[Paper Review] Digital Twin Brain: a simulation and assimilation platform for whole human brain

Wenlian Lu, Longbin Zeng+13 more|arXiv (Cornell University)|Aug 2, 2023

Molecular Communication and Nanonetworks13 citations

TL;DR

DTB is a GPU-based platform that simulates whole human-brain-scale spiking networks (86B neurons, 47.8T synapses) using sMRI-DTI-PET personalized structure, achieving real-time-ish wall times for various firing rates and enabling data assimilation to align with BOLD signals.

ABSTRACT

In this work, we present a computing platform named digital twin brain (DTB) that can simulate spiking neuronal networks of the whole human brain scale and more importantly, a personalized biological brain structure. In comparison to most brain simulations with a homogeneous global structure, we highlight that the sparseness, couplingness and heterogeneity in the sMRI, DTI and PET data of the brain has an essential impact on the efficiency of brain simulation, which is proved from the scaling experiments that the DTB of human brain simulation is communication-intensive and memory-access intensive computing systems rather than computation-intensive. We utilize a number of optimization techniques to balance and integrate the computation loads and communication traffics from the heterogeneous biological structure to the general GPU-based HPC and achieve leading simulation performance for the whole human brain-scaled spiking neuronal networks. On the other hand, the biological structure, equipped with a mesoscopic data assimilation, enables the DTB to investigate brain cognitive function by a reverse-engineering method, which is demonstrated by a digital experiment of visual evaluation on the DTB. Furthermore, we believe that the developing DTB will be a promising powerful platform for a large of research orients including brain-inspiredintelligence, rain disease medicine and brain-machine interface.

Motivation & Objective

• Demonstrate a whole-brain spiking neuronal network model based on personalized sMRI-DTI-PET data.
• Show that memory access and communication, not computation, bottleneck brain-scale simulation on HPC/GPU systems.
• Develop partitioning and data-structuring techniques to balance load and minimize inter-GPU traffic.
• Propose a mesoscopic data assimilation framework to infer parameters from BOLD signals and enable cognitive-like dynamics.

Proposed method

• Model the brain as a hierarchical random graph with voxel-level LIF neurons and four synapse types (AMPA, NMDA, GABAa, GABAb).
• Use sMRI-DTI-PET data to define sparse, heterogeneous connectivity across cortex, sub-cortex, brainstem, and cerebellum.
• Implement a GPU-based HPC simulation with three-thread parallelism handling computation, intra- and inter-GPU communication.
• Introduce an N-way partitioning algorithm to minimize inter-GPU traffic while respecting GPU capacity.
• Develop hierarchical mesoscopic data assimilation to estimate hyper-parameters from BOLD signals via the Balloon-Windkessel model.

Experimental results

Research questions

• RQ1 Can a whole-human-brain-scale spiking network be simulated efficiently on GPU-based HPC given the brain's sparse and heterogeneous structure?
• RQ2 What partitioning and data structures optimize inter-GPU communication for such a large-scale, voxel-based network?
• RQ3 Can a mesoscopic data assimilation framework infer biologically meaningful parameters to reproduce cognitive-like dynamics from BOLD data?
• RQ4 How does personalized anatomy (sMRI-DTI-PET) affect simulation performance and functional inference compared to homogeneous networks.

Key findings

• The platform simulates up to 86 billion neurons and 47.8 trillion synapses.
• Time-to-solution for 1 second of biological time is 65 seconds at ~7 Hz, 78.8 seconds at ~15 Hz, and 118.8 seconds at ~30 Hz firing rates.
• The simulation is memory-access- and communication-intensive rather than computation-intensive on HPC systems with heterogeneous brain structure.
• A novel partitioning algorithm reduces inter-GPU traffic variability and congestion compared to sequential mapping.
• A Pearson correlation over 0.65 is observed between DTB-derived and biological counterpart in visual evaluation tasks via HMDA data assimilation.