Key Takeaways
Adaptive and closed-loop control in T cell bioreactors primarily regulates environmental and metabolic parameters, not critical quality attributes such as phenotype and functional potency.
Current evidence shows benefits in stabilizing cell growth trajectories and improving reproducibility, but not in guaranteeing therapeutic quality outcomes.
Sensor limitations — especially for real-time viability, differentiation state, and phenotype — define the practical ceiling of closed-loop control in cell therapy manufacturing today.
Regulatory frameworks support data-driven control strategies only within validated control strategies governed by pharmaceutical quality systems, not autonomous decision making.
The future of closed-loop quality control depends more on advances in process analytical technology and structured decision frameworks than on control algorithms alone.
Promise vs. Practice in Closed-Loop CGT Manufacturing
Adaptive and closed-loop control systems are increasingly presented as solutions to one of the most persistent challenges in cell and gene therapy (CGT) manufacturing: biological variability. In particular, the concept of using real-time sensor data and automated feedback to continuously adjust bioreactor conditions presents a way to stabilize processes that are otherwise subject to donor-to-donor differences, dynamic cell behavior, and narrow operating windows. These ideas are often discussed alongside broader themes, such as process analytical technology (PAT), real-time release testing (RTRT), and digital manufacturing, creating the impression that CGT production is moving toward autonomous or self-correcting systems.
Regulatory frameworks, however, position closed-loop control very differently. Guidance from both U.S. and European authorities treats real-time data and adaptive control mechanisms as components of an overall control strategy rather than as independent decision-makers. Process data may be used to support evaluation of product quality, but only within predefined specifications, validated models, and established quality system governance. In this framing, closed-loop systems do not replace human or quality-system authority; they function as tools embedded within a pharmaceutical quality system (PQS) that remains responsible for release decisions and regulatory accountability.
At the same time, many claims made about “adaptive control” in CGT extend beyond what current measurement technologies and biological models can reliably support. While environmental parameters, such as pH, dissolved oxygen, and nutrient levels, can be monitored and adjusted in real time, key biological attributes, including viability, phenotype, and functional state, remain difficult to measure continuously with validated online sensors. Existing solutions are unevenly developed across process steps, and truly online measurement of critical biological characteristics remains limited.
This gap between promise and practice creates a tension that is central to the discussion of adaptive control in CGT bioreactors. On one hand, there is clear technical progress in integrating sensors, automation, and control algorithms into cell culture systems. On the other, there is a risk of overstating what these systems can currently achieve with respect to product quality and clinical relevance. Understanding where closed-loop control is already demonstrably effective and where it remains aspirational is essential for sponsors and manufacturers seeking to adopt these technologies in a regulatory-aligned and scientifically defensible way.
What “Adaptive Control” Means (Engineering vs. Regulatory Definitions)
In engineering terms, adaptive or closed-loop control refers to systems that continuously adjust operating conditions in response to measured process variables. Sensors collect data in real time, control algorithms interpret those data, and actuators modify inputs, such as feed rates, agitation, or gas flow to keep the system within defined targets. In bioreactor contexts, this approach is often associated with feedback control, model predictive control, and other strategies designed to compensate for process disturbances and biological variability.
Regulatory definitions frame these concepts more narrowly. RTRT is defined as the ability to evaluate and ensure the quality of an in-process and/or final product based on process data, typically using a valid combination of measured material attributes and process controls.1 This definition emphasizes that process data and control mechanisms are part of the evidence used to support quality evaluation, not substitutes for it. RTRT does not eliminate the need for specifications or acceptance criteria; it reconfigures how evidence is generated within an approved control strategy.
Crucially, RTRT must still satisfy the statutory and regulatory requirements for testing and release for distribution. FDA guidance explicitly states that RTR, as defined within the PAT framework, can meet the requirements of 21 CFR 211.165, but only when implemented with prior Agency approval and within a validated system of controls.2,3 This places adaptive control firmly inside the PQS, where models, sensors, and decision logic are subject to the same expectations for validation, documentation, and oversight as conventional testing methods.
