chemicalspace 0.1.1

An Object-oriented Representation for Chemical Spaces

  

ChemicalSpace

An Object-Oriented Representation for Chemical Spaces

chemicalspace is a Python package that provides an object-oriented representation for chemical spaces. It is designed to be used in conjunction with the RDKit package, which provides the underlying cheminformatics functionality.

While in the awesome RDKit, the main frame of reference is that of single molecules, here the main focus is on operations on chemical spaces.

Installation

To install chemicalspace you can use pip:

pip install chemicalspace

Usage

The main class in chemicalspace is ChemicalSpace. The class provides a number of methods for working with chemical spaces, including reading and writing, filtering, clustering and picking from chemical spaces.

Basics

Initialization

A ChemicalSpace can be initialized from SMILES strings or RDKit molecules. It optionally takes molecule indices and scores as arguments.

from chemicalspace import ChemicalSpace

smiles = ('CCO', 'CCN', 'CCl')
indices = ("mol1", "mol2", "mol3")
scores = (0.1, 0.2, 0.3)

space = ChemicalSpace(mols=smiles, indices=indices, scores=scores)

print(space)

<ChemicalSpace: 3 molecules | 3 indices | 3 scores>

Reading and Writing

A ChemicalSpace can be read from and written to SMI and SDF files.

from chemicalspace import ChemicalSpace

# Load from SMI file
space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space.to_smi("outputs1.smi")

# Load from SDF file
space = ChemicalSpace.from_sdf("tests/data/inputs1.sdf")
space.to_sdf("outputs1.sdf")

print(space)

<ChemicalSpace: 10 molecules | 10 indices | No scores>

Indexing, Slicing and Masking

Indexing, slicing and masking a ChemicalSpace object returns a new ChemicalSpace object.

Indexing

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

print(space[0])

<ChemicalSpace: 1 molecules | 1 indices | No scores>

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
idx = [1, 2, 4]

print(space[idx])

<ChemicalSpace: 3 molecules | 3 indices | No scores>

Slicing

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

print(space[:2])

<ChemicalSpace: 2 molecules | 2 indices | No scores>

Masking

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
mask = [True, False, True, False, True, False, True, False, True, False]

print(space[mask])

<ChemicalSpace: 5 molecules | 5 indices | No scores>

Deduplicating

Deduplicating a ChemicalSpace object removes duplicate molecules.
See Hashing and Identity for details on molecule identity.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space_twice = space + space  # 20 molecules
space_deduplicated = space_twice.deduplicate()  # 10 molecules

print(space_deduplicated)

<ChemicalSpace: 10 molecules | 10 indices | No scores>

Chunking

A ChemicalSpace object can be chunked into smaller ChemicalSpace objects.
The .chunks method returns a generator of ChemicalSpace objects.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
chunks = space.chunks(chunk_size=3)

for chunk in chunks:
    print(chunk)

<ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 1 molecules | 1 indices | No scores>

Drawing

A ChemicalSpace object can be rendered as a grid of molecules.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space.draw()

Featurizing

Features

A ChemicalSpace object can be featurized as a numpy array of features. By default, ECFP4/Morgan2 fingerprints are used. The features are cached for subsequent calls, and spaces generated by a ChemicalSpace object (e.g. by slicing, masking, chunking) inherit the respective features.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space_slice = space[:6:2]

# Custom ECFP4 features
print(space.features.shape)
print(space_slice.features.shape)

(10, 1024)
(3, 1024)

Custom featurizer

This should take in a rdkit.Chem.Mol molecule, and the numerical return value should be castable to NumPy array (see chemicalspace.utils.MolFeaturizerType).

from chemicalspace import ChemicalSpace
from chemicalspace.utils import maccs_featurizer

space = ChemicalSpace.from_smi("tests/data/inputs1.smi", featurizer=maccs_featurizer)
space_slice = space[:6:2]

# Custom ECFP4 features
print(space.features.shape)
print(space_slice.features.shape)

(10, 167)
(3, 167)

Metrics

A distance metric on the feature space is necessary for clustering, calculating diversity, and identifying neighbors. By default, the jaccard (a.k.a Tanimoto) distance is used. ChemicalSpace takes a metric string argument that allows to specify a sklearn metric.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi", metric='euclidean')

