Machine learning brings new insights to cell’s role in addiction, relapse

(Press-News.org) Object recognition software is used by law enforcement to help identify suspects, by self-driving cars to navigate roadways and by many consumers to unlock their cell phones or pay for their morning coffee.

Now, researchers led by the University of Cincinnati’s Anna Kruyer and the University of Houston’s Demetrio Labate have applied object recognition technology to track changes in brain cell structure and provide new insights into how the brain responds to heroin use, withdrawal and relapse. The research was published April 30 in the journal Science Advances.

Study background

Kruyer’s lab focuses on relapse to heroin use, as many overdose deaths occur when people overestimate their capacity for drug use during relapse. The team has developed an animal model of relapse over the past seven years, studying interactions between brain cells and the reward center of the brain that orchestrates the relapse process.

“We want to understand the neurons that are involved and all of the different cells and molecules that can shape that activity,” said Kruyer, PhD, assistant professor in the Department of Pharmaceutical Sciences in UC’s James L. Winkle College of Pharmacy. “The idea would be if you can interfere with relapse, you can help someone stay clean.”

While neurons are a more commonly studied brain cell, Kruyer has focused on another cell called an astrocyte. Astrocytes have many functions, including metabolic support for neurons, providing molecules that neurons turn into neurotransmitters, and shielding or uncovering different receptors during synaptic activity.

“Astrocytes are a kind of protective cell that can restore synaptic homeostasis,” Kruyer said. “They are super dynamic relative to the synapse, and they’re moving toward and away from the synapse in real time in a way that can impact drug seeking. So if you prevent this reassociation with synapses during relapse, you can increase and prolong relapse."

Labate is an applied mathematician with expertise in harmonic analysis and machine learning.

“A central focus of my research is the development and application of mathematical techniques to uncover meaningful patterns in non-Euclidean data, such as the analysis of complex shapes,” said Labate, PhD, professor in the University of Houston Department of Mathematics.  “The study of astrocytes provides an ideal setting for this type of investigation: these cells are highly heterogeneous, varying widely in size and shape, and are capable of dynamically remodeling their morphology in response to external stimuli.” 

A new approach with machine learning

While animal model studies have produced results, Kruyer and her colleagues faced a barrier in that the techniques used could not be translated for human subjects. To work around this issue, they focused on an astrocyte protein that essentially acts as the cell’s skeleton.

“We thought if we could figure out a way to translate what we’re seeing at the synaptic level to changes in the cytoskeleton, maybe we could see if astrocytes are doing something critical during relapse in humans,” Kruyer said.

A team of mathematicians led by Labate trained object recognition machine learning models on hundreds of astrocyte cells until the technology could accurately detect an astrocyte within an image, similar to how object recognition software works.

“Machine learning techniques have been widely applied in the literature to image classification tasks, where the objective is to assign each cell to a specific category,” Labate explained. “In such contexts, machine learning is particularly powerful for identifying image-based cellular features that are difficult to capture using traditional geometric descriptors, yet serve as effective discriminators between classes.” 

Once the program could identify astrocytes, the team trained it to analyze specific structures based on 15 different criteria, including astrocyte cytoskeletal density (similar to bone density), size, length versus sphericalness and number of smaller branches coming off of the main branch.

“You can think about this like if you gave a computer a bunch of images of street scenes, it would commonly see pedestrians, cars and buildings,” Kruyer said. “If you give a computer 1,000 images of astrocytes, there are things it would commonly see. This is the segmentation process whereby a computer can now start to make measurements of the different features of the astrocyte.”

Using all 15 measurements weighted by their importance in the computer’s precision to detect astrocytes, researchers developed a single metric to quantify the characteristics of each astrocyte.

“In previous work, I have utilized machine learning for both cell classification and segmentation problems,” Labate said. “In this paper, however, we address a more nuanced question: are there specific subpopulations of astroglia that exhibit more pronounced morphological changes compared to the rest? To investigate this, we introduced the concept of distance to compare the shape characteristics of individual astrocyte cells while accounting for the inherent heterogeneity within the population.”

Applying the model

After developing the machine learning model to identify astrocytes and report the new metric, the team looked at astrocytes specifically within an area of the brain called the nucleus accumbens (NAc) that is active during drug relapse.

The model was able to predict exactly where in the NAc an astrocyte came from based on its structure with 80% accuracy.

“This tells us that astrocyte structure varies by anatomy,” Kruyer said. “Astrocytes have been considered to be this homogenous type of cell, but this indicates to us that astrocyte structure varies significantly by location — perhaps the shape and the size have something to do with their function.”

Using animal models and the new knowledge gained from the computer models, the team found that astrocytes appear to shrink and become less malleable after exposure to heroin.

“These data suggest that heroin is doing something molecularly that makes astrocytes less able to respond to synaptic activity and maintain homeostasis,” Kruyer said.

“This paper exemplifies the strength of interdisciplinary collaboration, where innovative quantitative tools are developed or adapted to tackle complex biological questions,” added Labate. “The success of this research lies in the effective communication between disciplines and in our willingness to push the boundaries of traditional machine learning to address biologically meaningful and timely challenges.”

Next steps

Kruyer said she is most excited about the application of machine learning to a biological question, which eliminates human error and biases and makes the research more easily translatable from animal models to humans.

“We’re asking open-ended questions, and it’s giving us all of these really fine-grained detailed answers, and then what we do with that is up to us,” she said. “Human astrocytes are much larger, much more complex and way more abundant than in the animal models, so applying a tool like this is really cool to carry forward in humans.” 

Moving forward, the team wants to learn more about the specific mechanisms of astrocytes in each region within the NAc and train new models using human tissue samples. Long term, the knowledge gained could help develop new treatments for addiction focused on restoring or replacing astrocytes to their functions prior to being exposed to heroin.

Additionally, the machine learning method Labate’s team developed can be adapted and applied to other types of cells with intricate structures.

“By enabling precise quantification and comparison of single-cell morphological features, this approach opens the door to the development of novel techniques for identifying cellular or molecular biomarkers that reflect biological processes, disease states or responses to therapeutic interventions,” he said. “More broadly, our work introduces a new quantitative framework for uncovering and validating fundamental mechanistic models underlying complex brain conditions, such as addiction to drugs of abuse.”

Other coauthors include Michaela Marini, Yabo Niu, Heng Zhao, Anish Mohan and Nathan Koorndyk. This work was supported by grants from the Simons Foundation (MPS-TSM783 00002738 to Labate) and the National Institutes of Health (DA054339 to Kruyer and NS007453 to Koorndyk). All authors declare no competing interests related to this manuscript.

END