Recent literature also highlights the importance of distinguishing between different forms of data-driven manufacturing. For example, the concept of a “digital shadow” has been proposed as a means of mirroring cell growth behavior in real time using incoming process data. However, by definition, a digital shadow operates with one-way data flow and does not feed information back into the process for control actions. As such, it cannot be considered adaptive control, even though it may provide valuable insight into process behavior.4 This distinction underscores the risk of conflating monitoring, modeling, and control under a single umbrella of “closed-loop” manufacturing.
Hence, adaptive control has two overlapping but distinct meanings. From an engineering standpoint, it refers to dynamic adjustment of process parameters based on measured inputs. From a regulatory standpoint, it is acceptable only insofar as it functions within an approved control strategy that continues to meet release requirements and quality system governance. Understanding this dual definition is essential for evaluating claims about autonomy in CGT manufacturing. What is often described as “adaptive” in technical discussions may still be, in regulatory terms, a structured and highly constrained component of a broader quality-controlled process rather than a step toward self-governing production systems.
What Parameters Can Actually Be Controlled Today
Experimental and early manufacturing implementations of adaptive control in cell therapy bioreactors show that closed-loop strategies can influence aspects of cell growth and metabolic behavior, but their scope remains focused on a limited set of measurable variables. One of the clearest demonstrations comes from work applying model predictive control (MPC) to cell therapy cultures using lactate as a feedback signal. In this approach, cumulative lactate production is used as an indicator of cell growth, and feeding strategies are adjusted dynamically to guide the culture toward a predefined growth trajectory.5 This represents a form of adaptive control grounded in metabolic proxies rather than direct measurement of therapeutic function.
Closed-loop control has also been demonstrated in the context of T cell expansion systems. A published example of a closed-loop bioreactor for ex vivo T cell proliferation showed that automated control strategies could support enhanced cell expansion and more consistent culture conditions compared with less controlled environments.6 These results illustrate that feedback control can stabilize key aspects of the culture environment and influence population growth, particularly in processes where expansion efficiency is a central objective.
Beyond growth control, experimental systems have begun to incorporate online monitoring of metabolites and cell density as inputs for automated adjustment. In situ Raman spectroscopy has been used to track metabolite concentrations and enable on-line automated control of viable cell density in perfusion mammalian cell cultures, demonstrating that spectroscopic techniques can serve as real-time proxies for otherwise offline measurements.7 Similarly, glucose and lactate monitoring has been implemented in chimeric antigen receptor T cell (CAR-T) manufacturing contexts using single-use probes integrated into expansion systems, allowing continuous observation of metabolic trends during culture.8 These studies establish technical feasibility for feedback based on chemical and biomass-related parameters, even if they remain limited to controlled or developmental settings.
What these examples share is a focus on environmental and metabolic variables rather than on direct measures of product quality. Most current control strategies regulate parameters such as temperature, dissolved oxygen, nutrient availability, and cell density, while critical quality attributes (CQAs), such as phenotype, differentiation state, and functional potency, remain largely outside the reach of real-time control.9 In practice, adaptive control today is best understood as a means of stabilizing process conditions that are believed to influence quality, rather than as a mechanism for directly controlling quality itself.
This distinction is central to evaluating claims about closed-loop CGT manufacturing. The technologies now available can guide cultures toward desired growth and metabolic profiles and reduce operator dependence for routine adjustments. However, they do not yet provide continuous, validated control over the biological attributes that ultimately define therapeutic performance. As a result, adaptive control in current T cell processes operates as an engineering layer that supports consistency of conditions, while downstream testing and quality system oversight remain responsible for determining whether those conditions have produced an acceptable product.