Binary Operations

Single entries

Single entries as SMILES strings or RDKit molecules can be added to a ChemicalSpace object.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space.add("CCO", "mol11")

print(space)

<ChemicalSpace: 11 molecules | 11 indices | No scores>

Chemical spaces

Two ChemicalSpace objects can be added together.

from chemicalspace import ChemicalSpace

space1 = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space2 = ChemicalSpace.from_smi("tests/data/inputs2.smi")

space = space1 + space2

print(space)

<ChemicalSpace: 25 molecules | 25 indices | No scores>

And subtracted from each other to return only molecules in space1 that are not in space2.
See Hashing and Identity for more details.

from chemicalspace import ChemicalSpace

space1 = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space2 = ChemicalSpace.from_smi("tests/data/inputs2.smi")

space = space1 - space2

print(space)

<ChemicalSpace: 5 molecules | 5 indices | No scores>

Hashing and Identity

Individual molecules in a chemical space are hashed by their InChI Keys only (by default), or by InChI Keys and index. Scores do not affect the hashing process.

from chemicalspace import ChemicalSpace

smiles = ('CCO', 'CCN', 'CCl')
indices = ("mol1", "mol2", "mol3")

# Two spaces with the same molecules, and indices
# But one space includes the indices in the hashing process
space_indices = ChemicalSpace(mols=smiles, indices=indices, hash_indices=True)
space_no_indices = ChemicalSpace(mols=smiles, indices=indices, hash_indices=False)

print(space_indices == space_indices)
print(space_indices == space_no_indices)
print(space_no_indices == space_no_indices)

True
False
True

ChemicalSpace objects are hashed by their molecular hashes, in an order-independent manner.

from rdkit import Chem
from rdkit.Chem import inchi
from chemicalspace import ChemicalSpace

mol = Chem.MolFromSmiles("c1ccccc1")
inchi_key = inchi.MolToInchiKey(mol)

space = ChemicalSpace(mols=(mol,))

assert hash(space) == hash(frozenset((inchi_key,)))

The identity of a ChemicalSpace is evaluated on its hashed representation.

from chemicalspace import ChemicalSpace

space1 = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space1_again = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space2 = ChemicalSpace.from_smi("tests/data/inputs2.smi.gz")

print(space1 == space1)
print(space1 == space1_again)
print(space1 == space2)

True
True
False

Copy

ChemicalSpace supports copy and deepcopy operations. Deepcopy allows to fully unlink the copied object from the original one, including the RDKit molecules.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

# Shallow copy
space_copy = space.copy()
assert id(space.mols[0]) == id(space_copy.mols[0])

# Deep copy
space_deepcopy = space.copy(deep=True)
assert id(space.mols[0]) != id(space_deepcopy.mols[0])

Clustering

Labels

A ChemicalSpace can be clustered using by its molecular features. kmedoids, agglomerative-clustering, sphere-exclusion and scaffold are the available clustering methods. Refer to the respective methods in chemicalspace.layers.clustering for more details.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
cluster_labels = space.cluster(n_clusters=3)

print(cluster_labels)

[0 1 2 1 1 0 0 0 0 0]

Clusters

ChemicalSpace.yield_clusters can be used to iterate clusters as ChemicalSpace objects.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
clusters = space.yield_clusters(n_clusters=3)

for cluster in clusters:
    print(cluster)

<ChemicalSpace: 6 molecules | 6 indices | No scores>
<ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 1 molecules | 1 indices | No scores>

KFold Clustering

ChemicalSpace.splits can be used to iterate train/test cluster splits for ML training. At each iteration, one cluster is used as the test set and the rest as the training set. Note that there is no guarantee on the size of the clusters.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

for train, test in space.split(n_splits=3):
    print(train, test)

<ChemicalSpace: 4 molecules | 4 indices | No scores> <ChemicalSpace: 6 molecules | 6 indices | No scores>
<ChemicalSpace: 7 molecules | 7 indices | No scores> <ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 9 molecules | 9 indices | No scores> <ChemicalSpace: 1 molecules | 1 indices | No scores>

Overlap

ChemicalSpace implements methods for calculating the overlap with another space.

