A machine learning pipeline for genome-scale ranking of metabolic interventions, demonstrated on aerobic succinate production in Escherichia coli (iJO1366).
Author: Daxlia
Repository: github.com/Daxlia/RéseauFlux
License: Modified MIT with Citation Requirement (seeLICENSE)
RéseauFlux integrates constraint-based metabolic modelling (COBRApy) with gradient boosting regression (GBR) to rank all gene knockouts, overexpressions, and downregulations across a genome-scale model — without running FBA for every candidate. The pipeline:
- Extracts biologically meaningful features per reaction (FVA range, betweenness centrality, shortest paths to target and biomass, subsystem one-hot encoding, flux entropy, etc.)
- Trains a GBR on rank-normalized succinate flux labels from ~300 enumerated interventions
- Predicts ranks for the remaining ~1,950 unenumerated reactions
- Validates top candidates with full two-step FBA (maximize growth → fix growth → maximize target)
- Extends to double interventions, evolutionary search, Monte Carlo robustness, and multi-target generalization
Key result: ATPS4rpp KO is ranked #1 by the model (16.38 mmol/gDW/h succinate, +100× WT), consistent with experimental literature.
pip install cobra highspy scikit-learn networkx numpy pandas matplotlib scipy
pip install straindesign # optional — enables OptKnock/StrainDesign comparisonPython ≥ 3.9. The iJO1366 model is downloaded automatically via cobra.io.load_model("iJO1366").
Tested with:
- COBRApy 0.29+
- scikit-learn 1.4+
- HiGHS solver (via
highspy)
python metabolic_ml_pipeline.pyAll outputs (CSV tables + summary plot) are written to the current directory and packaged into a timestamped ZIP archive at the end of the run.
| Parameter | Default | Description |
|---|---|---|
TARGET_REACTION |
"EX_succ_e" |
Exchange reaction to maximize |
MAX_ENUM_REACTIONS |
300 |
Interventions to FBA-enumerate for training |
GROWTH_FRACTION |
0.10 |
Minimum growth fraction in two-step optimizer |
N_TOP |
10 |
Top singles validated with FBA |
N_VALIDATE_DOUBLES |
25 |
Top ML-predicted doubles validated with FBA |
SEED |
42 |
Random seed |
| File | Contents |
|---|---|
ranking_all_interventions.csv |
Full ranked list of all (reaction, kind) pairs |
top_validated_strategies.csv |
Top singles with FBA-validated flux and growth |
combo_search.csv |
Top ML-predicted double interventions |
evolutionary_search.csv |
Triple KO candidates from evolutionary search |
monte_carlo_robustness.csv |
Robustness scores under ±20% parameter noise |
flux_control_coefficients.csv |
FCC analysis (production bottleneck identification) |
gene_knockouts.csv |
Gene-level knockout predictions |
ablation_study.csv |
GBR vs pairwise ranker vs ensemble comparison |
generalization_holdout.csv |
Holdout test results (70/30 reaction-group split) |
literature_comparison.csv |
ML predictions vs known experimental results |
multi_target_summary.csv |
Pipeline performance on L-malate and acetate targets |
pareto_front.csv |
Pareto-optimal strategies (flux vs growth) |
results_summary.png |
12-panel summary figure |
| Metric | Value |
|---|---|
| GBR cross-validation Spearman ρ | 0.557 |
| Holdout Spearman ρ (unseen reaction groups) | 0.602 |
| Top-K precision (CV) | 0.472 |
| Best single intervention | ATPS4rpp KO → 16.38 mmol/gDW/h |
| Best double intervention (ML-guided) | ATPM KO + O2tex KO → 16.80 mmol/gDW/h |
| Literature rank recall (top 10) | 2/3 known KOs ranked in top 10 |
| Multi-target (L-malate) ρ | 0.445 |
| Multi-target (acetate) ρ | 0.491 |
iJO1366 model
│
├─ [1] FVA + network analysis → feature matrix (18 features per reaction)
├─ [2] Two-step FBA enumeration → rank-normalized labels (y_rank)
├─ [3] GBR training (5-fold GroupKFold CV, grouped by reaction)
├─ [4] Rank prediction for all 2,583 reactions × 3 intervention types
├─ [5] FBA validation of top-10 singles
├─ [6] ML double-intervention prediction + FBA validation of top-25
├─ [7] Monte Carlo robustness, FCC, evolutionary search
└─ [Pub] Holdout generalization, ablation, literature comparison, multi-target
Ablation study results:
| Model | CV Spearman ρ |
|---|---|
| GBR-only | 0.557 |
| Ensemble (GBR + pairwise) | 0.358 |
| Pairwise-only | 0.159 |
The pairwise ranker was removed from the final pipeline after ablation showed it degrades performance.
If you use this code or results, please cite:
Daxlia. RéseauFlux: Machine Learning-Guided Ranking of Metabolic Interventions for Succinate Overproduction in Escherichia coli. Zenodo (2026). DOI: 10.5281/zenodo.19984811
This project was developed and executed on Kaggle Notebooks using two NVIDIA Tesla T4 GPUs provided free of charge. The authors thank Kaggle for the free compute resources.
Modified MIT with Citation Requirement. See LICENSE.