Dissemination of a tool for data-driven multiscale modeling of brain circuits

点击翻译按钮获取中文摘要

Project Summary Experiments aimed at discovering how the brain works generate vast amounts of data that span multiple scales: from interactions between individual molecules to waves of electrical activity across the entire brain. Computational modeling provides a way to integrate and make sense of these data. Through the parent grant U24EB028998 we are developing and disseminating NetPyNE, a tool for data-driven multiscale modeling of brain circuits. This tool provides a programmatic and graphical high-level interface to the widely-used NEURON simulator that facilitates the development, parallel simulation, optimization and analysis of biophysically detailed neuronal circuits. NetPyNE uses CoreNEURON, an improved simulation engine optimized for parallel simulation on both CPUs and GPUs. Significant progress has been made towards achieving the parent grant goal of transforming NetPyNE into a solid and well-tested tool with a fully-featured GUI, and widely disseminating the tool among the scientific community. This is supported by a growing user base, as evidenced by over 100 models being developed across more than 40 institutions worldwide, over 30 peer-reviewed publications making use of the tool. NetPyNE has also been integrated or interfaced with multiple community standards, tools and platforms, including the NeuroML and SONATA, the Open Source Brain, EBRAINS and The Neuroscience Gateway (NSG), HNN, SciUnit/SciDash, LFPy, and The Virtual Brain. This supplement proposal aims to explore and evaluate the use of cloud-based GPU resources to accelerate large-scale biophysically-detailed simulations of brain circuits using NetPyNE and the CoreNERON simulation engine. CoreNEURON focuses on improving performance by modernizing the legacy NEURON simulation engine to be optimized for parallel computation on modern architectures, including cloud GPUs. The yield of offloading these computationally intensive tasks from CPUs to GPUs has been demonstrated on several in-silico models with speedups of up to 40x. Currently, performance increases have only been implemented and evaluated for a handful of models. To facilitate the adoption of GPU utilization for large-scale modeling of brain circuits, we will evaluate the recently published NetPyNE-based somatosensory (S1) and auditory (A1) thalamocortical large-scale models on cloud GPU resources. We will first evaluate individual simulations on a single GPU node (Aim 1). Next, we will evaluate, for the first time, the use of clusters of GPU nodes to perform large parameter optimizations by running many large-scale simulations simultaneously (Aim 2). We will apply rigorous benchmarking measures, including computation time and memory usage, to evaluate the feasibility of this approach and characterize its benefits across use cases, including models of different sizes and different cloud configurations. This supplement will enhance the performance, interoperability and community adoption of NetPyNE, accelerate multiple NIH-funded research projects that use NetPyNE, and make cloud GPU technologies more accessible to under-resourced institutions and communities.