Novel visualization method helps make sense of large neuronal activity datasets

Recent technological advances opened exciting possibilities for neuroscience, enabling the collection of increasingly detailed neural data. Making sense of the large number of neural recordings gathered by neuroscientists worldwide, however, has so far proved more challenging.

Researchers at the Howard Hughes Medical Institute (HHMI) Janelia Research Campus have developed Rastermap, a new computational method that could help to better visualize recordings collected from many neurons at once. This method, outlined in a paper published in Nature Neuroscience, was initially applied to recordings gathered from the mouse and monkey cortex, rat hippocampus, zebrafish brain and even artificial neurons from neural networks.

“About 10 years ago, we were starting to have access to much larger datasets of hundreds, thousands and sometimes tens of thousands of simultaneously recorded neurons,” Marius Pachitariu, senior author of the paper, told Medical Xpress.

“This was motivated by an understanding that we need to observe many neurons ‘working together’ at the same time in a circuit to really understand some of the fundamental features of neural computation. Engineers worked with neuroscientists to create the kinds of recording devices that can monitor neural activity in this way, and computational experts created tools to process these vast amounts of data.”

Interdisciplinary collaborations between neuroscientists and engineers have led to the collection of countless neural recordings, in which the activity of many neurons is often detected simultaneously. In these recordings, each individual neuron has its own distinct activity patterns unfolding over time, typically consisting of tens of thousands of datapoints per neuron.

“In these recordings, every neuron constitutes a dimension of neural activity in a neural space, and when you have 10,000 neurons you have 10,000 dimensions,” explained Pachitariu. “The problem is that we are not very good at visualizing neural activity in such high-dimensional spaces. That was the motivation for creating Rastermap.”

The key objective of the recent study by Pachitariu and his colleagues was to develop a visualization method that allows neuroscientists to easily produce familiar-looking plots (i.e., raster plots), which clearly map large amounts of multi-neuron data. The method they developed, called Rastermap, primarily relies on an ordering algorithm.

“Suppose that you have 20 cones and have to order them based on similarity,” said Pachitariu. “First you might notice that they are of different sizes, thus ordering them based on that.

“Easy enough, but then you notice that they are also different random colors, and they would also look nice ordered by color. So, you adjust the ordering a little bit to put more similar colors next to each other, but then realize they also have slightly different shapes (e.g., some are flatter and some pointier) and you really want to take that into consideration as well, so you further change the ordering.

“Now, instead of 20 cones with the properties of size, color and aspect ratio, we have 50,000 neurons with more abstract properties, like firing rates, responses to external stimuli, correlations with the animal’s movements etc.”

Rastermap takes properties of individual neurons and tries to order them in ways that make sense. Its underlying algorithm processes data similarly to how humans would order cones in the analogy mentioned above. Starting from a random order, the algorithm continuously shifts neurons around, placing them closer to other neurons with similar activity patterns.

“Rastermap continues this process for very many iterations, using slightly smart algorithms, and at the end you have a nice ordering,” said Pachitariu. “Finally, what we do with this ordering is what matters the most: We use the order to display the neurons firing rates in a matrix, where we took each neuron’s firing trace as a function of time (a long horizontal trace through the matrix for each neuron) and we moved these around according to the ordering, so that neurons with similar traces are next to each other.”

Ultimately, Rastermap produces a neat plot, where groups of neurons with similar activity profiles are placed next to each other. This allows researchers to quickly make sense of dense and extensive neural data, which can in turn lead to new interesting discoveries.

“Our visualization method works well because neurons in the brain are not completely independent from one another: They share certain patterns of activity, but often the patterns they share are not with their nearest neurons in the tissue, but rather with neurons relatively far away that happen to have similar activity,” said Pachitariu.

“It also works well because single neurons tend to be quite noisy, so just looking at one of them in isolation does not really let you ‘see’ the responses to a specific stimulus or behavior, but when you have 20–50 of these neurons with similar patterns, their average is much easier to see on a single-trial basis. “

As part of their recent paper, Pachitariu and his colleagues used their method to visualize data collected in past studies, including simultaneous recordings of multiple neurons in the mouse cortex, as well as neurons recorded across the whole zebrafish brain.

In both these cases, Rastermap appeared to present the previously reported results in clearer and more visually appealing ways. The researchers have also started using Rastermap in other studies carried out in their lab, which yielded new interesting results.

“We think Rastermap will become increasingly useful as scientists record more and more neurons, which is bound to happen,” said Pachitariu. “We hope it will support a discovery-based approach to science, which has traditionally been a strong driver of progress in neuroscience, simply because we often do not know what neuronal properties to look for, and we stumble across interesting neural properties mostly by accident.

“Rastermap gives you a chance to do that kind of research in the era of large-scale neural recordings.”

The new visualization method introduced by this team of researchers could soon be used by other neuroscientists worldwide to make sense of large datasets tracking the activity of several neurons simultaneously. This could help to gather new insight about the function of specific neurons, as well as connections between different parts of the brain.

“Possibly one day, when large-scale recordings arrive to clinical settings, Rastermap could allow scientists to read out and interpret the patterns of neural activity in human brains, for example, to make things like BCI more effective and more easily interpretable,” said Pachitariu.

Building on their recent efforts, Pachitariu and his colleagues are now working to develop more visualization techniques that could advance neuroscience research. Concurrently, they are testing the methods they developed in collaboration with neuroscientists and medical researchers at the HHMI Janelia Research Campus.

“To quote a recent Nobel laureate: to deal with 14-dimensional spaces (or much larger), visualize (in your head) a three-dimensional space and say 14 to yourself really loudly,” added Pachitariu. “This is even much harder when you need to visualize a 50,000 dimensional space, so we need methods to span from spaces we cannot intuitively visualize to spaces we can.

“And we need to make sure we do not ‘throw the baby out with the bathwater’ so to say when we do these simplifications, because the easiest way to simplify is to just throw out most of your data. That is what PCA does to neural data for example. PCA is a simple and popular dimensionality reduction algorithm, but we probably need to move beyond that. “

Provided by
Science X

© 2024 Science X Network