Building A Self-Driving Lab For LNP Development

A conversation with Peter Sagmeister and Aniket Udepurkar, MIT, and Life Science Connect's Jon O'Connell

Lipid nanoparticles (LNPs) are becoming one of the most important enabling technologies behind nucleic acid therapeutics, but manufacturing them with consistent size, morphology, and quality remains a complex process development challenge.

As the field moves from discovery toward scalable, regulated production, developers need better ways to understand how formulation variables and process conditions shape critical quality attributes (CQAs) in real time.

Peter Sagmeister, a former postdoctoral associate, and Aniket Udepurkar,a postdoctoral associate in the Myerson Research Group at MIT, are working with colleagues to accelerate LNP process development through automation, process analytical technology (PAT), and advanced optimization. The "self-driving" system they’re building integrates manufacturing equipment, advanced analytics, control software, and database management to steer rapid experimentation and particle characterization toward optimal process conditions. The platform, currently undergoing peer review, is built to generate high-quality datasets that support more predictive control strategies based on real-time data.

We wanted to learn more about the team's approach, and Sagmeister and Udepurkar agreed to answer questions. They explain how Bayesian optimization can reduce experimental burden, and how automated design of experiments (DoE) could expand the platform’s future capabilities.

They also connect the work to CMC considerations, where real-time monitoring of CQAs such as particle size and polydispersity can help bridge early development and manufacturing expectations. Together, the conversation points toward a future in which LNP process development becomes faster, more material-efficient, and more aligned with the needs of autonomous manufacturing research.

Your system integrates several sophisticated components — the impinging jet mixer, real-time spatially resolved dynamic light scattering, control software, and database management. Can you talk about the virtues of these parts and how they work together in an integrated system?

Sagmeister: We have developed an automated, modular platform for engineering LNPs with controlled size and morphology. The platform follows the concept of a self-driving laboratory by integrating automated experimentation, PAT, and data-driven decision making into a single workflow.

The system is built entirely from commercially available hardware components, including syringe pumps, switching valves, and an inline dynamic light scattering (DLS) instrument serving as the PAT tool for real-time particle characterization. All hardware is connected through a common communication standard, OPC UA, and orchestrated by a central software platform that coordinates experiments, acquires process and analytical data, and stores all information in a structured database.

This integrated architecture enables three key capabilities.

First, automated execution of predefined experimental campaigns, including classical DoE studies as well as dynamic experiments in which process parameters are continuously varied to efficiently explore the design space.

Second, implementation of closed-loop optimization strategies, where real-time PAT measurements are used by decision-making algorithms, such as Bayesian optimization, to autonomously identify optimal process conditions.

Third, the resulting high-quality datasets provide the foundation for developing both data-driven machine learning models and first-principles process models, enabling improved process understanding, prediction, and control.

By combining automation, PAT, and advanced data analytics within a unified platform, the system provides an end-to-end framework for accelerated LNP process development, optimization, and autonomous manufacturing research.

Inline SRDLS during manufacturing is fairly novel. What does that spatial resolution give you that conventional offline or single-point DLS measurements cannot?

Udepurkar: Spatially resolved DLS (SRDLS) measures particle size at high concentration (typically used in industry) in batch mode or in-line. SRDLS also eliminates the need for dilution or sample preparation. A typical workflow for conventional DLS requires sample preparation and buffer exchange before the size can be measured and needs a workup of 20-30 mins. LNP can undergo significant change in size post formulation and buffer exchange, and this change cannot be captured with a conventional DLS. With measurements acquired every 10 seconds, SRDLS continuously tracks changes in LNP size in real time — a capability that conventional offline DLS fundamentally cannot provide. To date, SRDLS remains the only technique capable of inline LNP size measurement.

Leveraging this rapid, inline characterization, we could systematically investigate the influence of process parameters on LNP size. We coupled the impinging jet mixer system with SRDLS to develop a fully automated LNP pilot system — an outcome that would not have been achievable with conventional DLS.

How does the platform quickly narrow in on critical process parameters? How does it compare to traditional DoE-based development timelines?

