Brains and computers have a lot in common. Both systems produce great complexity and remarkable abilities from relatively simple building blocks, whether it’s the neuron or the transistor. Exploiting this similarity, scientists have long sought to create artificial intelligence that mimics the brain’s function, and machine learning experts drew inspiration from synaptic communication to create their own mathematical neural networks for solving difficult problems. Currently in Europe and the US, massive collaborations seek to simulate the entire human brain with computers to gain insight into disease, behavior, and consciousness.
But in the laboratory of CI senior fellow Wim van Drongelen, researchers are using computer models to evaluate what happens when the brain “crashes” in a seizure. With a combination of animal studies, electrical recordings from human brain tissue, and computer models involving millions of connected, artificial neurons, the team searches for a deeper understanding of what tips normal brain activity into an epileptic seizure. The pursuit has led them to start developing a new code for simulating the brain’s intricate network dynamics, called Kunstkop -- Dutch for “artificial head.”
“Although the activity of single neurons is important, our goal is to focus more on the aggregate behavior of large populations of thousands and millions of cells,” said van Drongelen, Professor of Pediatrics at the University of Chicago and Technical & Research Director of the Pediatric Epilepsy Center. “It’s important to study what the populations are doing and how they interact to create seizure activity.”
The nature of epilepsy requires the Kunstkop developers, led by the CI’s Lorenzo Pesce, to think big. While epilepsy may originate in genetic mutations or deficient proteins at the single cell level, a seizure is a crowd phenomenon, with millions -- if not billions -- of connected neurons synchronizing into an aberrant pattern. In order to study these large-scale, network dynamics and match them up with clinical seizure data recorded using electroencephalograms (EEGs) scientists will need to construct equally massive and complex models.
“Since epileptic seizures are coordinated or compound activities of really large networks, we need really large computational models,” said van Drongelen. “In the past, we have not gone much beyond hundreds of thousands [of neurons], and I think we reached one million once. What we want to be able to do with our current approach is go up to ten million [neurons], so that we can mimic a piece of brain that is similar to a piece that generates a couple of EEG signals. That’s the current goal.”
A simulation of a 100,000-neuron network (courtesy van Drongelen lab)
In order to simulate the network activity of millions of computerized neurons, scientists must use parallel computing, where thousands of CPUs run calculations simultaneously instead of sequentially. But it’s not a given that a simulation of a single cell or a small network will work when scaled up to millions of neurons and thousands of processors. So in a recent paper for Computational and Mathematical Methods in Medicine, Pesce, van Drongelen, and other CI researchers including Mark Hereld and Rick Stevens tested the performance of two single-cell simulators when they are multiplied hundreds of thousands times over.
In the paper, Pesce and colleagues examined the performance of two programs, pNeo and Verdandi (both downloadable here), when run on two powerful parallel supercomputers: the CI’s Beagle, and Argonne’s Fusion. Both simulators showed linear scaling up to networks as big as 400,000 neurons and 1.5 billion synapses, indicating that it is feasible to run simulations of this size or larger on supercomputers, projecting a run time of 4 to 25 hours for networks of one to ten million neurons.
But while pNeo and Verdandi performed well in these scaling trials, they were not originally designed to be easily modifiable for exploring hypotheses about network activity and epilepsy. Researchers in the van Drongelen laboratory would like to test different cell types and different connectivity structures to see what components of the system contribute to seizure generation. So Kunstkop will be built with object-oriented programming, allowing users to make modular changes to the simulation, and will be optimized for parallel computing to make simulations of large networks feasible on today’s supercomputers.
“We need to have a lot of flexibility,” Pesce said, “and we need to have lots of performance, because if you need to run ten million neurons, with complex neurons in a complex network, our estimate is that we need to run it on hundreds or thousands of cores, at least.”
For the van Drongelen group, the Kunstkop model will close an experimental loop that also contains studies of neurons from animal and human brains as well as neuronal networks grown in cell culture. Recording the population activity in these laboratory setups can help fine-tune the performance of the model, and hypotheses tested out in the computer simulations can inform experimental manipulations in the laboratory. As the cycle turns, it will generate new insight on the origin of epilepsy and, hopefully, new treatments for patients with the condition.
“We don’t really know which part is necessary, how many types of neurons we need, how many of those neurons, or what kind of connectivity,” Pesce said. “The idea is to create a model that allows to bridge the various scales and allow to generate hypotheses that are hopefully helpful in understanding and perhaps curing some kinds of epilepsy.”