New Math Tools Offer Truer Picture of Brain Activity

Rockefeller University

When a firework bursts in the sky, what often follows is a cloud of smoke in the shape of the firework—a visual echo that lingers long after the spectacular explosion itself is over.

A similar phenomenon happens in the brain when a neuron fires: The explosion of activity causes a change in calcium levels, resulting in a fluorescence of light that neuroscientists use as a proxy for brain activity. Unfortunately, however, much like that cloud of smoke, the calcium indicator only approximates the original's shape, size, and origin.

Neuroscientists have tried to correct for this data problem using a mathematical algorithm borrowed from astrophysics. Merav Stern, a physics fellow at Rockefeller University, has a better idea: three entirely new computational algorithms that recover the "fireworks" in the data. She published her work in Cell Reports Methods.

"All the methods were superior to using fluorescence as a measure of the neural spiking rate, offering a more reliable depiction of the underlying neural dynamics," she says. "And because each algorithm has its unique strengths, neuroscientists can choose the method that works best for their data set."

A loan from astrophysics

Wide-field calcium imaging, captured by one-photon microscopy, is a common technique for recording brain activity across large cortical areas in a largely noninvasive way, allowing researchers to observe how behavior connects to brain activity and vice versa via light. The method relies on changes in calcium levels that occur during cellular activity, often via the synthetic protein GCaMP, which fluoresces neon green during the calcium shift.

"It's a powerful tool, but it also has limitations," notes Stern, a theoretical physicist who uses mathematical modeling to better understand the complex networks of the brain. "Conventional One-photon imaging can capture large cortical areas but can't record single neurons. The light signals can also get contaminated by blood flow or the lag time between neural firing and calcium fluorescence, which can result in data that shows neural activity in the wrong temporal locations. This leads to events such as 'ghost correlations,' which suggest simultaneous activity in different brain regions when really there is no overlap."

Neuroscientists have long attempted to mathematically correct for these complications using the Lucy-Richardson algorithm. Named after the two researchers who independently developed it in the early 1970s for use in astrophysics, where vast distances and powerful forces result in blurry images of space objects, this type of so-called deconvolution has been very useful when observing the universe but has proven less so in neuroscience.

One drawback is that it relies on the assumption that statistical "noise" in the data falls into a Poisson probability distribution, which Stern suspected may not hold true for noise in brain activity.

"While the noise sources in these space signals are relatively well established, the noise sources in wide-field imaging data are often a mix of signals or even unknown," she says. "They can be caused by tissue scattering the light, by the camera itself, or have other origins."

Many neuroscientists have tried to engineer their way around these problems by developing ever-better brain imaging technology, but Stern wondered if there was a way to get around having to use fluorescence as a proxy for neural activity by creating algorithmic alternatives tailored to the needs of brain imaging.

"To return to the fireworks analogy, I wondered, what if you could use mathematical tools to see the actual fireworks?" she says.

A vast improvement

Using a combination of real and synthetic brain activity data sets, Stern developed more than a dozen new algorithms before landing on three—and adapting a fourth—that were vastly superior to Lucy-Richardson when it came to cleaning up the statistical noise in wide-imaging calcium data.

Key to their creation was her discovery that the noise in brain data did not fall into a Poisson distribution but instead the most common one in nature: the bell curve, also known as Gaussian or normal distribution. Basing her models on that assumption allowed her to create models that recovered the neural activity itself rather than using calcium fluorescence as a proxy for it.

Stern named these algorithms dynamical binning, first differences, and continuously varying; she also adapted the Wiener Filter–based approach. While first differences and dynamical binning had excellent results in select applications, both continuously varying and the adapted Weiner Filter–based approach excelled across diverse types of data sets.

Moreover, beyond wide-field imaging, Stern says the methods could be used to analyze any data that includes the aggregated activity of neurons through light, such as neuropil recordings, fiber photometry, and non-spiking neurons, the last of which she's currently analyzing using data from Rockefeller's Bargmann lab . Some scientists have already used her algorithms—which can be downloaded from GitHub—to analyze and publish new and improved brain imaging data.

"I give people the tools, and they can judge for themselves which one would fit their data the best," she says. "Whichever algorithm they choose, it should improve the conclusions that can be drawn from brain activity recordings."

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.