Artificial intelligence (AI) is fast becoming the smart choice for process development and biomanufacturing control. But taking full advantage will require better use of data and regulatory support.
In its simplest form, AI is a system that can perceive its environment and automatically take actions based on that information. The approach has been used in drug discovery for a decade or so. According to a 2018 news article, Pfizer, Sanofi, and Roche subsidiary Genentech had all started working with AI developers to focus and streamline discovery efforts.1
The rationale is simple: AI is good at finding patterns in complex data. Also: biological sequences and disease mechanisms can be converted into hugely complicated datasets.
Not just for discovery
Most AI-driven activities have ranged from searches for new therapeutic targets—one such search led Berg to a drug candidate for pancreatic cancer2—to studies of the role genetic variation3 plays in disease progression.
More recently, AI has started to appear in the biopharmaceutical plant. According to Vikas Revankar, head of Software and Automation at MilliporeSigma, this development is driven by interest in continuous manufacturing.
“The biomanufacturing industry is evolving from a standard batch processing paradigm to a more continuous operation supported by inline sensors and process analytics to enable real-time product release,” Revankar says. “Data-driven continuous operations will help to significantly improve product quality, reduce production costs and shorten the time to market.”
The increasing complexity of production processes is another factor, observes Krist V. Gernaey, PhD, head of the process and systems engineering center at the Technical University of Denmark.
“The most obvious advantage is that AI can potentially detect patterns in datasets that are difficult to observe for an operator or process chemist,” Gernaey explains. “In this way, one could gain new information about production processes and, of course, exploit that later on toward improving such a process.
“When properly implemented, the use of AI should in general also have the benefit of making use of collected data, instead of just storing the data as historical documentation about a production process that has been completed.”
Jens Smiatek, PhD, an AI and data management expert at the University of Stuttgart, is of a similar opinion about the role of AI in biomanufacturing: “In the plant, AI can be used for laboratory automation, efficient document processing, and process control and steering. Such applications mainly correspond to the improvement of daily business as well as wet-lab and manufacturing work.”
Smiatek cites machine learning (ML)—a technology that provides systems the ability to “learn” by analyzing data over time—as a form of AI that industry is adapting and using to automate production processes.4
“In terms of efficient and improved unit operation models, ML provides a plethora of novel approaches,” he asserts. “In principle, there exist several standard models for distinct unit operations, ranging from the use of differential equations, in terms of mechanistic models via metabolic flux models, to molecular modeling approaches.”
Industry’s approach has been to combine mechanistic models, which are based on the assessment of parameters, with artificial neural networks (ANNs), which “learn” processes.
“Experimental data is evaluated by an ANN in accordance with a high-dimensional multivariate regression approach, which allows the modeler an accurate determination of relevant parameters for the combined mechanistic models,” he elaborates. “For example, the calculation and prediction of key parameters for cultivation or fermentation processes as well as filtration approaches can be improved significantly.”
Another major advantage of AI is its ability to shorten process development. According to Smiatek, process development modeling provides significantly higher precision and agreement with experimental data than do older modeling techniques.
“ML methods in terms of supervised and unsupervised learning techniques are often used for the classification and analysis of experimental data,” he notes. “Hence, there exist a plethora of potential applications for ML methods in terms of modeling, analysis, and execution of bioprocesses.”
In support of this view, Smiatek cites a recent report that ANN has been used to predict the solvation energies and entropies for distinct ion pairs in various protic and aprotic solvents.5 “The high accuracy of such approaches,” he argues, “provides even deeper insights into the underlying molecular mechanisms or process correlations, which helps to improve drug product formulations, and to enable the more efficient design of novel drug molecules or bioprocesses.”
“I see introduction of AI in the biopharmaceutical industry as a process that will take place in several phases,” Gernaey says. “I believe that the first phase will have focus on use of AI for predictive maintenance, that is, the monitoring of specific critical pieces of equipment for changes in behavior.”
Such monitoring can prompt the servicing of equipment whenever deterioration in performance is detected. Otherwise, servicing may be reactive, instigated by the unscheduled breakdown of equipment, which leads to loss of production capacity and, potentially, loss of valuable product.
According to Gernaey, this type of predictive maintenance can be considered a “local” AI application. “Some of these predictive maintenance projects have already been implemented in the bio-based industry,” he remarks.
“That first phase will build confidence with industry,” he continues. “In a second phase, more complex challenges will be tackled, where data from significant parts of a production process can be processed at once, in an attempt to extract knowledge on process performance that can be used toward improving process operations in real time.”
The trend toward real-time maintenance and process improvement is already impacting technology design. “Like our customers, we are looking at applying AI-based condition monitoring technologies,” Revankar remarks. “Many areas such as service and maintenance could benefit.
“As biomanufacturing starts embracing strategies toward continuous processing, the condition of the underlying assets forms an important and fundamental factor. AI-based technologies stand to offer some immediate and tangible benefits there.”
The biopharmaceutical industry often lags other industries when it comes to new technology, and to a degree, this is true for AI in biomanufacturing. However, for some ancillary tasks, the approach is already well established.
“AI and ML approaches in the biopharmaceutical industry are still in their infancy,” Smiatek observes. “However, useful applications already include laboratory automation, detection of impurities in raw materials, process control, and improved modeling techniques.”
