BULMA: A Stress-Responsive Atlas of ABC Transporters in S. cerevisiae

Harrizi S., Nait Irahal I., Kabine M.
Laboratoire Santé, Environnement et Biotechnologie (LSEB), Université Hassan II de Casablanca

---

Overview

BULMA (Biologically-informed Unified Learning for Multi-stress Atlas) is a reproducible computational framework that integrates:

1. Predictive Atlas — Two-tower MLP predicting ABC transporter–compound interactions using ESM-2 protein embeddings + ChemBERTa ligand embeddings
2. Causal Ranking — Doubly-robust DR-Learner (EconML) estimating causal treatment effects per transporter across stress conditions
3. Active Learning — Uncertainty / Diversity / Hybrid strategies to maximize discovery efficiency
4. Stress Transfer — Generalization from ethanol→oxidative→osmotic conditions

Key findings:

• Maps interactions between 28 ABC transporters × 260 compounds under 3 stress conditions
• Identifies ATM1, MDL1 as top stress-consistent effectors (SIMS score)
• Active learning achieves ≥1.2× label-efficiency over random acquisition
• 47 causal stress–transporter relationships separated from 83 spurious correlations

---

Quickstart (5 minutes)

git clone https://huggingface.co/datasets/BULMA/yeast-abc-atlas
cd yeast-abc-atlas
pip install -r requirements.txt
python scripts/run_pipeline.py --task atlas --cfg env/config.yaml

Or open the notebook directly in Colab:

---

Repository Structure

bulma/
├── README.md
├── requirements.txt
├── env/
│   ├── config.yaml                    # All hyperparameters (seed, dims, epochs)
│   └── environment.yml                # Conda environment
├── notebooks/
│   └── BULMA_full_pipeline.ipynb      # ★ End-to-end reproducible notebook
├── data/
│   ├── raw/                           # FASTA sequences, SMILES
│   ├── processed/                     # Embeddings + labels (auto-generated)
│   └── external/                      # HIP-HOP / SGD benchmark files
├── src/
│   ├── utils/
│   │   ├── io.py                      # Config, seed, I/O helpers
│   │   ├── metrics.py                 # AUROC, AUPRC, bootstrap CI
│   │   ├── plots.py                   # Heatmap, PR curve, waterfall, AL curve
│   │   ├── reproducibility.py         # Grouped CV, leakage-safe pipeline,
│   │   │                              #   calibration, AL logging, CORAL/GroupDRO
│   │   └── env_report.py              # GPU/version reporter, timer, peek()
│   ├── atlas/
│   │   ├── dataset.py                 # PairDataset (protein × compound)
│   │   ├── model_mlp.py               # Two-tower MLP
│   │   ├── train_eval.py              # Training + cold-start CV evaluation
│   │   └── inference.py               # Cartesian-product scoring
│   ├── causal/
│   │   ├── causal_rank.py             # DR-Learner + trimming + placebo
│   │   └── robustness.py              # Section 3.8 robustness add-on + plots
│   ├── active_learning/
│   │   └── al_loop.py                 # 5 strategies: random/uncertainty/
│   │                                  #   diversity/causal/hybrid
│   ├── stress_transfer/
│   │   ├── transfer_eval.py           # Section 5: ethanol→oxidative transfer
│   │   └── causal_al_stress.py        # Section 5: causal-guided AL under stress
│   ├── analysis/
│   │   └── wow_pack.py                # CT-map, SIMS, Discovery Frontier, Uplifts
│   └── validation/
│       ├── lit_crosscheck.py          # Literature anchor gene validation
│       ├── interaction_sensitivity.py # YAP1/PDR1 interaction sensitivity
│       └── external_benchmark.py      # HIP-HOP / MoAmap / SGD concordance
├── scripts/
│   ├── run_pipeline.py                # Unified CLI (atlas / causal / al / all)
│   ├── make_mock_data.py              # Offline synthetic data generator
│   ├── compute_embeddings_protein.py  # ESM-2 embeddings (--mock flag)
│   ├── compute_embeddings_compound.py # ChemBERTa embeddings (--mock flag)
│   ├── snq2_glutathione_test.py       # SNQ2 endogenous substrate test
│   ├── package_release.py             # Section 8: manifest + ZIP release
│   ├── data_curation/
│   │   ├── fetch_abc_sequences.py     # UniProt/SGD harvest + ESM-2 embed
│   │   ├── build_compound_library.py  # Compound library generation
│   │   ├── build_labels.py            # Interaction label table
│   │   ├── build_causal_table.py      # Causal covariates table
│   │   ├── clean_ligands.py           # NaN imputation + RDKit canonicalize
│   │   └── sync_labels.py             # Re-sync labels to valid IDs
│   ├── figures/
│   │   ├── pub_figure_suite.py        # Main publication figures (complete)
│   │   ├── pub_figure_final.py        # Final refined figures (hierarchy spine)
│   │   ├── ct_map_elegant.py          # CT-map: elegant alternatives
│   │   ├── ct_map_heatmap.py          # CT-map: polished contour heatmap
│   │   ├── sims_figure.py             # SIMS waterfall + rank concordance
│   │   ├── supp_figures.py            # Fixed supplementary figures
│   │   └── pipeline_schematic.py      # Graphical pipeline overview
│   └── tables/
│       ├── pub_tables.py              # Main + supplementary tables (CSV+LaTeX)
│       └── pub_tables_enhanced.py     # Enhanced concordance + ATE tables
└── results/                           # Auto-generated JSON snapshots + figures

