Overview of the results described in Neural Ordinary Differential Equations Inspired Parameterization of Kinetic Models
We used a HPC cluster to run all experiments.
Overview of scripts needed to get results reported in paper results
Initialization bounds latin hypercube sampling (Figure 2) in EXP4_Glycolysis_Fitting_Datasets
If you want to reproduce the results from the 25 SBML models which are reported in EXP1_initialization_bounds, you need to run 2207_trainer_script.py. This can be run through the command line. For example:
python3 scripts/simulated_data_trainer_script.py -n Becker_Science2010.xml -t_end 15 -s 100 -i 1 -b 10 -o "results/EXP1_initialization_bounds_lhs/Becker_Science2010"
Glycolysis model fitting EXP4_Glycolysis_Fitting_Datasets
For the model shown in figure 4, run the following command for 8000 iterations of gradient descent. This should take roughly 4 hours.
python3 scripts/simulated_data_trainer_script.py -n 8000 -d "datasets/VanHeerden_Glucose_Pulse/FF1_timeseries_format.csv"
Fitting multiple datasets as described in the Supporting Information can be performed using the 1709_train_gp_twodatasets.py and 1709_train_gp_all_datasets.py scripts in a similar fashion.
Figure 2A,B,C: Fig2ABC_parameterization_analysis_SBML.ipynb Figure 2D,E,F) Fig2DEF_lossplots_and_timeseries_examples.ipynb
Figure 3A,B,C,D) Figure3ABCD_parameter_distance_to_optimum.ipynb
Figure 4) Figure4_glucpulse_fit_assessment.ipynb
Figure SI4: Figure4_glucpulse_fit_assessment.ipynb