molcluster

A collection of tools to cluster molecules for cheminformatics applications.

How to use

pip install molcluster

You can use any function to generate descriptors for the molecules in the dataset. For instance, we could use Morgan fingerprints from RDkit to generate a vector of 1024 bits for each molecule.

from molcluster.unsupervised_learning.clustering import KMeansClustering, HDBSCANClustering, ButinaClustering, HierarchicalClustering
from molcluster.unsupervised_learning.transform import UMAPTransform, PCATransform

data = pd.read_csv('../data/fxa_processed.csv')

X = np.array([Chem.AllChem.GetMorganFingerprintAsBitVect(x, radius=1024) for x in list(map(Chem.MolFromSmiles, data.processed_smiles.values))])

Dimensionality reduction

Principal component analysis (PCA)

pca_reducer = PCATransform(X)

pca_embeddings = pca_reducer.reduce(n_components=2)
pca_embeddings[0:5]

array([[1.2142797 , 0.46797618],
       [1.44474151, 0.64233027],
       [1.51234623, 0.87651611],
       [3.77443183, 1.29613805],
       [3.654247  , 1.80719829]])

UMAP

umap_reducer = UMAPTransform(X)

umap_embeddings = umap_reducer.reduce(n_neighbors=50, min_dist=0.25, metric='euclidean')
umap_embeddings[0:5]

array([[ 1.5952768 ,  4.4337296 ],
       [ 1.5278653 ,  4.5167828 ],
       [ 1.3860604 ,  4.543414  ],
       [ 1.7233835 , -1.6080631 ],
       [ 0.79702693, -1.1479477 ]], dtype=float32)

Clustering

Kmeans clustering with 10 clusters

clustering_kmeans = KMeansClustering(X)
labels = clustering_kmeans.cluster(n_clusters=10)
labels[0:5]

array([0, 0, 0, 3, 3], dtype=int32)

Using the elbow method to select the optimal number of clusters

clustering_kmeans.elbow_method(n_clusters=np.arange(2, 20))

Butina clustering with similarity threshold > 0.7

mol_list = data.processed_smiles.values

clustering_butina = ButinaClustering(mol_list)
labels = clustering_butina.cluster(sim_cutoff=0.7)
labels[0:5]

[34, 34, 34, 1, 131]

HDBSCAN clustering

clustering_hdbscan = HDBSCANClustering(X)
labels = clustering_hdbscan.cluster(min_cluster_size=5,min_samples=1,metric='euclidean')

np.unique(labels)[0:5]

Agglomerative clustering (e.g. using Ward’s method)

clustering_agg = HierarchicalClustering(X)

labels = clustering_agg.cluster(n_clusters=None, distance_threshold=0.25, linkage='ward')
labels[0:5]

Plotting a dendrogram

clustering_agg.plot_dendrogram(truncate_mode="level", p=5)