traditional statistical techniques as a method of dealing with complex nonlinear relationships in product formulation have been reported.¹⁰ In the past, neural networks, GAs and neurofuzzy logic have been used to obtain improved understanding of formulations, and assist with formulation design and process optimization. Applications based on these technologies range from predictive models for controlled and immediate release dosage forms through to process control for granulation, topical formulations and film coating.

In the early 1990s, neural networks were applied to model the in vitro release of hydrophilic matrix formulations.^4,11,12 More recently, neural networks have been used to predict in vitro dissolution rate as a function of product formulation changes and to optimize the release profile of theophylline tablets comprising fast and slow release fractions.^13,14

A number of researchers have used neural network technology to model the relationships between formulation composition, process conditions (e.g., compression force) and tablet properties (e.g., tensile strength, disintegration time and dissolution profiles) for conventional immediate release tablets.^15–17 ANNs have also been successfully applied to model and optimize different immediate release tablet formulations, such as rapidly disintegrating tablets, high-dose plant extract tablets and intact tablets.^18–20

For pharmaceutical solid dosage forms, wet granulation is an important and complex process for which several material and process variables affect the quality of the final drug product. Murtoniemi et al. used neural networks to model a fluidized bed granulation process relating two granule properties (granule size and friability) to three process variables (inlet air temperature, atomizing air pressure and the quantity of binder solution).^21–23 The advantages of neural networks were clearly illustrated by the reliable predictions from models, generated from a relatively small number of experiments.

In another study performed by Kesavan and Peck,²⁴ neural networks were successfully applied to predict granule properties as well as tablet properties with respect to five processing conditions in fluid bed granulation. Moreover, neural networks were also used to obtain a deeper understanding of roller compaction,²⁵ where models successfully detected that the speed of the horizontal screw and the air pressure dramatically influenced granule quality while the vertical screw had no significant influence on ribbon quality.

In addition to neural networks, GAs have been used together with neural networks to predict the phase behaviour of colloidal delivery systems and to optimize the properties of acetaminophen tablets.^26,27 Colbourn and Rowe used commercial software embedded with neural networks, GAs and fuzzy logic to model and optimize a tablet formulation.²⁸ Their work demonstrated the benefit of utilizing different data mining technologies in combination to address pharmaceutical formulation challenges.

Improving the understanding of extrusion/spheronization

Neurofuzzy logic has been used to generate a detailed understanding of the extrusion/spheronization process comprising several key stages, each of which involves a number of relatively important process variables (Figure 1).²⁹ In this study, the aspect ratio of the produced spheroids was regarded as one of the key quality attributes of the finished product.

Figure 1 Extrusion/spheronization process and methodology (modified from reference 29).

Neurofuzzy logic software was used to generate predictive models from 56 experimental data records from which a set of "if then" rules were extracted. These rules illustrated the cause and effect relationship between key process variables and the quality of pellets. The rules indicated that a good quality product could be achieved by controlling three major factors:

• Mean load for extrusion was identified as the most influential factor and generally had a positive effect on aspect ratio of the pellet.

• To produce spheroids with most desirable characteristics, the model suggested that a small amount of water was required for wet massing alongside a low mean load for extrusion.

• It was also suggested that unacceptable products were likely when a high mean load and a short die hole length for extrusion were used.

The models from which the rules were extracted were also demonstrated in the form of 3D response surface diagrams (Figure 2). Clearly, the knowledge generated from neurofuzzy logic models can be used to streamline this multistage process.

Figure 2 Example of a 3D hypersurface generated by the neurofuzzy logic model.

Challenges and future aspirations

With new regulatory philosophies challenging the old scientific norm and helping to simplify the route to postapproval manufacturing changes, pharmaceutical companies have a notable financial driver for generating greater understanding of their products. The challenge of understanding multistage and multivariant processes is, however, one of great complexity, which cannot be solved by using conventional and well-established techniques alone. The fishbone diagram (Figure 3) emphasizes the complexity of a typical wet granulated tablet manufacturing process. Such processes comprise a number of variables, for which the relationships with finished product quality may not conform to conventional mathematical thinking. Experimental programmes during development are also likely to be constrained by limitations in raw materials and time caused by accelerated development planning.

