NeuronTransportGALDS

Overview

This repository contains the official code for the paper "GALDS: A Graph-Autoencoder-based Latent Dynamics Surrogate model to predict neurite material transport". It includes all necessary scripts for data pre-processing, model training, prediction, and post-processing to replicate our results.

Workflow & Usage

The workflow is divided into four main stages: data generation, preprocessing, model training, and post-processing.

1. Data generation

To simulate the baseline neurite transport, we follow the methodology from [1] using the code available at the NeuronTransportIGA repository.

For experiments involving a spatially asymmetric attachment rate, we have modified the original simulation code. These modifications are provided in the /IGA directory. Please consult the updated manual.txt within that folder for detailed instructions on running these custom simulations.

2. Data preprocessing

After generating the simulation data (VTK files), the next step is to preprocess it for the model. This involves extracting nodal data from the neurite cross-sections (using 17 and 24 nodes, as described in the paper) and consolidating them into compact .h5 files.

The necessary scripts are located in the /preprocess folder.

To run preprocessing:

python3 tree_extraction.py --case <case_name> --thread 20 --transport 1

Key Parameters:

--case: Name of the simulation case to process.
--thread: Number of parallel threads to use for faster processing.
--ns: (Optional) 1 to extract Navier-Stokes results.
--transport: 1 to extract concentration results (required for GALDS). --k: 1 to extract k values
--remove_old: 1 to clean the target directory before generation.
--overwirte: 1 to overwrite existing files.\

Note: Proper structuring of the raw simulation files is crucial. Please refer to the README.md inside the /preprocess directory for the required file hierarchy.

3. Model training

The training scripts for all components of the GALDS architecture are located in the /training directory. This includes individual scripts for the:\

1. Autoencoder (Sec 4.1)
2. Latent Space Transformation Module (Sec 4.2)
3. Latent Space System Dynamics Module (Sec 4.3)

To train all modules sequentially, simply execute the main shell script.

To start training:

sh train_all.sh

4. Result Post-processing

The trained GALDS model outputs predictions in a point cloud format. To facilitate analysis and visualization, we project these points back onto the original neurite mesh using linear interpolation.

The scripts for this final step are available in the /post_processing directory.

To run post-processing:

sh post_process.sh

Python Libraries & Dependencies

This project relies on several key Python libraries. The core dependencies are listed in requirements.txt.

[1] A. Li, X. Chai, G. Yang, Y. J. Zhang. An Isogeometric Analysis Computational Platform for Material Transport Simulation in Complex Neurite Networks. Molecular & Cellular Biomechanics, 16(2):123-140, 2019.