Zero-shot modelling discovers structurally unprecedented antibiotics against Escherichia coli

It contains the notebooks, input data, ranked compound predictions, FS-Mol training tasks, and generated figures used for zero-shot antibacterial compound prioritization.

Repository tree

.
├── experiments/
│   └── Smiles and Activity.xlsx          # Experimental compounds and activity annotations
├── fsmol/
│   └── train/
│       └── CHEMBL*.jsonl.gz              # FS-Mol training tasks
├── ranking/
│   └── predictions_lifechem_ecoli.csv    # Ranked LifeChem predictions for E. coli activity
├── scripts/
│   ├── predict.ipynb                     # Primary prediction workflow
│   ├── analysis.ipynb                    # Assay analysis and candidate extraction
│   ├── comp.ipynb                        # Comparative analysis
│   └── fs_mol_similarity.ipynb           # FS-Mol similarity analysis
├── figures/
│   ├── plots.ipynb                       # Figure generation notebook
│   ├── hit_sar_cluster_plot.ipynb        # Hit/SAR cluster figure workflow
│   ├── derivatives_umap.*                # Derivative UMAP figure
│   └── prediction_distributions_umap.*   # Prediction UMAP figure
└── README.md

Reproducing analyses

Run the notebooks from the repository root in this order:

1. scripts/predict.ipynb
2. scripts/analysis.ipynb
3. scripts/comp.ipynb
4. scripts/fs_mol_similarity.ipynb
5. figures/plots.ipynb
6. figures/hit_sar_cluster_plot.ipynb

The notebooks expect a Python/Jupyter environment with common cheminformatics and machine-learning packages, including pandas, numpy, matplotlib, seaborn, scikit-learn, rdkit, umap-learn, lightgbm, xgboost, tqdm, fsmol, and twinbooster.

Citation

If you use this repository, please cite the associated publication:

Zero-shot modelling discovers structurally unprecedented antibiotics against Escherichia coli.