Datamol reactions and bundled data (0.13.0)

Sources: reaction API, data API, and released source.

Reaction outputs and validation

dm.reactions.apply_reaction(rxn, reactants, product_index=None, single_product_group=False, as_smiles=False, rm_attach=False, sanitize=True) returns all product groups by default. With product_index=None, each group is a list of product molecules. product_index=0 instead selects the first product slot from each group, producing a flat list. No matches return []; failed sanitization can leave None entries. single_product_group=True randomly chooses a group; it is not a chemically ranked or deterministically first product. Retain all groups for enumeration.

import datamol as dm
from rdkit.Chem import rdChemReactions
# Computational substitution pattern; no reaction conditions or yield are implied.
rxn = rdChemReactions.ReactionFromSmarts("[C:1](=[O:2])[OH:3]>>[C:1](=[O:2])[Cl:3]")
reactant = dm.to_mol("CC(=O)O")
groups = dm.reactions.apply_reaction(rxn, (reactant,), single_product_group=False)
products = [p for group in groups for p in group if p is not None]
product_smiles = sorted({dm.to_smiles(p) for p in products})
assert product_smiles == ["CC(=O)Cl"]
assert dm.reactions.apply_reaction(rxn, (dm.to_mol("CC"),)) == []

rxn_from_smarts, rxn_from_block, and rxn_from_block_file initialize reaction objects; rxn_to_smarts, rxn_to_block, and rxn_to_block_file export them. can_react(rxn, mol) checks reactant-pattern matching. is_reaction_ok(rxn) invokes RDKit preprocessing/sanitization; it does not validate synthetic feasibility, conditions, selectivity, safety or yield. Atom mapping tracks atoms in a computational template; it does not imply atom-balanced laboratory chemistry. Deduplicate products by an explicit stereochemistry/tautomer policy and keep input/product-group provenance.

These simple templates were executed on matching toy reactants; they are illustrative chemical transformations, not validated synthesis instructions:

amide_rxn = rdChemReactions.ReactionFromSmarts(
    "[N;H1,H2:1].[C:2](=[O:3])[OH]>>[N:1][C:2](=[O:3])"
)
suzuki_rxn = rdChemReactions.ReactionFromSmarts(
    "[c:1][Br].[c:2][B]([OH])[OH]>>[c:1][c:2]"
)
esterification = rdChemReactions.ReactionFromSmarts(
    "[C:1][OH:2].[C:3](=[O:4])[Cl]>>[C:1][O:2][C:3](=[O:4])"
)

For a library, preserve (source_id, product_group, product_slot) and log both empty matches and invalid products; do not silently append nested groups as though they were Mol.

Bundled datasets

All three accept as_df=False for molecules. These are bundled toy snapshots, not a fresh download of the originating datasets. Inspect target definitions, units, duplicates, license and source provenance before drawing research conclusions. A bundled split is an example boundary, not evidence that it is scaffold-disjoint or suitable for your task.

import numpy as np
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_absolute_error
sol = dm.data.solubility()
train = sol.loc[sol["split"] == "train"]
test = sol.loc[sol["split"] == "test"]
# One molecule per call: dm.to_fp does not featurize a pandas Series.
X_train = np.stack([dm.to_fp(m, radius=2) for m in train["mol"]])
X_test = np.stack([dm.to_fp(m, radius=2) for m in test["mol"]])
model = RandomForestRegressor(n_estimators=20, random_state=42, n_jobs=1)
model.fit(X_train, train["SOL"].astype(float))
predictions = model.predict(X_test)
mae = mean_absolute_error(test["SOL"].astype(float), predictions)
print(f"Toy held-out MAE: {mae:.3f}")

This small executed baseline checks the API and feature/label alignment. It is not a benchmark claim or evidence of prospective predictive performance.