Overlap

The molecules of a ChemicalSpace that are similar to another space can be flagged. The similarity between two molecules is calculated by the Tanimoto similarity of their ECFP4/Morgan2 fingerprints.

from chemicalspace import ChemicalSpace

space1 = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space2 = ChemicalSpace.from_smi("tests/data/inputs2.smi.gz")

# Indices of `space1` that are similar to `space2`
overlap = space1.find_overlap(space2, radius=0.6)

print(overlap)

[0 1 2 3 4]

Carving

The overlap between two ChemicalSpace objects can be carved out from one of the objects, so to ensure that the two spaces are disjoint for a given similarity radius.

from chemicalspace import ChemicalSpace

space1 = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space2 = ChemicalSpace.from_smi("tests/data/inputs2.smi.gz")

# Carve out the overlap from `space1`
space1_carved = space1.carve(space2, radius=0.6)

print(space1_carved)

<ChemicalSpace: 5 molecules | 5 indices | No scores>

Dimensionality Reduction

ChemicalSpace implements methods for dimensionality reduction by pca, tsne or umap projection of its features.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
proj = space.project(method='pca')

print(proj.shape)

(10, 2)

Picking

A subset of a ChemicalSpace can be picked by a number of acquisition strategies.
See chemicalspace.layers.acquisition for details.

from chemicalspace import ChemicalSpace
import numpy as np

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

space_pick_random = space.pick(n=3, strategy='random')
print(space_pick_random)

space_pick_diverse = space.pick(n=3, strategy='maxmin')
print(space_pick_diverse)

space.scores = np.array(range(len(space)))  # Assign dummy scores
space_pick_greedy = space.pick(n=3, strategy='greedy')
print(space_pick_greedy)

<ChemicalSpace: 3 molecules | 3 indices | No scores>
<ChemicalSpace: 3 molecules | 3 indices | 3 scores>

Uniqueness and Diversity

Uniqueness

The uniqueness of a ChemicalSpace object can be calculated by the number of unique molecules.

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")
space_twice = space + space  # 20 molecules
uniqueness = space_twice.uniqueness()

print(uniqueness)

0.5

Diversity

The diversity of a ChemicalSpace object can be calculated as:

• The average of the pairwise distance matrix
• The normalized Vendi score of the same distance matrix.

The Vendi score can be interpreted as the effective number of molecules in the space, and here it is normalized by the number of molecules in the space taking values in the range [0, 1].

from chemicalspace import ChemicalSpace

space = ChemicalSpace.from_smi("tests/data/inputs1.smi")

diversity_int = space.diversity(method='internal-distance')
diversity_vendi = space.diversity(method='vendi')
print(diversity_int)
print(diversity_vendi)

# Dummy space with the same molecule len(space) times
space_redundant = ChemicalSpace(mols=tuple([space.mols[0]] * len(space)))

diversity_int_redundant = space_redundant.diversity(method='internal-distance')
diversity_vendi_redundant = space_redundant.diversity(method='vendi')

print(diversity_int_redundant)
print(diversity_vendi_redundant)

0.7730273985449335
0.12200482273434754
0.0
0.1

Advanced

Layers

ChemicalSpace is implemented as a series of layers that provide the functionality of the class. As can be seen in the source code, the class simply combines the layers.

If only a subset of the functionality of ChemicalSpace is necessary, and lean objects are a priority, one can combine only the required layers:

from chemicalspace.layers.clustering import ChemicalSpaceClusteringLayer
from chemicalspace.layers.neighbors import ChemicalSpaceNeighborsLayer


class MyCustomSpace(ChemicalSpaceClusteringLayer, ChemicalSpaceNeighborsLayer):
    pass


space = MyCustomSpace(mols=["c1ccccc1"])
space

<MyCustomSpace: 1 molecules | No indices | No scores>

Development

Installation

Install the development dependencies with pip:

pip install -e .[dev]

Hooks

The project uses pre-commit for code formatting, linting and testing. Install the hooks with:

pre-commit install

Documentation

The documentation can be built by running:

cd docs
./rebuild.sh