Inside the human brain, beneath an impenetrable skull, billions of neurons communicate constantly across an incomprehensibly complex, ever-changing signaling network. That makes the brain a uniquely difficult organ to study.
For many years, scientists have used genetically-encoded protein-based fluorescent indicators (GEFIs) – or protein sensors, for short – to map and make sense of the nonstop neuronal activity that goes on in our brains and to understand the breakdowns that lead to neurodegenerative diseases.
Andre Berndt, PhD, an associate professor of Bioengineering and faculty member in the Institute for Stem Cell and Regenerative Medicine (ISCRM), leads a research lab determined to push the boundaries of protein-based sensor technology. For Berndt, that means creating a real-time view of neural patterns down to the single-cell level, like a traffic camera with the capacity to track the live movement of individual cars.
Protein sensors are programmed to flash in response to certain biological activities, for example, a change in levels of a particular chemical, like calcium. For that reason, protein sensors must be engineered to perform specific tasks, which requires an ability to predict the necessary structure of the protein, a job that scientists like Berndt are naturally assigning to powerful computers.
In a 2023 paper published in the journal ACS Sensors, the lab unveiled a tool, dubbed Opto-Mass, that is capable of simultaneously testing thousands of versions of a protein sensor in mammalian cells, and identifying a variant in a library of options that is best suited for a particular experiment. This level of optimization is many magnitudes faster than current trial-and-error methods.
Berndt’s team followed up that study with a paper, published the following spring in Nature Computational Science, that described a trained learning model with the power to correctly predict variants of the calcium indicator GCaMP with record-setting speed and accuracy, outperforming all previous generations of these sensors, a breakthrough that marked a significant leap in functional protein engineering and accelerates the development of high-performing optogenetic sensors with broad applications across diverse protein engineering challenges.
Closing the Performance Between Green and Red Sensors
Now, a five-year, $11 million grant from the NIH BRAIN Initiative will enable the Berndt Lab, in collaboration with researchers from Princeton University, the Cleveland Clinic, Northwestern University, and the University of Maryland to enhance protein sensor technology farther still. The outcome will give scientists improved tools to observe brain activity in real time with even greater precision.
Funding will allow the researchers to address a performance gap between two types of sensors: green calcium sensors that glow when calcium levels rise inside a cell and red calcium sensors that turn red when calcium levels rise. Green sensors are typically brighter and faster for detailed, close-up measurements, while red sensors are better for imaging deeper in tissues and for use in combination with other tools.
The team aims to engineer new red calcium and neuromodulator sensors that match or exceed the performance of green ones by combining the Opto-MASS platform, which can test thousands of sensor variants at once, with machine learning models to predict which designs will work best. The most promising sensors will then be rigorously tested in real biological systems. By dramatically speeding up sensor development and improving performance, this work could give scientists much more powerful tools to study brain function and neurological disease.
Automating a Brute Force Process
According to Dr. Berndt, the first step is to create a dataset encompassing known attributes of thousands of sensor variants, a task that will be led by Dr. Doug Fowler, a professor in the UW department of Genome Sciences. “Machine learning models depend on data,” says Berndt. “You can have the best models, but if the data we use to train our model is not good, it won’t accurately predict which mutations in the protein-based sensors lead to the gain-of-function variants we need.”
Over the course of the grant, the researchers hope to screen up to 100,000 sensor variants using their high-throughput technology. A consumer analogy might be shopping for a car in a market where thousands of optimizable options exist. In this scenario, the buyer would use the machine learning model to identify or “engineer” the right vehicle for a particular set of needs, like fuel efficiency, ruggedness, cargo space, or comfort.
For the researchers, the important features – or phenotypes – are qualities like brightness, affinity, or penetration. These attributes are referred to as gain-of-function mutations for the sensor variant. The trick is to reach a point where scientists can harness the best of both types of sensors to better track and image neuronal activity in the brain.
“Right now, you can use green and red sensors in combination and image them together in real time, but with limitations.” says Berndt. “It’s important to know that the red sensors are performing just as well as the green sensors, and that any differences in the signal you are seeing are biological readings and are not related to the sensor itself.”
In practical terms, the goal is to be able to radically accelerate the process of engineering optimal sensors for studying how the brain works. During the grant period, researchers in the lab of Dr. Christina Kim at Princeton will test the sensors developed by machine learning models in the brains of mice.
“We are setting out to develop an integrated platform that combines machine learning and high-throughput screening to bring protein sensor engineering to the next level,” says Bernt. “If we’re successful, we’ll be able to automate what has been a brute-force, labor-intensive approach and provide the neuroscience field with a broader range of sensors that help them understand how to help patients suffering from neurodegenerative disorders.”
.
.