Thunder shares its underlying structure with software that the tech industry uses to customize online advertisements or to recommend music on streaming services such as Spotify.
Like the Web, neuroscience is flush with data, but the data collection threatens to outpace the ability to make sense of the bounty.
For example, sophisticated new techniques allow neuroscientists to collect massive amounts of complex data about the brain. These data may hold clues to how neurons interact with each other on a whole-brain level and how disruptions in the neural circuits relate to conditions such as autism.
But these recordings from thousands or even millions of neurons yield terabytes of data. Analysis on that scale requires tools beyond what neuroscientists are used to.
In the new study, published 27 July in Nature Methods, researchers present a toolkit that may help neuroscientists use the power of distributed computing — spreading data across many computers for rapid parallel processing.
Distributed computing is not a new concept, but Thunder uses a software system called Spark that saves the initial dataset for quick retrieval. This data caching eliminates the bottleneck of having to reload the dataset before each new analysis. Spark is open-source and can be run on a private computing cluster, available at most universities, or on a cloud computing cluster, available to virtually anyone with an Internet connection.
“A friend in the Bay Area told me [Spark] was likely to be the next big thing,” says lead investigator Jeremy Freeman, a group leader at the Janelia Farm Research Campus of the Howard Hughes Medical Institute in Maryland.
Researchers may need to inspect complex data from many statistical angles, so they require something fast and powerful. Freeman and his colleagues realized Spark could be customized for neuroscience. By last summer, just a couple of months after he had first heard of Spark, Freeman had Thunder up and running.
Running on Spark, Thunder can implement many of the statistical analyses commonly applied to neural data and can easily be modified to perform additional functions.
The input data can be any unit that evolves over time — such as the voxels, or volume-pixels, from calcium imaging or functional magnetic resonance imaging. Thunder turns this data into a form that can be cached and queried multiple times without reloading. This allows researchers to run a variety of statistical analyses, often in seconds or minutes.
Autism research stands to benefit from analyzing enormous datasets to reveal the functional connections among neurons, says John P. Cunningham, member of the Grossman Center for the Statistics of Mind at Columbia University.
“But the question is, how do you gain those scientific insights, how do you bridge that gap between the data and those huge advances?” Cunningham says.
To demonstrate Thunder’s potential, Freeman’s team analyzed functional brain images from mice and larval zebrafish. They engineered both sets of animals to bear molecules that fluoresce when neurons fire. They then looked at the brains of mice running on a treadmill and fed the data into Thunder to map how their neural activity changed with running speed.
For zebrafish larvae, they paired light-sheet microscopy of the brain with measurements of motor neuron activity in the tail and used Thunder to map responses to the larvae’s visual environment.
The light-sheet microscopy system used on the zebrafish is one example of a new imaging technology that provides high-resolution views of the brain. The method, detailed in a study published in the same issue of Nature Methods, allows imaging of the whole brain during visually driven behavior2. It employs two laser beams, one scanning the larval zebrafish brain from the front, between the eyes, and the other scanning from the side but shutting off when passing the eyes. Previous such techniques could not study visually driven behavior because the laser scanned over the eyes and compromised the visual stimuli.
Combining light-sheet microscopy with Thunder’s analytical capabilities, the researchers identified neurons that fire when zebrafish swim in particular directions and another set of neurons that do so when the fish are at rest.
Alipasha Vaziri’s team at the University of Vienna has developed a light-field microscopy technique for imaging neurons at lower spatial resolutions but higher speed than the new light-sheet microscopy technique. They have scanned the whole brains of larval zebrafish and the entire nervous system of the nematode C. elegans and are set to use Thunder to analyze their data.
“I’m happy to see that someone has done this,” says Vaziri, a group leader at the university’s Max F. Perutz Laboratories. “All the elements of the library have been available individually, but not in such a coherent platform.”
Cunningham agrees Thunder is a timely and well-conceived development. “They’ve gone to great lengths to make Thunder as usable as possible. In terms of distributed computing platforms, it’s quite elegant and quite usable,” he says.
Not all neuroscientists will have the background in programming language to implement Thunder right away. “It requires a computationally savvy scientist to use it,” he says. But “it’s an idea whose time has come.”
1: Freeman J. et al. Nat. Methods Epub ahead of print (2014) PubMed
2: Vladimirov N. et al. Nat. Methods Epub ahead of print (2014) PubMed