Sagmeister: Rather than systematically exploring the entire design space, including regions that are unlikely to produce the desired CQAs, Bayesian optimization focuses on selecting the most informative next experiment based on all previously acquired data. By balancing exploration of uncertain regions with exploitation of promising conditions, it efficiently converges toward the global optimum while requiring significantly fewer experiments than traditional approaches.

Compared with conventional DoE-based development, the platform offers two major advantages. First, automated experiment execution eliminates the need for continuous operator involvement, allowing experiments to be performed around the clock and substantially reducing development timelines. Second, the platform generates structured, high-quality datasets that are automatically stored and readily available for future optimization tasks. As more experimental data are collected, the underlying surrogate models become increasingly informative, enabling the decision algorithm to identify promising process conditions more efficiently and accelerate subsequent optimization campaigns.

How will Bayesian optimization and automated DoE enable future integrations? What could they open access to that the existing platform can't yet do autonomously?

Sagmeister: Bayesian optimization and automated DoE primarily enable autonomous experimentation. Instead of relying on a researcher to manually select the next experimental conditions, the platform can automatically determine and execute the most informative experiments based on statistical methods and real-time process data.

This has two major advantages. First, Bayesian optimization can identify optimal process conditions with significantly fewer experiments than traditional trial-and-error approaches by focusing on the most promising regions of the design space. Second, because experiment planning, execution, and data acquisition are fully automated, the platform can operate continuously, enabling 24/7 experimentation and substantially increasing experimental throughput.

While the flexibility of any automated platform is still constrained by its available hardware — for example, the formulations, lipids, or buffers that can be selected automatically — these limitations also exist in manual workflows. The key difference is that the automated platform can perform experiments continuously and make data-driven decisions without requiring constant human intervention, accelerating process development and optimization.

Notably, you flag regulatory relevance as a design goal. How did you build that consideration into the platform architecture, and how do you see tools like this fitting into a CMC or quality by design framework?

Udepurkar: LNP size and polydispersity index are the main CQAs for the drug product. Inline PAT is beneficial to monitor these CQAs and ensure batches meet specifications every time. The automated system can also help us understand the relationship between critical process parameters (CPPs) and CQAs. By systematically generating data across the design space, these relationships can be quantified and used to develop robust control strategies for manufacturing. While online and inline process analytical technologies are increasingly capable of monitoring attributes such as particle size, other critical CQAs, including encapsulation efficiency, particle morphology, and payload distribution, still lack robust real-time measurement methods. Developing predictive models that link readily measurable CPPs to these more challenging CQAs could therefore enable their indirect monitoring and control during manufacturing.

The work is demonstrated on model drug delivery systems. Could you extend the platform to more complex or sensitive LNP formulations? If so, how?

Udepurkar: Yes, the platform can be extended to more complex or sensitive LNP formulations. Because the system is modular, additional hardware can be integrated depending on the application. For example, incorporating an autosampler would enable automated screening of different lipid compositions, formulations, or buffer systems without manual intervention. This would significantly expand the accessible formulation space while maintaining the benefits of automated experimentation.

About The Experts:

Peter Sagmeister is an Austrian scientist, inventor, and entrepreneur specializing in process chemistry, automation, AI, and drug delivery technologies. He earned his Ph.D. from the University of Graz and completed a postdoctoral appointment at the Massachusetts Institute of Technology (MIT), where he developed autonomous technologies for process development in small-molecule pharmaceuticals and lipid nanoparticles. He has collaborated with leading pharmaceutical companies and academic institutions, authored peer-reviewed publications, co-invented patented technologies, and received multiple scientific and innovation awards.

Aniket Udepurkar is a postdoctoral associate in the group of Proessor Allan S. Myerson at MIT, working on mRNA and lipid nanoparticles in an FDA-funded project. He holds a B.Tech. in chemical engineering from IIT Guwahati and both an M.Sc. and Ph.D. in chemical engineering from KU Leuven, where he worked in the lab of Professor Simon Kuhn. His doctoral research focused on ultrasonic microreactors for the synthesis of drug-laden polymeric nanoparticles and included a visiting appointment at MIT with Professor Klavs F. Jensen. His work spans flow chemistry, nanoparticle synthesis, and drug delivery. He has authored multiple peer-reviewed publications and presented at several international conferences.