The biggest adoption driver for AI and ML is process analytical technology (PAT), which is already established. Many drug firms already have the information technology infrastructure in place to implement AI.
“Large amounts of sensor data can be efficiently processed,” notes Smiatek, who adds that mathematical models are also available that can support the use of digital twins for bioprocess control, steering, and prediction.
Another adoption driver is process control. “As one starts getting into the process space—especially in process monitoring—use of AI-based models stands to offer solutions to monitor performance vis-à-vis patterns and anomalies that start evolving across batches,” Revankar explains. “Using AI-based tools to process data across multiple sources and to facilitate the contextualization of data across batches can expedite the release of manufacturing batches.”
Adoption of AI would be simplified if regulations were clarified. “The main challenges for ML- and AI-based methods are missing guidelines in terms of GMP application,” Smiatek comments. “It is unclear if a proposed model or digital twin would satisfy GMP requirements in addition to scientific validation criteria.
“AI and ML applications are mainly used where GMP requirements do not apply, for example, in early development. However, reasonable guidelines could help extend the use of ML or AI approaches to GMP environments.”
This view is shared by Gernaey, who proposes that regulators, like industry participants, need time to adapt. “I have the impression that regulators push for more efficient production processes,” he says. “But of course, they also require time to adapt to a new reality where potentially more and more complex algorithms are used to convert data to information. What we need, I think, are published case studies where the benefit of applying AI is clearly demonstrated, including details about the methods that have been used.”
In addition, the training of staff in the biopharmaceutical industry could be revised to include AI-enabled technologies. “The main difficulty in exploiting AI isn’t about technology, but more about the requirement that people cultivate different skillsets,” Gernaey maintains. “Alternatively, the solutions that are implemented have to be made user-friendly in such a way that the operator/process chemist can readily use the information that is generated.”
To that end, industry could make greater use of early drug and process development data to train AI models. “Pharmaceutical companies usually produce a lot of relevant research and development data,” Smiatek points out. “With regard to an efficient use in terms of advanced ML approaches, the corresponding data sets need to be saved and structured reasonably in modern data storage systems.
“Specifically, ML approaches require a lot of structured data, such that the cleansing of data sheets consumes a lot of working time. Thus, ML approaches could benefit from the use of efficient data management systems, such as electronic lab notebooks (ELNs) or laboratory information management systems (LIMSs), and the application of FAIR (findable, accessible, interoperable, reusable) principles.”
Big data, big opportunity
Because biomanufacturers are familiar with PAT, they are well placed to use AI and ML. “Most often, ML techniques perform best if a lot of data is available,” Smiatek says. “The biopharmaceutical industry can indeed produce large data sets.”
Not all applications require more than limited amounts of data. Such applications include active learning. But PAT applications are more demanding. They have encouraged biomanufacturers to deploy sensor technologies and collect high-throughput experimental measurements.
With PAT applications, ML as well as other advanced statistics methods can be used to reveal hidden correlations and to improve product or process understanding. In principle, the algorithms needed to support these methods are already available from open or commercial source code platforms compatible with Python, Matlab, or R. The remaining difficulty, then, is efficient data storage and efficient structuring. This difficulty can be addressed by following the FAIR principles.
“New technologies are always welcome,” Smiatek adds. “But at the moment, companies should have enough to do applying their ML and advanced statistics methods onto current or archived data sets.”
This opinion is shared by Revankar, who notes that drug companies should use their data to train models. “To benefit from all the stated promises of AI,” he says, “manufacturers need to learn how to maximize the utilization of the datasets already at their disposal and identify, access, and integrate new data sources.”
There is also evidence that regulators expect more biopharmaceutical companies to use AI in manufacturing. A few years ago, the U.S. Food and Drug Administration launched the emerging technologies program.6 The aim was to promote innovation in general, including the use of novel technologies such as digital twins, advanced modeling, and other ML-based techniques.
Such regulatory programs meet with Smiatek’s approval. “As long as strict guidelines are unavailable for ML applications in terms of the bioprocess development or manufacturing,” he states, “such programs may help to establish novel approaches in companies and to increase the confidence in their benefits.”
- Fleming N. How artificial intelligence is changing drug discovery. Nature 2018; 557: S55–S57.
- Berg. FDA Orphan-Drug Designation of BPM31510 for the Treatment of Pancreatic Cancer [press release]. Issued: January 22, 2018. Accessed: December 31, 2020.
- Yu P, Wilhelm K, Dubrac A, et al. FGF-dependent metabolic control of vascular development. Nature 2017; 545: 224–228.
- Smiatek J, Jung A, Bluhmki E. Towards a Digital Bioprocess Replica: Computational Approaches in Biopharmaceutical Development and Manufacturing. Trends Biotechnol. 2020; 38(10): 1141–1153.
- Yang J, Knape MJ, Burkert O, et al. Artificial neural networks for the prediction of solvation energies based on experimental and computational data. Phys. Chem. Chem. Phys. 2020; 22(42): 24359–24364.
- Emerging Technology Program. Food and Drug Administration website. Updated: October 10, 2019. Accessed: December 31, 2020.