Teaching AI to uncover hidden disease signals in neuronal networks

Flagship Integrative Neuromedicine

Neurological disorders can change the way brain cells communicate. Yet when researchers study networks of neurons, these disease-related changes can be difficult to distinguish from normal biological variation, because no two living neuronal networks behave in exactly the same way. Even networks grown under similar conditions may differ in their number of cells, connectivity, maturation and spontaneous activity. As a result, researchers may overlook subtle effects of disease or mistake natural variation for a disease-related change.

Researchers within the Convergence Health & Technology Flagship Integrative Neuromedicine are developing a new way to bring such hidden functional differences to light. By combining advanced imaging, light-based stimulation, computational modelling and artificial intelligence, they aim to move beyond simply observing neuronal networks. Instead, they want to actively challenge them and learn from how they respond. The aim is to develop a more informative way to study how neurological disorders affect the functioning of neuronal networks. If the method proves successful in disease models, it could eventually also help researchers evaluate whether potential treatments restore network function. The platform is still at an early stage and must first be tested in living neuronal networks.

From observing a network to actively challenging it

Researchers usually study neuronal networks by observing their spontaneous activity or by applying a stimulus selected in advance. The new platform takes a more active approach by repeatedly observing and challenging the network. The approach can be compared to a physiological stress test. Two networks may appear similar when left undisturbed but respond differently when challenged. These responses could provide a functional phenotype: a profile of how a network functions under challenge. The researchers want to find out whether such profiles can help distinguish disease-related changes from normal biological variation.
.
.

How does the AI-guided stress test work?

The researchers combine several technologies. Neurons communicate through rapid electrical signals, produced by changes in voltage across their cell membranes. Voltage imaging uses fluorescent sensors whose brightness changes with this voltage. A microscope records these changes, allowing researchers to follow electrical activity across many cells at once. Computational models help describe how this activity develops and spreads through the network. Using optogenetics, researchers can then stimulate selected cells or parts of the network with precisely controlled light.
.
.The distinctive element is the use of reinforcement learning, a form of artificial intelligence that learns from the effects of previous actions. Rather than following a fixed sequence, the AI uses what it has observed so far to select the next light stimulus. The resulting response is recorded and analysed, after which the cycle begins again. This creates a closed loop in which each response helps shape the next step of the experiment. In this way, the platform is designed to actively search for the experiment that best reveals the underlying dynamics of a particular network.

Biology never gives us two identical networks. If we only look for a fixed pattern, we risk confusing normal variation with disease. Our aim is to let AI actively probe a network and identify which responses reveal a real change in its underlying dynamics.

Daan Brinks

Associate Professor of Imaging Physics at TU Delft and TU Delft lead of Integrative Neuromedicine

Moving separate technologies to the first closed-loop experiment

The project is now making the transition from separate technical demonstrations to an integrated biological experiment.

One important building block is the Octoscope developed at TU Delft (link). This compact microscope can both record electrical activity through voltage imaging and deliver precisely patterned light stimulation. Its imaging and stimulation capabilities have already been demonstrated in laboratory samples and living organisms.

In parallel, in collaboration with Wendelin Böhmer and Laurens Engwegen at TU Delft, reinforcement-learning agents have been tested in simulated neuronal networks. These agents learned to control activity in networks they had not encountered during training, despite differences in their hidden connectivity and topology. This work is available as a preprint (link). A methodological follow-up (link) has extended this ability to generalise across previously unseen systems.

A third building block is the rapid and reliable analysis of voltage-imaging data. Together with the lab of Zhenyu Gao, the Erasmus MC lead of the Flagship, and Riu Silva the researchers have systematically evaluated the reliability of existing voltage-imaging analysis methods (link). They examined how different steps in the analysis can affect the reliability and interpretation of the results. Building on this work, a follow-up real-time analysis platform is now operating in live mice. Real-time analysis is essential for the closed loop. The platform must be able to interpret the network’s response while the experiment is still running so that the AI can decide what stimulus to apply next.

