Integrating Advanced Technologies for Pharma Analysis

Future and media concept | Image Credit: ©peshkova - stock.adobe.com

Use of process analytical technology (PAT) and advanced digital solutions such as artificial intelligence (AI) and machine learning (ML) is a crucial component of the move to Pharma 4.0. Online/inline/at-line PAT enables real-time monitoring of processes during development and manufacturing. That information can then be applied to support process control, leading to more consistent performance. When PAT is integrated with AI/ML algorithms, it is possible to develop predictive models that enable improved control, support more effective decision-making, and reduce risks, leading to optimized operations, enhanced customer experiences, and better financial outcomes, according to Edita Botonjic-Sehic, head of process analytics, data engineering and data science at ReciBioPharm.

Improving soft sensor performance

Several PAT tools are particularly well-suited for integration with AI and ML algorithms. At the top of the list are soft sensors, which integrate a software layer based on a mathematical model that is employed to monitor process states, observes Stacy Shollenberger, senior manager, process analytical technology with MilliporeSigma.

Analysis of the results is complex, however, and requires a high level of expertise to identify, correlate, and quantify the signal and use it for predictive analysis, according to Kaschif Ahmed, principal data scientist with ReciBioPharm. “The best integration solution is one that has an extreme amount of flexibility for measurement and analysis and can easily be validated without the constant need for calibration,” he adds.

AI/ML models can be directly integrated inside the PAT tool, providing a data-driven means to estimate or predict parameters that are otherwise difficult to measure or cannot be measured directly, says Moritz von Stosch, chief innovation officer of DataHow.

Shollenberger points to spectroscopic sensors including Raman, nuclear magnetic resonance (NMR), and near-infrared (NIR) technologies, which offer near-instantaneous measurement of key metabolites and nutrients (e.g., lactate, glucose). However, the spectra generated by these techniques require interpretation using chemometric models leveraging AI/ML algorithms tailored for chemical analysis. Mass spectrometry, she notes, can also greatly benefit from AI/ML interpretation of the produced spectra.

High-performance liquid chromatography (HPLC) and ultra-high performance liquid chromatography (UHPLC), meanwhile, can also be coupled to online analyzers that rely on models and used to monitor product purity in near real-time, according to Shollenberger.

Importantly, in comparison to standard linear regression approaches, ML can increase prediction accuracy when nonlinearities arise (e.g., due to superimpositions in the spectra), says von Stosch. Fusion of data from different sensors is also possible, enabling capture of the overall state of the process better and reducing the impact of measurement noise. “By fusing sensor information with predictions coming from process models, the performance of such approaches can be extended, (e.g., in the case of low signal-to-noise ratios),” he remarks.

Depending on the monitoring scope, Shollenberger notes that hybrid models combining AI/ML with mechanistic understanding can be better suited for real-time monitoring than AI/ML models alone because the former generally better capture the complexity of bioprocesses, leading to more reliable, interpretable, and generalizable simulation results. Upstream processes in particular stand to gain the most from hybrid models, she says, as they are characterized by the high degree of unpredictability inherent in living cells.

There are applications for AI/ML outside of soft sensors as well. For example, van Stosch notes that the advent of deep learning, a specific ML method, has also made real-time-based image processing possible. “Today, ML methods are used to count cells or particles, classify different types of cells or particles, inspect vials for release, and other applications,” von Stosch comments.

Universally applicable

PAT tools integrated with AI/ML-driven models can be used in many different applications across modalities, the drug development and manufacturing cycle, and unit operations. “Such integrated tools are suitable for all types of processes, including those that produce small molecules and biotherapeutics and for both upstream and downstream unit operations at process development and manufacturing scale,” states Botonjic-Sehic.

“Real-time operation and immediate decision-making using PAT tools and digital twins will increase throughput and reduce failures during process development and manufacturing,” adds Ahmed. For these and other reasons, Botonjic-Sehic notes that the use of PAT and digital infrastructure is supported by regulators, with some guidelines available to help the industry move in this direction.

