mhfp 1.2

Molecular MHFP fingerprints for cheminformatics applications

MHFP

MHFP6 (MinHash fingerprint, up to six bonds) is a molecular fingerprint which encodes detailed substructures using the extended connectivity principle of ECFP in a fundamentally different manner, increasing the performance of exact nearest neighbor searches in benchmarking studies and enabling the application of locality sensitive hashing (LSH) approximate nearest neighbor search algorithms. To describe a molecule, MHFP6 extracts the SMILES of all circular substructures around each atom up to a diameter of six bonds and applies the MinHash method to the resulting set. MHFP6 outperforms ECFP4 in benchmarking analog recovery studies. Furthermore, MHFP6 outperforms ECFP4 in approximate nearest neighbor searches by two orders of magnitude in terms of speed, while decreasing the error rate.

Associated Publication: https://chemrxiv.org/articles/A_Probabilistic_Molecular_Fingerprint_for_Big_Data_Settings/7176350

Materials

The performance of MHFP has been avaluated using Open-source platform to benchmark fingerprints for ligand-based virtual screening and ChEMBL24. Python (pyspark) scripts for the ChEMBL24 benchmark can be found in this repository.

Dependencies

MHFP requires Python 3.x.

MHFP requires the cheminformatics library RDKit.

Installation

MHFP can be installed using the Python packet manager pip:

pip install mhfp

Usage

Usage is straightforward:

from mhfp.encoder import MHFPEncoder

mhfp_encoder = MHFPEncoder()

fp_a = mhfp_encoder.encode('CCOC1=C(C=C(C=C1)S(=O)(=O)N(C)C)C2=NC(=O)C3=C(N2)C(=NN3C)C(C)(C)C')
fp_b = mhfp_encoder.encode('CCCC1=NN(C2=C1NC(=NC2=O)C3=C(C=CC(=C3)S(=O)(=O)N4CCN(CC4)C)OCC)C')
fp_c = mhfp_encoder.encode('O=C(OC)C(C1CCCCN1)C2=CC=CC=C2')

print(MHFPEncoder.distance(fp_a, fp_a))
print(MHFPEncoder.distance(fp_a, fp_b))
print(MHFPEncoder.distance(fp_a, fp_c))
print(MHFPEncoder.distance(fp_b, fp_c))

#> 0.0
#> 0.45849609375
#> 0.97998046875
#> 0.97216796875

LSH Forest

As the fingerprints are created using the locality sensitive hashing scheme MinHash, they are ready to be used with LSH-based algorithms:

from mhfp.encoder import MHFPEncoder
from mhfp.lsh_forest import LSHForestHelper

mhfp_encoder = MHFPEncoder()

fp_a = mhfp_encoder.encode('CCOC1=C(C=C(C=C1)S(=O)(=O)N(C)C)C2=NC(=O)C3=C(N2)C(=NN3C)C(C)(C)C')
fp_b = mhfp_encoder.encode('CCCC1=NN(C2=C1NC(=NC2=O)C3=C(C=CC(=C3)S(=O)(=O)N4CCN(CC4)C)OCC)C')
fp_c = mhfp_encoder.encode('O=C(OC)C(C1CCCCN1)C2=CC=CC=C2')
fp_q = mhfp_encoder.encode('COC1=C(C=C(C=C1)S(=O)(O)N(C)C)C2=NC(=O)C3=C(N2)C(=NN3C)C(C)(C)C')

lsh_forest_helper = LSHForestHelper()

fps = [fp_a, fp_b, fp_c]

for i, fp in enumerate(fps):
    lsh_forest_helper.add(i, fp)

lsh_forest_helper.index()

print(lsh_forest_helper.query(fp_q, 1, fps))

#> [0]

Documentation

MHFPEncoder

The class for encoding SMILES and RDKit molecule instances as MHFP fingerprints.

__init__(n_permutations = 2048, seed = 42)

Parameter	Default	Description
n_permutations	2048	Analogous to the number of bits ECFP fingerprints are folded into. Higher is better, lower is less exact.
seed	42	The seed for the MinHash operation. Has to be the same for comparable fingerprints.

MHFPEncoder.encode(in_smiles, radius = 3, rings = True, kekulize = True, sanitize = True)

Parameter	Default	Description
in_smiles		The input SMILES.
radius	3	Analogous to the radius for the Morgan fingerprint. The default radius 3 encodes SMILES to MHFP6.
rings	True	Whether rings in the molecule are included in the fingerprint. As a radii of 3 fails to encode rings and there is no way to determine ring-membership in a substructure SMILES, this considerably increases performance.
kekulize	True	Whether or not to kekulize the molecule before extracting substructure SMILES.
sanitize	True	Whether or not to sanitize the molecule (using RDKit) before extracting substructure SMILES.

Returns a numpy.ndarray containing the fingerprint hashes.

MHFPEncoder.encode_mol(in_mol, radius = 3, rings = True, kekulize = True)

Parameter	Default	Description
in_mol		The input RDKit molecule instance.
radius	3	Analogous to the radius for the Morgan fingerprint. The default radius 3 encodes SMILES to MHFP6.
rings	True	Whether rings in the molecule are included in the fingerprint. As a radii of 3 fails to encode rings and there is no way to determine ring-membership in a substructure SMILES, this considerably increases performance.
kekulize	True	Whether or not to kekulize the molecule before extracting substructure SMILES.

Returns a numpy.ndarray containing the fingerprint hashes.

MHFPEncoder.distance(a, b)

Parameter	Default	Description
a		A numpy.ndarray containing fingerprint hashes.
b		A numpy.ndarray containing fingerprint hashes.

Returns a float representing the distance between two MHFP encoded molecules.

LSHForestHelper

If you know your way around Python, I suggest you use the LSHForest class directly. See code for details.

__init__(dims = 2048, n_prefix_trees = 64)

Parameter	Default	Description
dims	2048	Has to be equal to the number of permutations n_permutations set when initializing MHFPEncoder
n_prefix_trees	64	The number of prefix trees. Lower number results in lower memory usage but less quality

LSHForestHelper.add(key, mhfp):

Parameter	Default	Description
key		The key that will be retrieved by a query. It's recommended to use integers.
mhfp		The numpy.ndarray containing the fingerprint hashes.

LSHForestHelper.index():

Has to be run after adding entities and before running queries!

MHFPEncoder.query(query_mhfp, k, data, kc=10)

Parameter	Default	Description
query_mhfp		A numpy.ndarray containing fingerprint hashes.
k		The number of nearest neighbors to be retrieved.
data		Data mapping keys to fingerprints. See LSH Forest example for details.
kc	10	Initially k is multiplied by kc and k · kc (ANN) results are returned. The top k are then selected from these k · kc using linear scan.