Hundreds of different types of neurons make up the neural circuits in our brains. Over the years, scientists have discovered many details about those different cell types, including their electrical properties and the early genetic indicators that dictate what type of neuron they will eventually become. Yet combining all of that varied information into a clear model that accurately reflects how neurons produce brain processing remains an outstanding challenge.
Now a team of scientists led by Caltech and Cedars-Sinai has developed a new artificial intelligence framework that can accurately, quickly, and efficiently create virtual models of brain neurons. The tool, dubbed NOBLE (Neural Operator with Biologically-informed Latent Embeddings), could accelerate discoveries in brain function research and ultimately lead to better treatments for brain disorders.
The researchers presented NOBLE at the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS) in San Diego.
"Not many are aware that Caltech is the birthplace of NeurIPS, and its goal was to bring neuroscience and AI under one umbrella," says Anima Anandkumar, Bren Professor of Computing and Mathematical Sciences at Caltech, an affiliated faculty member of the Tianqiao and Chrissy Chen Institute for Neuroscience at Caltech, and an author of the new work. "That's precisely what NOBLE accomplishes. It is the first large-scale deep-learning framework that combines mathematical models of bio-realistic neurons with experimental validation."
NOBLE is based on neural operators, which are Anandkumar's specialty. Neural operators are a special type of neural network that works with continuous mathematical functions rather than discrete data points. They allow researchers to examine factors in a system at varying scales and resolutions to provide a wider view of how the data are behaving much faster than standard methods, offering key benefits to scientific problems.
"Computational modeling of brain neurons has become an important tool for studying their activity and interactions," says Costas Anastassiou, an associate professor of neurology, neurosurgery and biomedical sciences at Cedars-Sinai and an author of the new paper. "But traditional models are hindered by limitations including the cost of computer resources, data availability, and cumbersome handling. Our new framework tackles this problem by operating at speeds thousands of times faster than existing methods while remaining so biologically accurate that it can capture the variability of actual brain neurons. The framework can generate an unlimited number of virtual neurons, better reflecting the diversity and variability of actual biological neurons."
The paper, "NOBLE—Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models," appears in the NeurIPS proceedings. Additional authors are Valentin Duruisseaux, a postdoctoral scholar research associate at Caltech; Luca Ghafourpour of ETH Zurich; Bahareh Tolooshams of the University of Alberta and Alberta Machine Intelligence Institute; and Philip H. Wong of Cedars-Sinai Medical Center. The researchers are funded by the Bren Endowed Chair at Caltech, the Office of Naval Research, the AI2050 Senior Fellow program at Schmidt Sciences, and the National Institutes of Health.