Real-time process optimization relying on process models is one specific activity that can particularly benefit from AI/ML at the upstream level, according to Shollenberger. “Integration of AI/ML/hybrid models with PAT in process development helps to accelerate process understanding, including the impact of different critical process parameters (CPPs) on product critical quality attributes (CQAs), and facilitates the extrapolation of process knowledge across different scales,” she says.

Process development activities can greatly benefit from more continuous tracking of the process state and the ability to make decisions during or after the run, rather than having to wait for the analytics for several hours or days, agrees von Stosch. “In addition to accelerating process development, PAT tools with integrated AI/ML models can render on-the-fly process optimization possible,” he states.

Such knowledge management is also crucial for a smooth process scale-up and transfer from development to pilot and manufacturing scales, all while ensuring stable and effective performance, Shollenberger contends.

As examples, Shollenberger highlights the use of multivariate statistical process control methods within a quality-by-design approach to process development to maintain CPPs and CQAs within the identified design space and new AI/ML design-of-experiment (DoE) tools such as the Bayesian DoE to reduce experimental effort and development costs by providing experiments that are the most informative for characterizing processes. Finally, she points to the use of digital twins for predicting process performance at production scale by mirroring and extrapolating insights gained from smaller-scale experiments. “This approach allows for a more accurate representation of how processes will behave in full-scale production, ultimately streamlining process development and enhancing overall efficiency,” Shollenberger comments.

In addition to supporting the scale-up from process development to production by learning to adapt scale-dependent process parameters and optimizing their configurations, AI/ML/hybrid models supporting real-time or near real-time monitoring and control can greatly improve process efficiency, reliability, yield, and product quality, says Shollenberger. PAT and digitally driven infrastructure can, agrees Botonjic-Sehic, reduce uncertainty by providing insights into process conditions through simulators enhanced with integrated AI/ML and PAT real-time data. “Real-time release testing will also be enabled by AI/ML with PAT due to the ability to achieve rapid quality verification,” she notes.

Improved responsiveness and better decision-making

The integration of AI, ML, and hybrid models with PAT tools offers numerous benefits that extend beyond simple process monitoring. Having an infrastructure that includes novel PAT tools, ML models, and automation allows researchers to identify their targets of interest faster and more accurately during process development and manufacturing, according to Ahmed. “Having predictive models available for each step in the process enhances process understanding and ultimately allows businesses to deliver drugs to market more quickly with greater quality while also reducing waste and cost,” he states.

Increased (near) real-time insight into the evolution of processes makes several actions possible, adds von Stosch. He highlights the ability to take corrective actions either manually or via closed-loop control, make more informed decisions about when to harvest or release material for further processing, introduce changes that afford greater insights into the process, and ultimately increase process capability.

Specifically, according to Shollenberger, chemometric models can interpret spectroscopic data and predict in near real-time process variables of interest. Advanced control techniques based on these predictive models can anticipate future deviations, enabling timely interventions and process adjustment. “It is the combination of real-time deviation monitoring and process optimization that enhances process understanding and control, leading to better quality, increased speed, and greater flexibility in the manufacturing process,” she concludes.

Several notable challenges

The potential presented by integrating AI/ML with PAT technologies is limited to some degree by several challenges to successful implementation of these solutions. Complexity, connectivity, and data availability and quality are the top issues.

“The complexity of model building and maintenance, particularly in the highly regulated biopharmaceutical environment, necessitates extensive qualification and validation efforts,” says Shollenberger. Identifying the right PAT tools that are suitable for the right level of analysis, developing soft sensors specifically targeting each CQA in the process, and building in the analytics and logic to enable process control must be achieved first, according to Botonjic-Sehic. Appropriate automation solutions that support collection and use of real-time information to predict process performance and identify failures as they occur must also be developed. Then once deployed, validation, compliance and data integrity components must be well understood before a drug is processed for release, she comments.

In addition, Shollenberger notes that executing models in a real-time process control environment involves managing time latencies and ensuring seamless connectivity and full compatibility between sensors, other equipment, software, and all interfaces.

