RDKit fingerprint to numpy array benchmarking

[hadim]

A quick benchmark about RDKit fingerprints conversion to numpy array:

from rdkit.Chem import DataStructs
import datamol as dm

smiles = "Cc1nc2ccc(NC(=O)[C@H](C)Oc3ccc4c(c3)OCO4)cc2c(=O)n1-c1ccccc1"
mol = dm.to_mol(smiles)

bfp = dm.to_fp(mol, as_array=False)
bfp
# <rdkit.DataStructs.cDataStructs.ExplicitBitVect at 0x7f5482ea7990>

%%timeit
fp1 = dm.fp_to_array(bfp)  # the datamol way
# 10.9 µs ± 331 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

%%timeit
fp2 = np.array(bfp)  # the RDKit way (1)
# 853 µs ± 6.4 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%%timeit
fp3 = np.zeros((2048,), dtype=int)
DataStructs.ConvertToNumpyArray(bfp, fp3)  # the RDKit way (2)
# 34.1 µs ± 100 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)

# sanity check
fp1 = dm.fp_to_array(bfp)  # the datamol way
fp2 = np.array(bfp)  # the RDKit way (1)
fp3 = np.zeros((2048,), dtype=int)
DataStructs.ConvertToNumpyArray(bfp, fp3)  # the RDKit way (2)

np.allclose(fp1, fp2), np.allclose(fp1, fp3), np.allclose(fp2, fp3)
# (True, True, True)

In conclusion, dm.fp_to_array seems to be faster than the RDKit alternatives (at least on my machine) with ExplicitBitVect. I haven't run the benchmark for the other types of fingerprints.

[hadim]

A numpy conversion has been merged into upstream rdkit. See rdkit/rdkit#3863 for details.

[hadim]

I guess we should run the benchmark again.