Sensor Reality: What Cannot Be Measured in Real Time
The effectiveness of adaptive control in cell therapy manufacturing is fundamentally constrained by what can be measured reliably and continuously during culture. While physical and chemical parameters, such as temperature, pH, dissolved oxygen, and certain metabolites, are routinely accessible through in-line or at-line sensors, many biologically meaningful attributes remain difficult to observe in real time. Among the most significant limitations is the challenge of online viability measurement. Reviews of sensing technologies note that, despite decades of development in bioprocess monitoring, few commercial methods exist for robust, real-time determination of cell viability in industrial settings, and most viable cell assessments still rely on offline sampling and laboratory-based assays.10
This limitation has direct implications for closed-loop control strategies. If viability itself cannot be measured continuously with validated sensors, then adaptive systems must rely on proxy variables, such as metabolite concentrations, capacitance signals, or growth trends. These proxies can be useful for stabilizing process conditions, but they do not provide direct confirmation of biological health or functional integrity. As a result, feedback control based on these measurements can regulate the environment in which cells grow without necessarily ensuring that the cells remain in a desired biological state.
Additional gaps remain related to more complex biological attributes. Analyses of monitoring solutions for CAR-T processes emphasize that tools for tracking differentiation state, phenotype, and functional readiness of T cells are unevenly developed and often limited to offline characterization steps.9 Similarly, while mechanical integration and process control have advanced, corresponding progress in real-time biological measurement has lagged behind, particularly for attributes tied directly to therapeutic function.11
These gaps underscore a critical boundary between engineering control and biological understanding. Closed-loop systems can adjust nutrient supply, gas exchange, and agitation in response to sensor inputs, but they cannot yet respond directly to shifts in T cell phenotype or differentiation trajectory because those states are not observable in real time with sufficient reliability. In practical terms, this means that adaptive control currently manages the conditions believed to influence quality rather than the quality attributes themselves.
Recognizing this distinction is essential for realistic expectations of closed-loop manufacturing in CGT. The absence of robust online measurements for viability, phenotype, and functional state does not invalidate the use of adaptive control, but it limits its scope. Until sensing technologies evolve to capture these attributes continuously and with regulatory confidence, adaptive control will remain an indirect tool that stabilizes the process environment while leaving final assessment of biological quality to established analytical methods and quality system oversight.
Evidence of Benefit: What the Data Actually Show
The strongest evidence for adaptive and closed-loop control in cell therapy manufacturing comes from targeted demonstrations showing improved regulation of growth-related variables rather than direct control of therapeutic quality attributes. In one peer-reviewed study focused on cell therapy applications, model predictive control (MPC) was used to guide cell growth by adjusting feeding strategies based on cumulative lactate production. Lactate served as a measurable proxy for cellular activity, and the control system successfully steered the culture along a predefined growth trajectory. This work provides concrete proof that algorithm-driven feedback can stabilize and shape growth dynamics under controlled conditions.5
Additional experimental evidence supports the feasibility of closed-loop control in T-cell expansion systems. A published example of a closed-loop bioreactor for ex vivo T cell culture reported that automated regulation of culture conditions enabled enhanced T cell proliferation and more consistent expansion performance. The study illustrates that feedback control can reduce variability in environmental conditions and support reproducible cell growth, particularly in processes where expansion efficiency is a primary objective.6 Together, these findings show that adaptive control can exert meaningful influence over parameters linked to productivity and culture stability.
However, the scope of demonstrated benefit remains relatively narrow. These studies focus on growth control and proliferation outcomes rather than on downstream attributes, such as phenotype, differentiation state, or functional potency. The evidence base therefore supports adaptive control as a tool for managing process behavior, not as a mechanism for guaranteeing final product quality. This distinction is critical when interpreting claims about “autonomous” or “self-correcting” CGT manufacturing.
Industry publications extend these findings by asserting broader operational advantages, describing improved reproducibility, reduced processing bottlenecks, and more responsive adjustment of culture conditions through the use of in-line monitoring and feedback strategies.12,13 While these accounts align conceptually with the peer-reviewed demonstrations of growth control, they are largely descriptive and forward-looking, and they do not provide the same level of experimental validation or regulatory context.
Taken together, the available data indicate that adaptive control can deliver tangible benefits in stabilizing culture conditions and guiding cell growth trajectories. What they do not yet show is a consistent ability to control the biological attributes that define therapeutic performance. The current evidence therefore supports a measured conclusion: closed-loop systems improve how processes are run, but they do not yet redefine how product quality is assured.
