Modern Control Strategies for Biopharmaceutical Monitoring

Lonza's Atul Mohindra, head of R&D at Lonza Integrated Biologics, discusses how process analytical technology (PAT) is reshaping biopharmaceutical manufacturing. According to Mohindra, modern control strategies integrate spectroscopic tools like Raman and near-infrared spectroscopy (NIR) with chemometric models to enable real-time monitoring of critical process parameters (CPPs).

Mohindra addresses key technical challenges: bioreactor environments create high signal-to-noise ratios due to cell density and bubbles, requiring multivariate analysis and partial least squares modeling to extract reliable signals. Beyond single-parameter control, Lonza combines PAT data with traditional measurements (pH, DO, temperature) to build soft sensors offering a full-process view and early detection of out-of-trend events.

On scale-up, he explains that multi-scale training datasets help models remain robust across clones and operating conditions, with reference-measurement verification during site transfers or reactor geometry changes. Lonza's digital twin approach pairs mechanistic models with machine learning for complex, nonlinear dynamics to continuously recalibrate as PAT-generated data accumulates.

PharmTech: How does a PAT control strategy integrate multiple sensor modalities (e.g. NIR, Raman probes, and soft sensors), and how are data streams from these fused in real time to generate actionable process insights?

Mohindra (Lonza): A modern PAT control strategy is built on the integration of multiple novel sensor modalities, including spectroscopic tools such as Raman spectroscopy or NIR to improve process monitoring and control. For instance, Raman spectroscopy can be translated into metabolite concentration measurements using chemometric models that enable continuous monitoring. These continuous measurements of CPPs, provide the means to improve control of select critical quality attributes (CQAs) more consistently. As an example, Lonza’s newly developed intensified fed-batch platform leverages Raman models developed for glucose, phenylalanine, and methionine to control 3 nutrient feed rates during high-inoculum-density cultures without the need for manual interventions.

Beyond upstream process monitoring, chemometric models can also be used to measure CQAs in downstream applications. In addition to Raman spectroscopy, other PAT tools, including NIR and Fourier transform infrared spectroscopy (FTIR), can be applied to measure product concentration, aggregates, and excipients in the purified process streams during the appropriate downstream unit operations. Maintaining sufficient control of these process parameters supports right-first-time manufacturing by enabling target endpoints more consistently.

The ultimate goal is not just to measure and control single CPPs but integrate the breadth of data generated by novel PAT solutions with more traditional data streams (i.e., pH, DO, temperature) in order to develop soft sensors that can map out the whole process from start to finish. These soft sensors add another layer of process control, enabling real-time monitoring, capable of monitoring a process, real-time identification of out-of-trend (OOT) events, and timely process adjustment to bring these OOT events back within control. By aligning and processing these novel PAT and traditional data streams in real time within Lonza’s data platform, a unified view of the process is created. The output is a set of actionable variables, such as nutrient concentrations, metabolic indicators, and cell culture health, that can be used to inform operators or directly drive automated control strategies.

What are the biggest data quality and signal-to-noise challenges you face when deploying in-line spectroscopic sensors in a bioreactor environment, and how do chemometric models or multivariate analysis help overcome them?

Using in-line spectroscopic measurements in bioreactors can be inherently challenging. The process environment is complex due to high cell densities, bubbles, and other physiochemical effects making for high signal to noise ratios and therefore difficult to interpret data. This makes it difficult to directly interpret raw spectra, particularly for low-concentration analytes embedded in a dynamic process environment.

Chemometric approaches, such as multivariate analysis and partial least squares modeling, are essential to extract relevant information from this noise by identifying patterns that correlate with concentration changes of key analytes. For example, this strong performance model (R²>0.90) has been successfully applied to monitoring glucose, phenylalanine, and methionine, in order to deliver accurate and reliable measurements that would be difficult to achieve with other methods.

Robust model performance is also achieved through careful preprocessing and training on diverse datasets that capture expected variability across clones, scales, and conditions. Ongoing monitoring of model output and periodic updates help ensure these models continue to remain stable over time.

How do PAT-enabled glucose, lactate, and dissolved oxygen measurements translate into automated control decisions, and what measurable impact have you seen on titer and yield at commercial scale?

PAT-enabled measurements are first translated into actionable process variables using chemometric model-based approaches, allowing real-time estimation of nutrient levels (e.g., glucose and lactate) and metabolic states. These values are then compared to predefined process set points, with any deviation triggering automated real-time adjustments, such as changes in feed rates or aeration strategies.

This enables a more dynamic control approach that responds to actual cellular demand, rather than relying on fixed feeding strategies used in traditional and intensified fed-batch processes operated by Lonza. The result is tighter control of the culture environment, reduced variability, and improved process efficiency. In practice, this has led to a significant reduction in potential process deviations, while maintaining consistent product titer, yield and product quality at scale. In addition, spectroscopic measurements used for trending monitoring can alert operators to early signs of process drift, allowing for intervention.

Scaling a PAT-integrated fed-batch process from development to commercial manufacturing introduces significant variability in sensor response and process dynamics. How can one manage scale-dependent PAT model drift, and what does the revalidation workflow look like when transferring processes between sites or bioreactor geometries?

Scale-up can introduce variability in both process behavior and sensor response, which can lead to drift in PAT model performance if not properly managed. One key strategy is to develop models using multi-scale datasets that capture expected process variability early on. This approach has been shown to improve the robustness of Raman-based models for nutrients across different operating conditions for a diverse panel of clones producing monoclonal antibodies and novel molecular formats.

During scale-up or site transfer, model predictions are typically verified against reference measurements to ensure performance remains within expected ranges. Where needed, models can be adjusted or updated, supported by ongoing performance monitoring and predefined model evaluation metrics to enable consistent performance across scales and sites.

Digital twins for bioprocesses often blend mechanistic kinetic models with data-driven machine learning layers. How do you decide where first-principles biology ends and where machine learning takes over in your twin architecture, and how do you keep the twin calibrated as experimental data accumulates through early development?

Digital twin architectures in bioprocessing typically combine 2 types of models: mechanistic models, which describe well-understood biological and physical relationships, and data-driven machine learning components, which can capture more complex or nonlinear behavior. The balance between the two depends on the level of process understanding, data availability, and the intended use of the model.

Mechanistic models are used where reliable established principles exist. For example, this occurs in mass balances, reactor dynamics, or well-characterized relationships such as substrate uptake and growth kinetics within defined operating ranges. Machine learning is also applied where large amounts of data are available, but the system dynamics are too difficult to model explicitly.

PAT plays a critical role by providing high-frequency, real-time data streams that feed and update the digital twin. As more experimental data becomes available, the twin is continuously calibrated through parameter updates or model retraining, allowing it to evolve alongside the process and improve predictive accuracy over time.

About Atul Mohindra

Atul Mohindra is vice president of R&D for Lonza Integrated Biologics where he brings more than 20 years of experience in cell biology, bioprocessing, media development, and process scale up experience to translate scientific innovation into scalable manufacturing processes.

In this role, he leads a team focused on developing and implementing innovative bioprocessing and analytical solutions across the product lifecycle. One current focus for our team is integrating technologies such as advanced process analytical technologies, automation, and digital tools to improve process understanding, accelerate development timelines, and enable more robust and predictable biologics manufacturing.