Developing a better understanding of many brain and neurological diseases relies in part on counting and mapping the millions of specialized microscopic junctions in the brain, called synapses. Synapses enable neurons to communicate with one another, transmitting electronic and chemical signals that form the foundational networks underlying thought, movement, and memory.
But counting and mapping these synapses is a painstaking, laborious process.
Now, a team of Tufts University School of Medicine scientists has developed new, open-source machine learning software that can count and map hundreds of thousands of these synapses in minutes, a task that would take humans hundreds or even thousands of hours to complete. Importantly, the new tool, called SynAPSeg, can complete the task with human-level accuracy.
“The new platform, which we are making freely available, opens exciting new avenues to more quickly understand how neural circuits form, work, and change with age and disease,” says Alexei Bygrave, assistant professor of neuroscience at the School of Medicine and corresponding author on a paper detailing the results published in PLOS Computational Biology.
Previously, the process of counting and mapping large swaths of tissue was monumentally slow. This typically involves using histology methods to label components of the synapse, so they can then be visualized with high-powered fluorescent microscopes. For neuroscientists, it demanded hundreds of hours of tedious manual labor to map and count synapses in just one section of the brain. Such an effort was notoriously prone to eye strain, inconsistency, and subjective bias by those doing the work.
The new platform developed out of the researchers’ specific research interest in inhibitory interneurons. While most brain cells are excitatory, actively prompting neighboring neurons to fire, about 15 to 20 percent are inhibitory, regulating overall activity and suppressing unwanted noise.
Tools have existed that do a good job of analyzing excitatory synapses, because they protrude from the neuron’s dendrite, a short, branched extension of a nerve cell.
“The excitatory synapses looks a bit like microscopic brussels sprouts while still on their stalks,” says Bygrave. “But inhibitory synapses—and excitatory synapses located within inhibitory interneurons—typically sit flat against the dendrite, closely packed together.” Traditional software registers them as a blur, requiring a human’s eye to parse out their numbers in detail, he says.
Pascal Schamber, a Ph.D. student in neuroscience at Tufts Graduate School of Biomedical Sciences and lead author of the study, had to develop an enormous amount of high-quality training data to teach a computer model to distinguish individual synapses packed in these tight clusters. That meant that over a period of many weeks, Schamber and four other team members sat at their computer screens, and using a computer mouse manually outlined individual synapses from brain-image scans that contained approximately 10,000 synapses in total. More than one team member was assigned to mapping each synaptic region to ensure that any bias or disagreement among those doing the work was reconciled.
“The resulting data was then fed into the computer model to teach it how to differentiate each individual synapse like a human would,” says Schamber.
The completed tool does more than count synapses, Bygrave notes. It helps manage and structure data and allows scientists to analyze properties of synapses under various conditions and when using different labeling tools and strategies.
To test the tool’s utility in real-world research, the team evaluated a specific subpopulation of inhibitory neurons that express the protein parvalbumin within the hippocampus of the brain.
“The hippocampus is essential to learning and memory,” says Bygrave. “Comparing young adult mice with older specimens, the analysis pinpointed a measurable decline in synaptic density in the aging mouse brains. This finding aligns with a long-standing hypothesis that this phenomenon may contribute to age-related cognitive decline.”
Beyond its immediate experimental results, the researchers believe the real value of their project lies in the open-source release of both the software code base and the extensive, hand-drawn training data set that created it. Both are available on public repositories like GitHub for the global scientific community.
“In the future, researchers can use this dataset to train even more advanced diagnostic tools, which we hope will accelerate broader work by neuroscientists to map the inner workings of the brain,” says Schamber.
Co-authors of the study include Madison Pelletier, Helene Hartman, Olivia Friedman, Shiyu Zhang, Allison Blais, Seyun Oh, all from the Department of Neuroscience at Tufts University School of Medicine, and Haining Zhong of Oregon Health and Science University.