Regulatory Perspective: Control Strategy vs Autonomous Manufacturing
Regulatory frameworks consistently position adaptive and closed-loop control as elements of an approved control strategy rather than as a pathway to autonomous manufacturing. Guidance on PAT makes clear that RTRT, when used, must be implemented with prior agency approval and within a validated system that demonstrates how process data and controls ensure product quality. In this context, adaptive control does not operate independently of regulatory oversight; it must be justified through defined models, qualified sensors, and documented decision logic that can be reviewed and inspected.2
Quality decisions associated with adaptive control remain embedded within the PQS and governed by quality risk management principles. International guidance emphasizes that manufacturing control strategies must be supported by systematic risk assessment, ongoing process performance monitoring, and formal change management. These requirements apply equally to conventional batch processes and to those incorporating advanced control algorithms. Models and automated adjustments therefore become part of the life cycle knowledge base rather than substitutes for it, and they are subject to the same expectations for validation, documentation, and continuous improvement.14,15
European regulatory guidance reinforces this distinction between data-driven support and autonomous decision making. RTRT is defined as the use of process data, typically in combination with measured material attributes and process controls, to evaluate and ensure product quality. This framing underscores that process data may contribute to release justification, but only as one component of an overall quality evaluation framework. It does not eliminate the need for acceptance criteria or transfer authority for release from quality systems to automated tools.
Together, these perspectives draw a clear boundary between control strategy and autonomy. Adaptive control may adjust process parameters in real time, but responsibility for determining whether a batch is suitable for clinical or commercial use remains with quality organizations operating under established regulatory expectations. The use of process data can strengthen the scientific basis for those decisions, yet it does not replace the requirement for PQS governance, risk-based evaluation, and accountable quality assurance. In this sense, regulatory frameworks support innovation in control methods while preserving the principle that manufacturing decisions must remain transparent, validated, and ultimately human-governed.
Biological Variability vs. Engineering Control
The most persistent limitation on adaptive control in T cell manufacturing is not the absence of control algorithms but the biological variability those algorithms are asked to manage. T cell cultures are not simply expanding populations that respond predictably to nutrients and oxygen; they are dynamic systems in which differentiation state, phenotype distribution, and functional readiness can shift in response to subtle process conditions. Many of these shifts are clinically meaningful, but they remain difficult to observe continuously during culture. Reviews of monitoring solutions for CAR-T manufacturing emphasize that tools for tracking phenotype and differentiation-related attributes are often confined to offline characterization steps, leaving limited ability to monitor these biological states in real time.9
This constraint is compounded by the challenge of measuring viability and other indicators of biological health online. Even in broader bioprocess contexts, reviews note that online viability determination remains difficult and that few commercial methods exist for robust real-time viability measurement at industrial scale.10 For T cell processes, this means that adaptive control systems are frequently dependent on proxy measurements (e.g., metabolites, growth trends, or indirect biomass estimates) that may correlate with certain outcomes but do not provide direct insight into whether cells remain in the desired functional state. As a result, control loops can stabilize environmental conditions without necessarily stabilizing the biological properties that determine therapeutic performance.
The emerging use of digital representations of process behavior further illustrates the boundary between insight and control. A digital shadow, for example, can mirror cell growth dynamics in real time using process data inputs, offering a structured way to interpret ongoing culture behavior. However, by definition, a digital shadow involves one-way data flow and does not feed information back into the process for corrective action. It can support situational awareness and decision making, but it does not constitute adaptive control and cannot compensate for measurement gaps on its own.4
These issues define the practical ceiling of engineering control in T cell manufacturing today. Closed-loop systems can regulate environmental variables and reduce operator dependence for routine adjustments, but they cannot directly control phenotype, differentiation trajectory, or functional potency when those states are not observable in real time. Until sensing technologies mature to provide continuous, validated measurement of biologically meaningful attributes, adaptive control will remain an indirect strategy that manages conditions believed to influence quality, while final assurance of biological performance continues to rely on established analytical methods and quality system oversight.
