Isogeometric analysis-based physics-informed graph neural network for studying traffic jam in neurons
Isogeometric analysis-based physics-informed graph neural network for studying traffic jam in neurons
复制标题
DOI：
10.1016/j.cma.2022.115757
复制
发表时间：
2023-01
影响因子：
7.2
通讯作者：
Angran Li;Y. Zhang
中科院分区：
文献类型：
--
作者：
Angran Li;Y. Zhang
马达驱动的细胞内转运在支持神经元细胞的生存和功能中起着至关重要的作用，马达蛋白和微管（MT）结构协同作用，迅速将必需物质运送到神经元的正确位置。运输的中断可能导致各种神经退行性疾病的发生。因此，研究神经元如何调节物质运输过程，更好地理解交通堵塞的形成是十分必要的。在我们早期的工作中，我们开发了一个pde约束优化模型和一个等几何分析（IGA）求解器来模拟MT减少和MT漩涡引起的交通拥堵。在这里，我们开发了一种新的基于iga的物理信息图神经网络（PGNN）来快速预测不同神经元几何形状的正常和异常运输现象，如交通堵塞。特别是，该方法借鉴了细胞内运输过程的IGA模拟，对正常运输和mt诱导的交通堵塞提供了准确的物质浓度预测。基于iga的PGNN模型包含模拟器，用于处理管道中正常和两种mt引起的交通堵塞的局部预测，以及另一个模拟器用于预测分叉中的正常运输。采用bsamizier提取方法将几何信息融入模拟器中，精确计算出带有PDE残差的物理信息损失函数。此外，采用GNN装配模型将局部预测装配到整个几何结构中，以处理不同的神经元形态。综上所述，与IGA模拟相比，训练良好的模型能够有效地预测交通阻塞和正常运输过程中运输速度和物质浓度的分布，平均误差小于10%。在几个复杂的神经元几何中证明了该模型的有效性。
The motor-driven intracellular transport plays a crucial role in supporting a neuron cell’s survival and function, with motor proteins and microtubule (MT) structures collaborating to promptly deliver the essential materials to the right location in neuron. The disruption of transport may lead to the onset of various neurodegenerative diseases. Therefore, it is essential to study how neurons regulate the material transport process and have a better understanding of the traffic jam formation. In our earlier work, we developed a PDE-constrained optimization model and an isogeometric analysis (IGA) solver to simulate traffic jams induced by MT reduction and MT swirl. Here, we develop a novel IGA-based physics-informed graph neural network (PGNN) to quickly predict normal and abnormal transport phenomena such as traffic jam in different neuron geometries. In particular, the proposed method learns from the IGA simulation of the intracellular transport process and provides accurate material concentration prediction of normal transport and MT-induced traffic jam. The IGA-based PGNN model contains simulators to handle local prediction of both normal and two MT-induced traffic jams in pipes, as well as another simulator to predict normal transport in bifurcations. Bézier extraction is adopted to incorporate the geometry information into the simulators to accurately compute the physics informed loss function with PDE residuals. Moreover, a GNN assembly model is adopted to tackle different neuron morphologies by assembling local prediction into the entire geometry. In summary, the well-trained model effectively predicts the distribution of transport velocity and material concentration during traffic jam and normal transport with an average error less than 10% compared to IGA simulations. The effectiveness of the proposed model is demonstrated within several complex neuron geometries.