Dipankar Kaul -Head GMP Audits, Asia-Pacific- Novartis Technical Operations, envisages the many ways in which AI can transform pharma manufacturing operations and the present perspective of the science of pharma development, manufacturing and quality assurance
As the pharma manufacturing plants operate on computerised/programmable logic-controlled machines, instruments or technologies within fixed operating parameters to produce products of standard quality and specifications, compatibility towards highly automated and robotic machineries is evident. Through integration with AI self-learning machines, these complex operations can be simplified to a greater degree. The further development of these technologies will facilitate ensuring that these operations become more intelligent and efficient; however, they pose a challenge for the policy makers and regulators to redefine our knowledge of the current good manufacturing practices.
Some of the AI applications useful in text, speech, and video recognition can be utilised further to interpret the pattern and convert it into a meaningful and chronological script of events. While this will certainly reduce the time and increase the efficiency, it will also enhance the reporting quality. Machine learning thus employs statistical applications to identify patterns in the data then make predictions using those patterns. The AI lead machine-learning processes are capable of reading and recognising the key process parameters responsible for raising product quality.
Academics, researchers and a few industry players have been working on AI for several decades now, from as far back as 1955 when John McCarthy coined its definition.  Substantial material has already been published on AI and its prospects; however, its usability in various industrial and commercial applications is yet to be fully applied and benefitted. As stated, not all technological revolutions are straightforward and thus, in like manner concerns continue to arise on the ways that AI will integrate within the domain of pharma engineering and regulatory compliance. At least, this is the dilemma confronting the current pharma scientists, engineers and manufacturers. There is an interesting cultural challenge to overcome, which is technological innovation that supports AI, while the pharma industry has been built on
a strong backbone of ‘traditional’ science that the regulators expect.
A cautious approach has been made by data scientists from various tech giants, who are continuously engaged in constructing some of the AI systems for the pharma industry which has a necessary fence of regulations and is controlled by the health authorities. The present need is to determine that AI solutions are both feasible for use, and safe for adoption.
Perhaps many such technologies had existed earlier, but today they are poised to become mainstream systems with advancements in technologies brought in by IT giants. At the same time, the increase in automation and emergence of new technologies within the pharma industry over the last decades have altered our perspective of the science of pharma development and manufacturing operations. Also, the global expectations of the health authorities, to rationalise and validate pharma processes right from the moment of their development and establish scientific rationales undergirding the principal quality attributes and controlling measures are gaining greater acceptance. Consequently, the systems used to automate the process steps during manufacture, which are uninterruptedly evolving with the implementation of new
instrumentation, are well within the gamut of regulatory compliance and this further diversifies the need for various autonomous technologies.
Integrating machine learning with pharma operations
Plant efficiency and reliability are often cited as the potential reasons for developing and applying AI techniques, developing a variety of algorithms and expert systems to the control and operations.  AI is otherwise referred to as a process of evolving machines or a device to communicate, learn, plan and solve problems in a manner similar to the ways humans do.
Machine learning lies at the core of AI. Learning in the absence of any kind of supervision requires an ability to identify patterns in the streams of inputs, whereas learning with adequate supervision involves classification and numerical regressions.  Although, over time, machines become increasingly capable, and some of the tasks are removed from the list of ‘intelligence’, for e.g. playing chess is no more considered by some as AI which continues to be debatable as AI continuously evolves. However, with various technology giants, robotics companies and IT start-ups expressing greater interest and venturing into AI, the boundary between automation and AI appears to be more distinct. Therefore, innovative ideas continue to appear almost every year, and inventing autonomous machines is signaling the beginning of the AI revolution.
AI and pharma product development
AI applications have already been established and proven for pharmaceutical development. There are various publications demonstrating successful AI interfaces to develop pharmaceutical formulations or predict high yielding combination of chemical reactions. Utilising AI in technology can save time, money and resources, while providing a better understanding of the relationships between the different chemical reactions, physical processes and other related process parameters.
