Pharma ML Tools

Use this skill when the user asks for compound-library profiling, chemistry ML feature generation, medicinal-chemistry screening, or benchmark dataset preparation.

Typical triggers:

• standardize and profile a compound library before QSAR or screening
• featurize molecules with molfeat for downstream ML
• pull benchmark-ready ADME, toxicity, DTI, or DDI datasets with PyTDC
• apply medicinal-chemistry rules or alert filters before prioritization
• compare scaffolds, duplicates, or diversity in a virtual-screening library
• prepare a docking or QSAR campaign with better compound hygiene

Environment Check

which python3 || true
python3 - <<'PY'
mods = ["pandas", "numpy", "datamol", "molfeat", "medchem"]
for name in mods:
    try:
        __import__(name)
        print(f"{name}: ok")
    except Exception as exc:
        print(f"{name}: missing ({exc})")
try:
    import tdc
    print("PyTDC: ok")
except Exception as exc:
    print(f"PyTDC: missing ({exc})")
PY

If a requested module is missing, say so explicitly. Do not claim the screen, featurization, or dataset pull completed.

Bundled Assets

• templates/datamol_library_profile.py
• templates/molfeat_featurize.py
• templates/pytdc_dataset_fetch.py
• templates/medchem_screen.py

Preferred Workflow

1. Normalize the input table first and identify the exact SMILES column.
2. Run datamol_library_profile.py before building models so duplicates, invalid structures, and scaffold concentration are visible.
3. Use medchem_screen.py before large docking or QSAR jobs to flag problematic chemotypes.
4. Use molfeat_featurize.py when the user needs model-ready features rather than only descriptor summaries.
5. Use pytdc_dataset_fetch.py when the user needs reproducible public benchmark datasets rather than ad hoc CSV collection.
6. Keep outputs under a dedicated directory such as ./pharma_ml/.

Library Profiling With Datamol

python3 templates/datamol_library_profile.py \
  --input libraries/kinase_hits.csv \
  --smiles-column smiles \
  --id-column compound_id \
  --output pharma_ml/kinase_hits_profile.csv \
  --summary pharma_ml/kinase_hits_profile.json

Use this first for:

• canonical SMILES and InChIKey generation
• invalid structure detection
• scaffold counts
• molecular-property summaries before modeling

Molfeat Featurization

python3 templates/molfeat_featurize.py \
  --input libraries/kinase_hits.csv \
  --smiles-column smiles \
  --id-column compound_id \
  --featurizer ecfp \
  --output pharma_ml/kinase_hits_ecfp.csv \
  --summary pharma_ml/kinase_hits_ecfp.json

Supported baseline featurizers in the bundled template:

• ecfp
• maccs
• rdkit2d

Use this for local QSAR, ranking, clustering, or embedding handoff.

PyTDC Benchmark Datasets

python3 templates/pytdc_dataset_fetch.py \
  --task adme \
  --dataset Caco2_Wang \
  --split-method scaffold \
  --out-dir pharma_ml/caco2_wang

Good use cases:

• ADME or toxicity baselines
• DTI or DDI dataset retrieval
• reproducible train/valid/test splits for benchmarking

Medicinal-Chemistry Screening

python3 templates/medchem_screen.py \
  --input libraries/kinase_hits.csv \
  --smiles-column smiles \
  --id-column compound_id \
  --output pharma_ml/kinase_hits_medchem.csv \
  --summary pharma_ml/kinase_hits_medchem.json

Use this for:

• Rule-of-Five and lead-like checks
• alert-oriented library triage
• quick pass/fail summaries before wet-lab nomination

Treat these filters as prioritization heuristics, not hard truth.

DiffDock Boundary

If the user asks for diffusion docking or deep pose generation, acknowledge that this runtime already includes docking-tools for Vina-style workflows, but DiffDock-class workflows require a heavier environment with PyTorch Geometric, model weights, and usually GPU acceleration. Do not pretend that support is bundled unless the environment is confirmed.

Output Expectations

Good answers should mention:

• the exact input file and SMILES column
• which template ran
• valid versus invalid molecule counts
• whether outputs are profiling, features, dataset splits, or medchem filters
• what files were written
• any module, network, or dataset-license caveats

Related Skills

For public APIs such as PubChem, ChEMBL, openFDA, ClinicalTrials.gov, or OpenAlex, activate pharma-db-tools. For RDKit descriptors, ADMET heuristics, DrugBank, QSAR, or structure-aware affinity, activate chem-tools. For docking and pose-level workflows, activate docking-tools.