Practical Use Cases Where Adaptive Control Makes Sense
Despite the limitations imposed by sensor capability and biological variability, there are well-defined areas in T cell manufacturing where adaptive control can be applied in a practical and defensible way. These use cases focus on stabilizing process conditions that are known to influence culture performance and that can be measured reliably in real time. Rather than attempting to regulate complex biological attributes directly, adaptive control in these contexts serves to reduce operational variability and support more consistent execution of established process designs.
One such application is feed- and metabolite-based control. In cell therapy cultures, metabolites, such as lactate, can be used as indicators of cellular activity and growth. Model predictive control strategies have been demonstrated that adjust feeding profiles dynamically based on cumulative lactate production, guiding cultures toward predefined growth trajectories. This approach allows nutrient delivery to respond to actual culture behavior rather than to fixed schedules, helping to compensate for variability between runs while remaining grounded in measurable process signals.5
Adaptive control is also well suited to environmental stabilization during T cell expansion. Closed-loop bioreactor systems have been shown to regulate key parameters, such as temperature, dissolved oxygen, and agitation, in a way that supports enhanced and more consistent T cell proliferation. By automating routine adjustments, these systems can reduce operator dependence and improve reproducibility of culture conditions across batches, particularly in processes where expansion efficiency is a primary performance driver.6
A further area of application is perfusion control using biomass proxies. Trade and industry sources describe the use of online measurements, such as capacitance-based estimates of viable cell density, to adjust perfusion and bleed rates in real time. These strategies aim to maintain cultures within targeted density ranges and prevent nutrient depletion or waste accumulation. While such approaches are more extensively documented in other mammalian cell culture contexts, they illustrate how adaptive control can be anchored to measurable proxies rather than to complex biological endpoints.12
Across these use cases, a common theme emerges: adaptive control is most effective when it is applied to variables that are both measurable and operationally meaningful. Feed rates, metabolite levels, and environmental conditions meet this criterion, whereas phenotype and functional potency do not. By focusing on stabilizing these controllable elements, manufacturers can use adaptive control to improve consistency of execution without overextending claims about its ability to govern biological quality. In this sense, practical deployment of adaptive control today is less about achieving autonomy and more about reinforcing disciplined, data-informed process control within established manufacturing frameworks.
What Would Be Required for True Closed-Loop Quality Control
Moving from adaptive control of process conditions to true closed-loop control of product quality would require advances that extend beyond control algorithms themselves. The central limitation today is not the absence of automation but the lack of real-time access to biologically meaningful attributes. Reviews of monitoring solutions for CAR-T manufacturing emphasize that tools capable of tracking differentiation state and phenotype during culture remain underdeveloped and are largely confined to offline characterization. Without continuous insight into these attributes, control systems cannot directly regulate the biological properties that define therapeutic performance.9
A second requirement is improved measurement of viability and related indicators of biological health. Current sensing technologies can infer aspects of cell growth and metabolism, but robust online determination of viability remains difficult, and few commercial methods exist that can be applied reliably at industrial scale. Until viability can be measured continuously and with regulatory confidence, adaptive systems must rely on indirect proxies rather than on direct indicators of whether cells remain fit for therapeutic use.10
Even with advances in sensing, true closed-loop quality control would still depend on validated control strategies embedded within formal quality system governance. Regulatory guidance makes clear that any approach using process data to support release or quality evaluation must operate within an approved control strategy, supported by defined models, qualified instruments, and documented decision logic. These elements must be subject to quality risk management, lifecycle monitoring, and change control rather than treated as experimental or autonomous features.
Collectively, these requirements point to a future in which adaptive control becomes more tightly linked to biological understanding and quality oversight. New PAT capable of monitoring phenotype and differentiation, more reliable online viability sensors, and rigorously validated control strategies would be necessary to shift adaptive control from stabilizing process conditions toward directly managing quality attributes. Until those capabilities mature, closed-loop quality control in T cell manufacturing will remain an aspirational goal that depends as much on advances in measurement science and regulatory integration as on progress in automation and control theory.