The right algorithms, which are a set of rules to be followed in order to calculate or perform problem-solving operations using computing devices, are fundamental to the designing of the AI architecture for a process industry. Neural networks are such rapidly growing technologies that can be applied to the development and processing of pharma substances and
products. Neural networks are the learning algorithms used within machine learning. In the recent years, neural networks have been demonstrated to be able to offer an alternative approach. Neural networks are mathematical constructs with their capacity to “learn” relationships within data, with no prior knowledge required from the user.
The current trends in pharma sciences portend a good outcome with the developments of information technology and AI. The “Quality by Design” cited in the ICH Q8 Guideline, provides enhanced scientific understanding of critical process and product qualities using the knowledge obtained during the life cycle of a product. In light of this development, “design space” is the area in which a product can be manufactured within acceptable limits.
The Neural Network generates and assesses a range of models to identify the one that will best fit the experimental data provided to it. As such, increasingly, artificial neural networks (often termed ANNs) are used to model complex behaviors in issues like formulation and processing of pharmaceuticals.  To create these spaces, artificial neural networks (ANNs) can be trained to emphasise the multidimensional interactions of the input variables and closely bind these variables to a design space. This assists in guiding the experimental design process to include interactions among the input variables, together with the modeling and optimisation of the pharma formulations. 
There are various machine learning algorithms viz. genetic algorithms or fuzzy logic which have proven to be effective and useful tools in predicting the results that arise from alterations in the input parameters, such as the formulations. Using this approach with neural networks can be productive as it provides “what if” predictions and optimisations.  The objective of this paper is to develop an integrated multivariate approach to obtain quality products based on a sound understanding of the cause–effect relationships between the formulation ingredients and product properties, employing trained ANNs and genetic programming. The data are generated through the systematic application of the principles of the design of experiments (DoE) and optimization studies using artificial neural networks and neuro-fuzzy logic programmes. 
The functioning supporting AI in product development arises from the predictive and deep learning algorithms which recognise patterns and continuously track the input versus output data and accordingly regulate the most suitable design space. This approach generates experimental data within all the possible sub-sets of the variables which could be referred to later on during the entire lifecycle of the product. This approach of using trained algorithms maps dealing with an input to an output and developing knowledge-space (a summary of all the knowledge obtained during product development) is in accordance with the approach of quality by design.
AI and pharma manufacturing operations
The pharma manufacturing plants which operate on computerised/programmable logic-controlled machines, instruments, gadgets or technologies are gearing up to incorporate innovations in therapies and new drug delivery systems. These plants operate within fixed operating parameters to produce products of standard quality and specifications, compatible with highly automated and robotic machineries. These complex operations can be further simplified by integrating with the AI-driven technologies or self-learning machines. These developing technologies will enable operations to become intelligent and efficient, although they pose a challenge for the policy makers and regulators to redefine the way we understand the current good manufacturing practices (cGMPs). As the GMPs require manufacturers of drugs and medical devices etc., to take proactive measures to ensure that their products are safe, pure, and effective, these pharma manufacturers are accountable to demonstrate that their technologies are suitably qualified and validated to consistently produce products of standard quality and efficacy.
Therefore, machine learning involves connecting good manufacturing practices (GMPs) with AI, in a way that the algorithms can recognise any inconsistency within the process and accurately detect them. A variety of sub-process viz. granulations, compression, blending for formulations and chemical synthesis for API manufacturing can be optimised on real-time basis to get standard quality of the in-process bulk. These machines will continuously track and control the process attributes and retain the right recipe for the formulation to maintain the dosage accuracy. What characterises these machines as being different from the present automated machines are the machine learning algorithms or mechanisms by which the continuously collected data will be utilised to make informed decisions on real-time basis. The synergy of robotics and AI can therefore revolutionise the entire spectrum of pharma operations as is possible in any other manufacturing sector.
Therefore, conventional manufacturing processes may undergo a paradigm change. Processes like granulation or compression of tablets or a progressive chemical synthesis in a chain of reactors can be rapidly achieved by eliminating or combining specific processing unit operations. This could involve either simply eliminating tasks that are no longer necessary or using deep learning machine algorithms for equipment that can operate machines to simultaneously perform more than one unit of operation at a time (multi-tasking).