---

Data

All processed data is available on this Hugging Face dataset page.

File	Description	Rows
data/processed/protein.csv	ESM-2 embeddings (1280-dim) for 28–38 ABC transporters	28–38
data/processed/ligand.csv	ChemBERTa embeddings (768-dim) for 260 compounds	260
data/processed/labels.csv	Binary interaction labels with provenance (assay, condition, replicate)	~9,360
data/processed/causal_table.csv	Expression + covariate table for DR-Learner	6,000

Protein sequences were fetched from UniProt (taxon 559292) for canonical ABC transporters. Compounds were drawn from a curated library of alcohols, aromatics, and heterocycles.

---

Reproducing Results

Step 1 — Install dependencies

pip install -r requirements.txt

Step 2 — Compute embeddings (or use precomputed)

python scripts/compute_embeddings_protein.py   # ESM-2, ~10 min on GPU
python scripts/compute_embeddings_compound.py  # ChemBERTa, ~5 min

Step 3 — Run the full pipeline

# Section 2: Atlas training + evaluation
python scripts/run_pipeline.py --task atlas --cfg env/config.yaml

# Section 3: Causal ranking
python scripts/run_pipeline.py --task causal \
    --causal_csv_in data/processed/causal_table.csv \
    --causal_out results/causal_effects.csv

# Section 4: Active learning
python scripts/run_pipeline.py --task al --cfg env/config.yaml

Step 4 — Notebook (recommended)

Open notebooks/BULMA_full_pipeline.ipynb — every cell is self-contained and runs top-to-bottom.

---

Reproducibility

• Global seed: 17 (set via utils.io.set_seed)
• All splits are deterministic (GroupKFold, cold-transporter, cold-ligand, cold-both)
• A results/env_manifest.json is written at runtime capturing library versions
• USE_MOCK = True generates synthetic data so the notebook runs without external files

---

Citation

@article{harrizi2025bulma,
  title   = {BULMA: A stress-responsive atlas of ATP-binding cassette transporters 
             in Saccharomyces cerevisiae using active learning and causal inference},
  author  = {Harrizi, Saad and Nait Irahal, Imane and Kabine, Mostafa},
  year    = {2025},
  note    = {Preprint}
}

---

License

MIT — see LICENSE.