Associated Publication: Young, J., Gu, L. & Zhou, R. "Predicting FOX gene candidates for oxic nitrogen fixation using multi-omic machine learning and comparative bioinformatics." Scientific Reports (submitted).

Setup

git clone https://github.com/jamesyoung93/FoxGenes_ML.git
cd FoxGenes_ML

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Repository Structure

FoxGenes_ML/
├── src/                           # Core module
│   ├── data_prep.py               # Data loading
│   ├── feature_sets.py            # Feature set definitions
│   ├── nested_cv.py               # Nested CV (not used for Table 1)
│   ├── bootstrap_holdout.py       # Repeated holdout evaluation (Table 1)
│   └── blocked_area_cv.py         # Location-blocked CV (Table S6)
├── data/
│   └── prepared_for_modeling.plus_conservation_v2.csv
├── outputs/                       # Manuscript results
│   ├── bootstrap_ablation/        # Table 1 & S5 metrics
│   ├── blocked_cv/                # Table S6 metrics
│   └── final_predictions/         # Trained models & pFOX scores
├── run_bootstrap_ablation_suite.py
├── run_blocked_area_cv.py
├── train_final_and_predict_unknowns.py
├── make_fig5_6panel.py
└── requirements.txt

Reproducing Results

Table 1: Model Performance (repeated stratified holdout)

Table 1 reports metrics from 20 repeated stratified 80/20 train-test splits using the principled_expression feature set (includes genomic position).

python run_bootstrap_ablation_suite.py \
  --data data/prepared_for_modeling.plus_conservation_v2.csv \
  --out outputs/table1_check \
  --train_fracs 0.8 \
  --n_splits 20

Pre-computed results are in outputs/bootstrap_ablation/principled_expression/summary.csv.

Table S6: Location-Blocked CV

Tests generalization across genomic regions (5 blocks × 4 partition offsets = 20 splits).

python run_blocked_area_cv.py \
  --data data/prepared_for_modeling.plus_conservation_v2.csv \
  --feature_set principled_expression \
  --n_groups 5 \
  --n_partitions 4 \
  --out outputs/blocked_cv_check

Final Models & Predictions

After evaluation, models were retrained on all 903 labeled genes with 5-fold CV hyperparameter tuning (60 iterations).

python train_final_and_predict_unknowns.py \
  --data data/prepared_for_modeling.plus_conservation_v2.csv \
  --feature_set principled_expression \
  --out outputs/final_check \
  --n_iter 60 \
  --cv_folds 5

Final trained models and genome-wide pFOX scores are in outputs/final_predictions/.

Figure 5: SHAP Analysis

python make_fig5_6panel.py \
  --model_joblib outputs/final_predictions/final_model_xgb.joblib \
  --feature_matrix data/prepared_for_modeling.plus_conservation_v2.csv \
  --predictions_all outputs/final_predictions/predictions_all_ensemble.csv \
  --outdir outputs/fig5

Feature Sets

Name	Description
principled_expression	RPKM + proteomics + promoter + conservation + position (main model)
principled_expression_no_position	Same without genomic coordinates (ablation)
baseline_full	All available features

Related Repositories

• FoxGenesApp — Interactive web tool for exploring predictions
• cyanobacteria-diazotrophic-proteome — RBH comparative proteomics pipeline

License

MIT