Datamol Cheminformatics Skill

Overview

Datamol is a Python library that provides a lightweight, Pythonic abstraction layer over RDKit for molecular cheminformatics. Simplify complex molecular operations with sensible defaults, efficient parallelization, and modern I/O capabilities. All molecular objects are native rdkit.Chem.Mol instances, ensuring full compatibility with the RDKit ecosystem.

Checked against: datamol 0.12.5 (PyPI stable, released 2024-06-10; still the current release as of August 2026). Examples target datamol 0.12.x. Since 0.10.0, modules are lazy-loaded by default (set DATAMOL_DISABLE_LAZY_LOADING=1 to disable). Since 0.12.2, RDKit is a direct PyPI dependency of datamol. Fingerprints use RDKit's rdFingerprintGenerator API (0.12.5+).

Key capabilities:

• Molecular format conversion (SMILES, SELFIES, InChI)
• Structure standardization and sanitization
• Molecular descriptors and fingerprints
• 3D conformer generation and analysis
• Clustering and diversity selection
• Scaffold and fragment analysis
• Chemical reaction application
• Visualization and alignment
• Batch processing with parallelization
• Cloud storage support via fsspec

Installation and Setup

Guide users to install datamol:

uv pip install datamol

RDKit is installed automatically with datamol. For remote file paths (S3, GCS, HTTP), install the matching fsspec backend:

uv pip install s3fs   # AWS S3
uv pip install gcsfs  # Google Cloud Storage

Import convention:

import datamol as dm

Core Workflows

Ten workflow areas, each with worked code, are documented in references/core_workflows.md:

Three end-to-end pipelines — load/filter/analyze, SAR by scaffold series, and virtual screening — are in references/workflow_patterns.md.

Parallelization

Datamol includes built-in parallelization for many operations. Use n_jobs parameter:

• n_jobs=1: Sequential (no parallelization)
• n_jobs=-1: Use all available CPU cores
• n_jobs=4: Use 4 cores

Functions supporting parallelization:

• dm.read_sdf(..., n_jobs=-1)
• dm.descriptors.batch_compute_many_descriptors(..., n_jobs=-1)
• dm.cluster_mols(..., n_jobs=-1)
• dm.pdist(..., n_jobs=-1)
• dm.conformers.sasa(..., n_jobs=-1)

Progress bars: Many batch operations support progress=True parameter.

Reference Documentation

For detailed API documentation, consult these reference files:

• references/core_api.md: Core namespace functions (conversions, standardization, fingerprints, clustering)
• references/io_module.md: File I/O operations (read/write SDF, CSV, Excel, remote files)
• references/conformers_module.md: 3D conformer generation, clustering, SASA calculations
• references/descriptors_viz.md: Molecular descriptors and visualization functions
• references/fragments_scaffolds.md: Scaffold extraction, BRICS/RECAP fragmentation
• references/reactions_data.md: Chemical reactions and toy datasets

Best Practices

1. Always standardize molecules from external sources:

   mol = dm.standardize_mol(mol, disconnect_metals=True, normalize=True, reionize=True)

2. Check for None values after molecule parsing:

   mol = dm.to_mol(smiles)
   if mol is None:
       # Handle invalid SMILES

3. Use parallel processing for large datasets — n_jobs/progress are accepted by the batch entry points listed under Parallelization, not by every function:

   desc_df = dm.descriptors.batch_compute_many_descriptors(mols, n_jobs=-1, progress=True)

4. Use cloud I/O only when requested — confirm remote write paths; install s3fs/gcsfs as needed:

   df = dm.read_sdf("s3://bucket/compounds.sdf")

5. Use appropriate fingerprints for similarity:
   • ECFP (Morgan): General purpose, structural similarity
   • MACCS: Fast, smaller feature space
   • Atom pairs: Considers atom pairs and distances
6. Consider scale limitations:
   • Butina clustering: ~1,000 molecules (full distance matrix)
   • For larger datasets: Use diversity selection or hierarchical methods
7. Scaffold splitting for ML: Ensure proper train/test separation by scaffold
8. Align molecules when visualizing SAR series

Error Handling

# Safe molecule creation
def safe_to_mol(smiles):
    try:
        mol = dm.to_mol(smiles)
        if mol is not None:
            mol = dm.standardize_mol(mol)
        return mol
    except Exception as e:
        print(f"Failed to process {smiles}: {e}")
        return None

# Safe batch processing
valid_mols = []
for smiles in smiles_list:
    mol = safe_to_mol(smiles)
    if mol is not None:
        valid_mols.append(mol)

Integration with Machine Learning

Datamol ships with scipy and scikit-learn as dependencies. Import them as normal PyPI packages — they are not scripts bundled in this skill.

import numpy as np

# Feature generation
X = np.array([dm.to_fp(mol) for mol in mols])

# Or descriptors
desc_df = dm.descriptors.batch_compute_many_descriptors(mols, n_jobs=-1)
X = desc_df.values

# Train model (scikit-learn PyPI package)
from sklearn.ensemble import RandomForestRegressor  # third-party library
model = RandomForestRegressor()
model.fit(X, y_target)

# Predict
predictions = model.predict(X_test)

Troubleshooting

Issue: Molecule parsing fails

• Solution: Use dm.standardize_smiles() first or try dm.fix_mol()

Issue: Memory errors with clustering

• Solution: Use dm.pick_diverse() instead of full clustering for large sets

Issue: Slow conformer generation

• Solution: Reduce n_confs or increase rms_cutoff to generate fewer conformers

Issue: Remote file access fails

• Solution: Install the matching fsspec backend (uv pip install s3fs or gcsfs) and verify only the provider credentials needed for that backend are set (see Installation and Setup above)

Composing with the rest of the bundle

• rdkit → instead: when you need control datamol does not expose — custom sanitization, partial sanitization, reaction fingerprints, pharmacophore features.
• medchem → after: rule-based triage (Lipinski/Veber/CNS, PAINS and NIBR alerts) on the standardized molecules produced here. Same maintainers, same Mol objects.
• molfeat → after: featurization for a model, from ECFP through pretrained ChemBERTa.
• chembl → before: measured bioactivity to standardize and cluster.
• admet-prediction → after: standardize and desalt here first. ADMET-AI predicts on the SMILES string as given, so a salt or mixture yields a number for the wrong species.
• chemical-space / generative-design → after: dm.pick_diverse and scaffold grouping are how you cut a generated or enumerated set down to what is worth making.

Additional Resources

• Datamol Documentation: https://docs.datamol.io/
• RDKit Documentation: https://www.rdkit.org/docs/
• GitHub Repository: https://github.com/datamol-io/datamol