How do you diagnose a psychiatric illness without a clear biological marker?
Researchers at Johns Hopkins University have grown patient-derived brain organoids to study neural firing patterns in schizophrenia and bipolar disorder.
By pairing these “mini-brains” with machine learning, the team could distinguish between healthy and patient samples with up to 92% accuracy, hinting at a path toward more objective diagnosis and treatment.
Why schizophrenia and bipolar disorder remain hard to diagnose
Schizophrenia and bipolar disorder affect millions of people worldwide and remain difficult to diagnose. Clinicians still rely on observing symptoms and behavior because there are no objective biomarkers. Medications are prescribed by trial and error, and it can take months before a patient finds something that works.
Postmortem brain studies have shown changes in both conditions. In schizophrenia, there are consistent signs of reduced activity in GABAergic neurons, which help balance brain activity. In bipolar disorder, glial cells are often reduced and genes involved in immune signaling and synaptic function appear disrupted. These findings give hints about what might be going wrong, but postmortem tissue offers only a static snapshot. It cannot show how neural networks form, change or misfire over time.
Molecular and genetic studies have highlighted differences between schizophrenia and bipolar disorder, yet little is known about how neural firing patterns separate the two.
“Schizophrenia and bipolar disorder are very hard to diagnose because no particular part of the brain goes off. No specific enzymes are going off like in Parkinson’s, another neurological disease where doctors can diagnose and treat based on dopamine levels, even though it still doesn’t have a proper cure,” said corresponding author Dr. Annie Kathuria, an assistant professor at Johns Hopkins University.
Induced pluripotent stem cells (iPSCs), created from a person’s skin or blood, can be turned into cerebral organoids – mini-brains that develop multiple types of neural cells and form connections. These models keep the patient’s genetic background and can be observed over months as they fire and wire up. They give researchers a way to study brain activity dynamically.
The new study set out to find electrophysiological “signatures” of each condition to understand their differences.
Brain organoids reveal neural firing in schizophrenia and bipolar disorder
Kathuria and the team used iPSCs from people with schizophrenia, bipolar disorder and healthy controls and then coaxed them into two types of brain models. The first were cerebral organoids, which capture features of the prefrontal cortex – the region tied to higher cognitive functions. The second were two-dimensional interneuron cultures, simplified sheets of cortical networks that are easier to record from.
To track how these neurons fired and talked to each other, the researchers placed them on chips fitted with grids of electrodes. These multi-electrode arrays detect tiny electrical impulses, acting like a very small electroencephalography. The team measured baseline activity, then applied a light electrical pulse to see how the networks responded when pushed.
“At least molecularly, we can check what goes wrong when we are making these brains in a dish and distinguish between organoids from a healthy person, a schizophrenia patient or a bipolar patient based on these electrophysiology signatures. We track the electrical signals produced by neurons during development, comparing them to organoids from patients without these mental health disorders,” said Kathuria.
The recordings were run through a computer analysis designed to pick out patterns in how signals flowed across the network. A machine learning model was then trained on these patterns to see if it could tell the difference between organoids from healthy donors and those from patients.
In the simpler two dimensional networks, control and schizophrenia samples could be separated with 94% accuracy at baseline, rising to 96% after stimulation. In the organoids, accuracy reached 83% across all 3 groups at baseline and 92% after stimulation.
The extra stimulation revealed deficits that were harder to see at rest. Schizophrenia organoids showed slower timing and weaker responses, while bipolar organoids showed irregular bursts of activity.
Overall, the accuracy outperformed structured clinical interviews, which usually sit at ~80%.
The future of schizophrenia and bipolar disorder treatments
Brain organoids could one day provide objective biomarkers for schizophrenia and bipolar disorder. Instead of relying on symptoms alone, doctors might be able to grow organoids from a patient’s own cells to confirm a diagnosis.
“Our hope is that in the future we can not only confirm a patient is schizophrenic or bipolar from brain organoids, but that we can also start testing drugs on the organoids to find out what drug concentrations might help them get to a healthy state,” said Kathuria.
The same models could then be used to test how that patient’s neurons respond to different drugs before treatment begins. That would avoid the current trial-and-error process, which can take months and often leaves patients struggling in the meantime.
“Clozapine is the most common drug prescribed for schizophrenia, but about 40% of patients are resistant to it. With our organoids, maybe we won’t have to do that trial-and-error period. Maybe we can give them the right drug sooner than that,” said Kathuria.
However, while the classification accuracy is promising, the approach is still far from being used in clinics. Future work will need larger, more diverse groups of patients and more standardized lab methods
Reference: Cheng K, Williams A, Kshirsagar A, et al. Machine learning–enabled detection of electrophysiological signatures in iPSC-derived models of schizophrenia and bipolar disorder. APL Bioeng. 2025. doi: 10.1063/5.0250559
This article is a rework of a press release issued by Johns Hopkins University. Material has been edited for length and content.