Figure 3 The fishbone diagram for a wet granulated tablet manufacturing process.

Rational DoE coupled with advances in informatics capabilities such as AI must, therefore, be exploited to generate sufficient understanding of processes and build reliable predictive models required by regulators. Only then can pharmaceutical companies fully reap the benefits of enlightened regulatory thinking.

Marcel de Matas is a senior lecturer in computational formulation science Qun Shao is a computational project officer Riddhi Shukla is computational formulation manager, all at the Institute of Pharmaceutical Innovation at the University of Bradford (UK).

References

1. Guidance for Industry PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (2004). www.fda.gov

2. Pharmaceutical cGMPs for the 21st Century: A Risk-Based Approach (2004). www.fda.gov

3. L. Eriksson et al., Mutli- and Megavariate Data Analysis. Part 1 Basic Principles and Applications (Umetrics Academy, Kinnelon, NJ, USA, 2006).

4. A. Hussain et al.,Pharm. Res., 8(10), 1248–1252 (1991).

5. K. Takayama et al., Adv. Drug Del. Rev., 55, 1217–1231 (2003).

6. F.J. Bartos, Control Engineering (2003). www.manufacturing.net/ctl/article/CA321037.html

7. K. Young, The Guardian, 9 September 2004. http://technology.guardian.co.uk

8. R.C. Rowe and R.J. Roberts, "Neural networks, Genetic Algorithms and Fuzzy Logic," in Intelligent Software for Product Formulation (Taylor & Francis, London, UK, 1998).

9. J.R. Koza, Introduction to genetic programming — Advances in genetic programming (The MIT Press, Cambridge, MA, USA, 1994).

10. R.C. Rowe and E.A. Colbourn, Chem. Educator, 8, 1–8 (2003).

11. A.S. Hussain et al., Proceeding of Pharmaceutical Technology Conference, 2, 74–76 (1992).

12. A.S. Hussain et al.,Drug Dev. and Ind. Pharm., 20, 1739–1752 (1994).

13. Y.X. Chen et al., J. Cont. Rel., 59, 33–41 (1999).

14. K. Takayama et al., J. Cont. Rel., 68, 175–186 (2000).

15. M. Turkoglu et al., Eur. J. Pharm. Biopharm., 41, 315–322 (1995).

16. J. Bourquin et al., Pharm. Dev. Tech., 2, 111–121 (1997).

17. J. Bourquin et al., Eur. J. Pharm. Sci., 6, 287–300 (1998).

18. H. Sunada and Y. Bi, Pow. Technol., 122, 188–198 (2002).

19. K. Rocksloh et al., Drug Dev. Ind. Pharm., 25, 1015–1025 (1999).

20. Y.X. Chen et al., Drug Dev. Ind. Pharm., 27, 623–631 (2001).

21. E. Murtoniemi et al., Laboratory Microcomputer, 12, 69–76 (1993).

22. E. Murtoniemi et al., Int. J. Pharmaceutics, 110, 101–108 (1994a).

23. E. Murtoniemi et al., Int. J. Pharmaceutics, 108, 155–164 (1994b).

24. J.G. Kesavan and G.E.Peck, Pharm. Dev. Technol., 1, 391–404 (1996).

25. S. Inghelbrecht et al., Int. J. Pharmaceutics 148, 103–115 (1997).

26. S. Agatonovic-Kustrin and R.G. Alany, Pharm. Res., 18, 1049–1055 (2001).

27. M. Turkoglu et al., Eur. J. Pharmaceutics and Biopharmaceutics, 48, 239–245 (1999).

28. E.A. Colbourn and R.C Rowe, Pharm. Tech. Eur., 8(8), 46–55 (1996).

29. Y. Tozuka, R.C. Rowe and P. York, poster presented at AAPS Annual Meeting (Nashville, TN, USA, 2005).