Modeling Neurons with AI Innovations
Over the years, scientists have identified hundreds of different types of neurons that form the neural circuits in our brains, revealing their electrical properties and the early genetic indicators that determine their eventual types. However, the challenge of synthesizing this vast array of data into a cohesive model that accurately reflects how neurons facilitate brain processing persists.
A breakthrough has been made by a team of researchers led by Caltech and Cedars-Sinai, who have developed an innovative artificial intelligence framework named NOBLE (Neural Operator with Biologically-informed Latent Embeddings). This tool promises to accelerate discoveries in brain function research and ultimately improve treatments for brain disorders by efficiently crafting virtual models of brain neurons.
The introduction of NOBLE was made at the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS) in San Diego. Anima Anandkumar, Bren Professor of Computing and Mathematical Sciences at Caltech, and a lead author of the study remarked, “Caltech, being the birthplace of NeurIPS, aims to bridge neuroscience and AI, and NOBLE exemplifies this ambition. It is the first large-scale deep-learning framework that merges mathematical models of bio-realistic neurons with experimental validation.” This statement underscores the significance of uniting these two cutting-edge fields.
At the core of NOBLE are neural operators, which Anandkumar specializes in. Unlike traditional neural networks that utilize discrete data points, neural operators function with continuous mathematical functions. This distinction allows for the examination of system factors at various scales and resolutions, presenting a comprehensive overview of data behavior and significantly enhancing the speed and efficiency compared to traditional methods.
As Costas Anastassiou, Associate Professor of Neurology, Neurosurgery, and Biomedical Sciences at Cedars-Sinai states, “Computational modeling of brain neurons is crucial for studying their activities and interactions. However, existing models face limitations such as high computational costs, data accessibility, and cumbersome management issues. Our innovative framework overcomes these challenges by operating thousands of times faster than current models while capturing the biological accuracies necessary to reflect the true variability of brain neurons. This advancement enables the generation of an unlimited variety of virtual neurons, portraying the rich diversity found in actual biological neurons.”
The research paper titled “NOBLE—Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models,” is featured in the NeurIPS proceedings. The research was conducted by a team that includes Valentin Duruisseaux from Caltech, Luca Ghafourpour of ETH Zurich, Bahareh Tolooshams from the University of Alberta, and Philip H. Wong from Cedars-Sinai Medical Center. The study received funding from 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.