Abstract
Neurons exhibit complex geometry in their branched networks of neurites which is essential to the function of individual neuron but also brings challenges to transport a wide variety of essential materials throughout their neurite networks for their survival and function. While numerical methods like isogeometric analysis (IGA) have been used for modeling the material transport process via solving partial differential equations (PDEs), they require long computation time and huge computation resources to ensure accurate geometry representation and solution, thus limit their biomedical application. Here we present a graph neural network (GNN)-based deep learning model to learn the IGA-based material transport simulation and provide fast material concentration prediction within neurite networks of any topology. Given input boundary conditions and geometry configurations, the well-trained model can predict the dynamical concentration change during the transport process with an average error less than 10% and 120–330 times faster compared to IGA simulations. The effectiveness of the proposed model is demonstrated within several complex neurite networks.
Introduction
Neurons are the long-haul truckers of the biological world, relying on intricate, branched networks called neurites to transport essential materials—such as proteins and organelles—from the cell body to distant synapses. Disruptions in this transport process are linked to neurodegenerative diseases like Alzheimer’s and Parkinson’s. To study these processes, researchers typically use Isogeometric Analysis (IGA) to solve partial differential equations (PDEs) describing reaction-diffusion-transport. However, IGA is computationally expensive; simulating a single complex 3D neuron can take hours or days on a supercomputer, creating a bottleneck for real-time biomedical applications.
Methods: A Modular GNN Approach
To overcome the computational limitations of traditional solvers, a modular Graph Neural Network (GNN) framework was developed. This approach treats the complex neuron structure as an assembly of fundamental building blocks.
Geometric Decomposition and Simulation
The framework decomposes any complex neurite network into two primary components: pipes (straight segments) and bifurcations (junctions). Two specialized GNN simulators, $G_S^p$ for pipes and $G_S^b$ for bifurcations, are trained to predict local transport dynamics.
Physics-Informed Learning
A key innovation is the use of physics-informed loss functions. Rather than relying solely on Mean Squared Error (MSE), the training process incorporates PDE residuals. This ensures the neural network respects the underlying physical laws of diffusion and transport, significantly improving accuracy. For instance, adding PDE residuals reduced the error in the pipe simulator from 10.9% to 6.1%.
GNN Assembly Model ($G_A$)
To reconstruct the full neuron, an assembly model uses a message-passing scheme to stitch local predictions together. A penalty term in the loss function ensures continuity in material concentration at the interfaces between pipes and bifurcations.
Key Results
The GNN-based model demonstrates remarkable performance across several metrics, offering a high-fidelity surrogate for traditional numerical methods.
Speed and Accuracy
The model achieves a massive speedup, running 120 to 330 times faster than the IGA solver. Simulations that previously required hours on a CPU cluster can now be completed in seconds on a GPU. Despite this speed, the Mean Relative Error (MRE) remains consistently under 10% across various complex networks.
Generalization and Transfer Learning
The model's modular nature makes it topology-agnostic, allowing it to predict transport in networks of arbitrary complexity. While trained primarily on zebrafish neurons from the NeuroMorpho.Org database, the model generalized well to mouse neurons. Using transfer learning—fine-tuning the pre-trained model with a small amount of species-specific data—the researchers achieved an average MRE of approximately 8.23% on mouse morphologies.
Conclusion and Future Work
This GNN framework represents a significant step toward real-time simulation of neuronal health. It successfully identifies regions of high material accumulation, which are potential sites for transport disruptions. Future work will focus on including physical constants (like diffusion coefficients) as input features and optimizing the model for extremely high bifurcation densities.
Summary updated