Online Fitting Connectome-constrained Drosophila Whole-brain Model Reproduces Critical Resting-state Dynamics

Qianzhu Li, Minghao Wang, Xinzhu Liao, KeZhang, Chaoming Wang

Keywords: whole-brain model; connectome; online learning; resting-state neural activity; critical dynamical state

Abstract: The rapid growth of large-scale synaptic connectome maps and neural activity datasets has created an urgent need for connectome-constrained whole-brain models that can fit and interpret experimentally recorded neural data. A promising approach to bridge this gap is to train biologically inspired models using backpropagation through time (BPTT), which enables data-driven optimization of unknown model parameters. However, BPTT is inherently an offline method, with memory requirements that grow linearly with simulation time, making it impractical for training large-scale whole-brain networks over biologically relevant timescales. To address this challenge, we introduce an online learning framework for fitting whole-brain models using online gradient-based optimization. By updating parameters in a strictly forward-time manner, our method reduces memory consumption to a single time step, scaling only with the number of parameters rather than the entire temporal sequence. Using this framework, we construct a Drosophila whole-brain network comprising over 130,000 neurons and millions of synapses, where the network topology is fixed from the FlyWire connectome, and unknown parameters such as synaptic weights and cellular time constants are optimized to match in vivo resting-state neural activity. Our results show that this approach enables the training of large-scale Drosophila models over experimental timescales on a single GPU, a feat that is computationally prohibitive with BPTT. Remarkably, the optimization not only captures target dynamics but also spontaneously produces synaptic weight distributions that closely match empirical connectome statistics and drives the network toward a hallmark feature of resting state -- critical dynamics. Together, this work establishes an online, scalable, and data-driven framework for integrating anatomical and functional datasets, paving the way toward mechanistic whole-brain models at unprecedented scales.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 24727