Conclusion – Engineering Tool, Not Biological Panacea
Adaptive control systems have demonstrated their value as engineering tools for stabilizing process conditions in T cell manufacturing, particularly in relation to growth dynamics and environmental regulation. Evidence from model predictive control strategies and closed-loop bioreactor implementations shows that feedback mechanisms can guide cell cultures along desired trajectories and support more consistent proliferation. At the same time, reviews of sensing technologies make clear that these systems primarily influence measurable process variables rather than CQAs like phenotype, differentiation state, and functional potency.
Regulatory frameworks reinforce this distinction between process stabilization and biological control. Guidance on RTRT and PAT positions data-driven control strategies as components of an approved control strategy rather than as steps toward autonomous manufacturing. Process data may be used to strengthen quality evaluation, but only within predefined specifications, validated models, and quality system governance. In this context, adaptive control enhances the evidence base for decision making without displacing the authority of the PQS or the role of quality assurance in release determinations.
The future trajectory of closed-loop control in T cell manufacturing will therefore depend less on increasingly sophisticated algorithms and more on advances in measurement science and decision frameworks. New PATs capable of monitoring biologically meaningful attributes, along with more reliable online indicators of viability and functional state, would be required to move beyond indirect proxies toward true quality-centered control. Equally important will be the integration of these tools into validated, risk-based control strategies governed by quality systems and life cycle management principles.
Seen through this lens, adaptive control is best understood not as a biological panacea but as a disciplined extension of process engineering. Its current strength lies in improving consistency of execution and reducing operational variability, not in guaranteeing therapeutic performance. As sensing technologies mature and regulatory-aligned decision frameworks evolve, adaptive control may play a larger role in quality assurance. For now, its value is clearest when it is applied where engineering control is possible, biological uncertainty is acknowledged, and regulatory accountability remains firmly in place.
References
1. Guideline on Real Time Release Testing. European Medicines Agency. 29 Mar. 2012.
2. Guidance for Industry: PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance. U.S. Department of Health and Human Services. Sep. 2004.
3. 21 CFR 211.165 — Testing and release for distribution. Code of Federal Regulations. Accessed 26 Jan 2026.
4. Egan, Joseph R, et al. “A digital shadow of CAR T cell expansion in a perfusion bioreactor: Informing optimal harvest times for autologous cell therapy.” Biotechnol. Prog. 41: w70045 (2025).
5. Van Beylen, Kathleen, et al. “Lactate-Based Model Predictive Control Strategy of Cell Growth for Cell Therapy Applications.” Bioengineering (Basel) 7: 78_(2020).
6. Li, Matthew, et al. “Closed Loop Bioreactor System for the Ex Vivo Expansion of Human T-Cells.” Cytotherapy. 21: 76-82_(2018_C).
7. Chen, Gong. “Viable cell density on-line auto-control in perfusion cell culture aided by in-situ Raman spectroscopy.” Biochemical Engineering Journal. 172: 108063_(2021).
8. Philip, B, et al. “Improving Cell Therapy Manufacturing Robustness with Real-Time Analytics.” Cytotherapy. 27: S154_(2025).
9. Shoshi, Arber, et al. “An Analysis of Monitoring Solutions for CAR T Cell Production.” Healthc. Technol. Lett. 12: e70012 (2025).
10. Rosner, Laura S, et al. “Sensors and Techniques for On-Line Determination of Cell Viability in Bioprocess Monitoring.” Bioengineering (Basel) (: 762_(2022).
11. Moutsatsou, P, et al. “Automation in cell and gene therapy manufacturing: from past to future.” Biotechnol. Lett. 41: 1245–1253_(2019).
12. Challener, Cynthia A. “Control Strategies for Perfusion Cell Culture.” BioPharm International. 2 Nov. 2020.
13. Aleynik, Boris and Mindy Miller. “Real-Time Analytics in Automated Cell Therapies Manufacturing.” American Pharmaceutical Review. 11 Mar. 2024.
14. ICH Q9 — Quality Risk Management. International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use. 1 Mar. 2025.
15. ICH Q10 — Pharmaceutical Quality System. International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use.
