chem-tools

Files (15)

• SKILL.md
• admet_screen.py · templates/
• assay_data_prepare.py · templates/
• binding_affinity_predict.py · templates/
• bioactivity_predict.py · templates/
• chem_ml_utils.py · templates/
• deepchem_featurize.py · templates/
• drugbank_lookup.py · templates/
• pdbbind_prepare.py · templates/
• protein_ligand_affinity.py · templates/
• protein_ligand_benchmark.py · templates/
• pyscf_single_point.py · templates/
• qsar_benchmark.py · templates/
• rdkit_descriptors.py · templates/
• virtual_screen.py · templates/

SKILL.md (first lines)

# Chem Tools

Use this skill when the user asks to:
- featurize molecules from SMILES, CSV, TSV, or text inputs
- use DeepChem for molecular ML preprocessing, fingerprints, or dataset preparation
- use PySCF for small-molecule HF or DFT calculations
- run ADMET triage or structural alert screening
- normalize ChEMBL, BindingDB, MoleculeNet, or generic assay tables into DrugClaw-ready datasets
- adapt PDBbind-style structure datasets into benchmark-ready CSV tables
- benchmark QSAR baselines with scaffold or random splits
- train or apply ligand binding-affinity models from labeled SMILES
- train or apply structure-aware protein-ligand affinity models from complexes
- benchmark structure-aware protein-ligand models on grouped or random splits
- train or apply bioactivity models from labeled SMILES
- rank a library for virtual screening with chemistry, activity, affinity, and docking signals
- search DrugBank from a local export or the online discovery API, export the matched drug structure, or read drug descriptive properties
- compute chemistry follow-up checks on docking hits
- inspect small-molecule descriptors, simple QM sanity checks, or ligand ranking features

## Environment Check

Do not assume the chemistry stack is available. Check first.

```bash
which python3 || true
python3 - <<'PY'
mods = ["deepchem", "pyscf", "rdkit", "numpy", "pandas", "sklearn"]
for name in mods:
    try:
        __import__(name)
        print(f"{name}: ok")
    except Exception as exc:
        print(f"{name}: missing ({exc})")
PY
```

If key modules are missing, say so immediately and recommend the unified `drug-sandbox` image documented in `docs/operations/science-runtime.md`.

## Bundled Assets

- `templates/deepchem_featurize.py`
- `templates/pyscf_single_point.py`
- `templates/rdkit_descriptors.py`
- `templates/admet_screen.py`
- `templates/assay_data_prepare.py`
- `templates/pdbbind_prepare.py`
- `templates/binding_affinity_predict.py`
- `templates/bioactivity_predict.py`
- `templates/drugbank_lookup.py`
- `templates/protein_ligand_affinity.py`
- `templates/protein_ligand_benchmark.py`
- `templates/qsar_benchmark.py`
- `templates/virtual_screen.py`

Use these templates instead of rewriting the same chemistry scripts from scratch.

## Preferred Workflow

1. Identify the input type: inline SMILES, CSV/TSV, text list, XYZ, or inline atom string.
2. Run the smallest deterministic template first.
3. For predictive work, separate descriptive heuristics from supervised models.
4. Save structured outputs such as `.npy`, `.csv`, `.joblib`, or `.json`.
5. Benchmark supervised models with scaffold or external splits before treating them as useful.
6. Report the exact featurizer, basis set, method, model algorithm, label definition, split strategy, and whether convergence or training succeeded.
7. Call out whether a model is ligand-only or structure-aware.
8. State clearly whether the result is descriptive, predictive, or quantum-mechanical.

## DeepChem

Use `templates/deepchem_featurize.py` for:
- circular fingerprints
- MACCS keys
- Mol2Vec fingerprints
- quick dataset preparation for downstream ML

Quick start:

```bash
python3 templates/deepchem_featurize.py \
  --smiles "CCO" "c1ccccc1" \
  --featurizer circular \
  --output-prefix chem/deepchem/demo
```

CSV input example:

```bash
python3 templates/deepchem_featurize.py \
  --input ligands.csv \
  --smiles-column smiles \
  --id-column ligand_id \
  --featurizer maccs \
  --output-prefix chem/deepchem/ligands
```

Deliverables:
- `.npy` feature matrix
- `.summary.csv` with per-molecule stats
- `.json` metadata with featurizer and shape

If the user asks for actual DeepChem neural models, verify the required backend first. Do not assume TensorFlow or PyTorch models are available just because `deepchem` imports.

## RDKit

Use `templates/rdkit_descriptors.py` for:
- common molecular descriptors
- Lipinski and Veber rule flags
- quick ligand triage before docking or QM follow-up

Quick start:

```bash
python3 templates/rdkit_descriptors.py \
  --smiles "CCO" "c1ccccc1O" \
  --output chem/rdkit/descriptors.csv \
  --summary chem/rdkit/summary.json
```

CSV input example:

From DrugClaw/DrugClaw, Apache-2.0.

More from DrugClaw/DrugClaw