While machine manufacturers mainly focus on eliminating the ‘Out-of-Specification’ (OOS) products, the AI-enabled machines will emphasise a means of rejecting the root-cause(s) for the output going ‘Out-of-Specification’. Understanding the logics and patterns, predicting the variations and adjusting the process beforehand will preempt unnecessary product failures. This will also minimise any redundancy in the process, improve the yield, ensure consistency and stabilise quality.
Besides these operational benefits, the AI-enabled processes will exhibit complete regulatory compliance. Regulatory compliance can be achieved and established through continuous monitoring, tracking not merely for the key process parameters alone but also by generating, compiling and evaluating the online process data. The challenge for the future is to first integrate the AI technologies with the existing machines so that the process data can be collected online, evaluated and the precise decision-set applied on a real-time basis.
As an intelligent machine which actively participates in the production process, in this context, should be able to exchange information and control these processes in real time, thus enabling the production to run in a fully automatic and compliant manner. The machine controls the digital receipt of the incoming order, and even the individual if necessary, as well as the product planning, request for the materials required, production as such, order handling and even product shipment. Once done, these intelligent machines will be able to take complete control of the pharmaceutical manufacturing operations such as:
The next challenge is to validate these systems and integrate the technologies to the total satisfaction of the individual manufacturer and of the health authorities, such as the US Food and Drug Administration.
Process validation for drugs involves legally enforceable requirements under CFR 501 (a) (2) (B), required by the GMP regulations in parts 210 and 211. Regulatory requirements require manufacturers to design a process, including operations and controls, which will result in a product fulfilling its critical quality attributes.
As an alternative technique for process validation where the manufacturing process performance is continuously monitored and evaluated and further tracked after performance qualification, there is a continuous quality verification (CQV) in which the manufacturing process (or supporting utility system) performance is uninterruptedly monitored, evaluated and adjusted as necessary. The CQV moves away from validation as a discrete exercise and is consistent with a lifecycle approach to process validation. These expectations of the regulatory agencies will evidently be satisfied by AI and its applications through machine learning tools.
Therefore, the innovations in AI, the cutting edge scientific and engineering knowledge in pharmaceutical manufacturing, together with the principles of quality management will be incorporated within the framework built on process understanding and risk-based regulatory decisions by the industry and health authorities.
Enabling AI in quality assurance & regulatory compliance
Based upon its wide scope as a human-technology interface (HTI), either through predictive text, speech, gesture or pattern recognition, AI finds use in diverse applications for quality assurance and quality systems management. These applications can come from real-time detection of non-compliances, risk assessment, implementing remedial actions or predicting variations and thereafter via reporting of these facts.
Different types of rapidly emerging AI technologies have further appeared to contribute to communication with machines, computers or other objects. Some of these technologies are beginning to transform the ways humans track, use and manage anything from match commentaries in sports, interpretation of legal judgment to product quality and customer outcomes. AI is becoming a popular means of making more informed decisions. Here, it is worth sharing one of the interesting cases during the 2015 Wimbledon Tennis Championship, where machine learning algorithms were used to automatically turn match statistics and sensor data collected during each game into automated news stories which read as if sports journalists had written them. 
Such applications which are useful in text, speech, and video recognition can be further utilised to interpret the patterns and convert them into chronological and meaningful script of events. While this certainly minimises the time and increases the efficiency, it also improves the reporting quality. Machine learning thus utilises the advanced applications of statistics to identify patterns in the data and then make predictions from those patterns. AI-directed machine learning processes have the capacity to read and recognise the principal process parameters responsible for building product quality.
A long time ago Google proved the value of a program me that can read text. Their search engine algorithm revolutionised internet search, and continues to do so with every advancement. However, it is one thing to detect whether a document contains a certain word or phrase and quite another to understand its context. The newer AI technologies are capable of accomplishing such deep learning, enabling the machines to understand the context and accordingly respond.
