Physics-informed machine learning for predicting intracellular metabolic fluxes from extracellular metabolite information

code/ common.py shared code: data loading, models, training, hyperparameter search, QP correction, metrics 01_hp_search_pinn.py hyperparameter search (nested leave-one-out, Optuna) for the PINN 02_hp_search_nsp.py hyperparameter search for the NSP-MLP 03_validate_pinn.py leave-one-out validation of the PINN (50 seeds, QP correction) 04_validate_nsp.py leave-one-out validation of the NSP-MLP 05_benchmark_fba_pfba.py FBA and pFBA with iCW773 06_benchmark_ecfba.py ecFBA with ecCGL1 07_collect_metrics.py per-sample overall and pathway-level metrics of the five models export_sbml.py writes the condensed CCM network as SBML and as a reaction tabledata/ raw_data.csv ¹³C-MFA fluxes of the 45 datasets (normalized, see below) Stoichiometric_Matrix_v87.csv stoichiometric matrix (58 metabolites x 87 reactions) dataset_metadata.csv source study, production target and gene deletions of each dataset reaction_list.csv the 87 reactions: stoichiometry, bounds, pathwaymodel/ CCM_Cglutamicum.xml condensed CCM network (SBML level 3, FBC 2)hyperparameters/ PINN_sample_hyperparams.csv hyperparameters selected in each of the 45 outer folds (used in the paper) NSP-MLP_sample_hyperparams.csvresults/ pinn/, nsp_mlp/ predicted fluxes (mean and SD over 50 seeds, before and after QP) and per-sample metrics fba/, pfba/, ecfba/ predicted fluxes and per-sample metrics of the benchmark models metrics/ metrics of the five models: RMSE.csv, WMAPE.csv, R2.csv, PCC.csv (V3-V27) and RMSE_/WMAPE_ files for glycolysis, PPP and TCA (inputs of the tables and figures)