The researchers are now integrating these components on the Octoscope. The next milestone is the first experiment in which voltage imaging, real-time analysis, reinforcement learning and optogenetic stimulation operate together in a closed loop in a cultured neuronal network. This experiment must demonstrate that the platform can observe a biological network, select a stimulus, record the response and adapt the next step while the experiment is running. Once the closed loop works reliably, the researchers can begin investigating whether the AI-guided stress test reveals meaningful differences between healthy and disease-related network models.

What could this lead to?

In the nearer term, the platform could give researchers a more informative way to evaluate cellular disease models. It could help them study how disease alters communication between cells and measure how networks respond to potential treatments. The platform could also help determine whether a disease model reproduces the relevant network-level phenotype, rather than only a molecular marker or change in cell appearance.

In the longer term, such functional profiles might contribute to distinguishing between disease subtypes, selecting treatments or recognizing disease-related changes earlier. Reaching that stage will require validation in specific disease models and, subsequently, in networks derived from patient cells.

The present work provides the technological and scientific foundation for exploring these future applications.

Linking cell behavior to underlying biology

The concept also builds on an earlier collaboration between the Daan Brinks lab at TU Delft and the Miao-Ping Chien lab at Erasmus MC. Their collaboration demonstrated that researchers can use live imaging to follow living cells over time and select cells based on their behavior, such as how they move or change shape. The selected cells can subsequently be analyzed using single-cell sequencing. This technique reveals molecular information about each individual cell, including which genes are active. This made it possible to link functional cellular phenotypes (what cells actually do) to their molecular profiles (the biological processes active inside them) (link). Miao-Ping Chien, a leading researcher in this field, now applies this approach in cancer research.
.
.The current project extends this underlying principle from cellular behaviors such as migration and morphology to electrical signaling and collective network dynamics. In other words, instead of studying the movement or shape of individual cells, the researchers examine how neurons communicate and function together as a network. Building on this, a possible future extension would be to combine these functional network phenotypes (patterns in how a network functions and responds) with genetic or molecular information. This could help researchers understand not only which biological changes are associated with disease, but also how those changes alter the functioning of a living network.

The added value of Convergence

Developing the platform requires expertise that does not usually sit within a single research group. TU Delft contributes imaging technology, voltage imaging, computational modelling and reinforcement learning. Erasmus MC contributes cellular neuroscience, biological disease models, real-time analysis and expertise in connecting dynamic cellular behavior to molecular information.
.
Convergence makes it possible to develop these elements as one experimental system rather than as separate technical components. In particular, the collaboration with the Gao lab provides the bridge from AI developed in simulated environments to biologically meaningful, live experiments, and the collaboration with the Chien lab to diagnostic contexts.
.
The involvement of Erasmus University Rotterdam adds another important perspective. Together with Maren Wehrle, the EUR lead of Integrative Neuromedicine, researchers regularly discuss the ethical implications of AI-guided functional profiling. They consider what future classifications could mean for patients and how patients can be involved as co-design partners. This includes deciding which questions to pursue, which outcomes matter and how the technology might ultimately be applied if the research moves closer to clinical use.

Collaborators

Daan Brinks — TU Delft; TU Delft lead of the Convergence Integrative Neuromedicine Flagship; BIOLab
Wendelin Böhmer — TU Delft; reinforcement learning; BIOLab
Laurens Engwegen — TU Delft; reinforcement-learning research; shared PhD researcher in the Brinks and Böhmer groups
Zhenyu Gao — Erasmus MC lead of Integrative Neuromedicine; cellular neuroscience and voltage-imaging analysis
Rui Silva — Erasmus MC and TU Delft; real-time analysis; shared PhD researcher in the Brinks and Gao groups
Miao-Ping Chien — Erasmus MC and Oncode Institute; functional genotype–phenotype coupling
Maren Wehrle — Erasmus University Rotterdam; ethics and co-design

Octoscope developed at TU Delft (link).