Effective training of AI/ML models, meanwhile, requires substantial volumes of experimental data to ensure reliable predictions, placing a considerable burden on operators to generate these datasets. “ML data quantity and quality requirements are rarely met today as current datasets often consist of infrequent data points, such as daily measurements as PATs are mostly offline, limiting the availability of high-frequency historical data for model training,” Shollenberger observes. Hybrid models, she says, are therefore a good alternative for reducing the data quantity required for model training.

On a positive note, Botonjic-Sehic emphasizes that once digital models are developed using historical run data, the robustness and validation is integrated into the process, resulting in faster manufacturing with better quality while decreasing the cost of operation.

Some real-world implementation

The implementation of PAT solutions and AI/ML-enabled PAT solutions in the pharmaceutical industry at production scale is more prevalent in the small-molecule segment. “The chemical nature and synthesis of small molecules are relatively straightforward, and well-established manufacturing technologies dominate the production environment. In the face of intense generic competition and cost pressures, however, drug makers are increasingly seeking innovative, speedy, and cost-efficient production approaches, and AI/ML-enabled PAT solutions have attracted significant interest,” Shollenberger explains.

Biotherapeutics producers are moving toward the use of integrated AI/ML/PAT, but at a slower pace due to the greater complexity of biomolecules, according to Ahmed. While there is significant potential for AI to enhance process understanding and control, enabling informed decisions in the future, today human involvement remains essential to address the current limitations of AI, Shollenberger adds. “Human expertise is still necessary to interpret analytical results accurately and to develop robust models that will leverage AI support in the future,” she says. In fact, she believes collaboration between human insight and AI capabilities will be crucial for advancing biomolecule production.

One example already under development is ReciBioPharm’s novel process for RNA-based vaccines and therapeutics, which it is developing in collaboration with the Massachusetts Institute of Technology using grant money from FDA. The continuous GMP process is extensively enabled by real-time analytical technologies leveraging AI/ML models, including new fit-for-purpose solutions. It is expected to be operational in mid-2025, according to Aaron Cowley, CSO, ReciBioPharm.

DataHow, meanwhile, released a spectral data analytics module within DataHowLab, its advanced process data analytics solution that deploys AI/ML with hybrid models (1).

Rapid advances expected

Integration of AI/ML within PAT tools will, according to Ahmed, eventually make real-time data analysis more autonomous and ultimately support real-time release and thus a paradigm shift in pharma manufacturing.

Indeed, Shollenberger believes significant advances in PAT capabilities will be achieved in the coming years, with integration of AI and ML representing a crucial next step toward the envisioned “facility of the future”.

In the short term, she expects the push for cost reduction and environmental sustainability to drive the miniaturization and cross-unit factorization of automated PAT devices, making these advanced PAT solutions more accessible and cost-effective. In addition, use of calibration-free or limited-calibration PATs could help to streamline the integration of these technologies into existing processes with lower cost.

Other advances will occur in knowledge management, as it is vital for collecting and structuring knowledge coming from different devices and PATs. “Scale agnostic, cross-process, cross cell line, and cross-proteins knowledge management is essential for enabling the application of AI and ML across various processes and steps,” Shollenberger states.

Once these advances are achieved, it will be possible to deploy integrated AI and ML solutions that facilitate near real-time quality assurance, process optimization, and continuous improvement. “These advances will help ensure that final products consistently meet regulatory and quality standards, enhancing the efficiency and reliability of pharmaceutical manufacturing,” observes Shollenberger. She cautions, however, that integration of AI and ML models into GMP processes introduces regulatory challenges related to data security, algorithm transparency, and the validation of AI-driven processes against existing industry standards, all of which must be addressed to ensure their effective use while maintaining compliance with regulatory requirements.

Reference

1. DataHow. DataHowLab Spectra. https://datahow.ch/datahowlab-spectra/ (accessed Nov. 7, 2024).

About the author

Cynthia A. Challener, PhD, is a contributing editor to Pharmaceutical Technology®.

Article details

Pharmaceutical Technology®
Vol. 48, No. 12
December 2024
Pages: 16–18

Citation

When referring to this article, please cite it as Challener, C.A. Integrating Advanced Technologies for Pharma Analysis. Pharmaceutical Technology 2024 48 (12).