Investigations and root-cause identification through pattern detection algorithms
According to Carmelo Rosa, Director Division of Drug Quality, US FDA, since the last 10 years, inadequate or the complete lack of thorough investigations have been true of the fiver top-most GMP violations / deficiencies cited during FDA inspections. The deficient Out-of-Specification (OOS) investigations represent risk to patients and thus compromise either safety or efficacy. The FDA continues to observe a lack of scientific rationale to support the investigative conclusions which primarily are as listed:
Algorithms are being developed which can determine whether a sentence is positive or negative, and its context within a document, besides other information.  Due to these developments in personalization and machine-human interactions, AI is more efficient than ever in pattern recognition, in which friendly machines mimic human investigation patterns to provide fast, easy and authentic service for the senior management. These applications, can thus be utilized for various routine investigations in the pharmaceutical operations during product failures, non-compliance or for Out-of-specification test results.
In order to automate processes involving human level of intelligence, some algorithms and technologies necessitating human-level intelligence are needed. These new algorithm-enabled machines will have the capacity to more accurately capture the events, key process attributes on a real-time basis and speedily identify the areas of anomalies in the process. Fortunately, researchers and companies have developed a myriad of techniques that can dramatically improve automation which is the strong point of AI – viz., business decision making. Such information will be chronologically transcribed, as well as have captured even minute technical details of the root-cause to create comprehensive and scientifically valid investigation reports.
Therefore, the organisation that implements meaningful real-time process surveillances and processes for their investigations will both strengthen their business base, as well as produce sustainable improvements in their productivity, regulatory compliance and product quality.
AI in pharma process risk assessment
AI advance technology platforms — in fact, cognitive computing, in particular, can address complex situations that are defined by ambiguity and uncertainty. Cognitive computing has commenced to make inroads into augmenting business decisions and power performances alongside with human thought process and traditional analytics.
In fact, the domain of risk management relates well to cognitive computing capabilities, as typical risk issues often include unlikely and/or ambiguous events. The cognitive capabilities, including data mining, machine learning, and natural language processing, are replacing traditional analytics and being applied to these massive data sets to enable indicators of known and unknown risks to be identified.
As the pharma products or processes include their own associated intricate risks, clearly evident from the fact that manufacturing involves various components, there is necessarily a certain degree of risk. This quality risk is only one component of the overall risk. Therefore, product quality must essentially be maintained throughout the lifecycle of the product in a manner that the attributes crucial to the quality of the drug product remain consistent with those used in clinical studies. 
Patient protection, performed by risk management in the quality system and manufacturing process, is being allocated supreme significance in the pharma industry. As revealed by the ICH Q9, several components are involved in the overall risk assessment model – risk identification, risk analysis, risk evaluation, risk control (risk reduction / risk acceptance), output after which a continuous review of events is performed as the risk review. Each component is evolved by utilising the massive data gathered from several different permutations and a combination of the process parameters, machine settings and associated variables. The pharma manufacturers simultaneously employ several quality management tools to produce an appropriate risk-assessment model for both their products and processes. Therefore, for patient protection, enormous amount of data is compiled and meaningfully evaluated through scientific knowledge. The degree of effort and documentation of the quality risk management processes must be commensurate with the level of the risk.
Using AI to manage risk is especially beneficial when unstructured data needs to be handled and evaluated. This implies the type of information that does not fit neatly into structured rows and columns. Cognitive technologies, like natural language processing (NLP), use advanced algorithms to analyse the text to draw insights and direction from unstructured data.
To summarise, at the present time a paradigm change is being observed, with engineering principles and product-process design becoming the main principle guiding pharma development and manufacturing. This implies, we are adopting a pattern of thinking, based on which pharma operations and related manufacturing processes are simultaneously and quantitatively being considered. However, we continue to be in the process of understanding and integrating various technologies currently available which can learn, predict and/or measure compound properties, operational attributes, and define and characterise their constitutive behaviour.
Although, the conclusion drawn is that a large part of the fundamental knowledge and technical tools in the AI space exist for the implementation of innovative pharma manufacturing operations, further work is definitely required, particularly at the interfaces between the pharma sciences, regulatory compliance and software engineering, in order to make substantial contribution to the successful operation of the pharma and chemical industries.
Bibliography & References