deepchem

name: deepchem description: Molecular ML with DeepChem - featurizers, models, and molecular property prediction

Overview

DeepChem is a Python library for molecular machine learning. It provides featurizers, datasets, model wrappers, and splitters for drug discovery and materials science tasks.

Featurizers

ECFP (Extended Connectivity Fingerprints): dc.feat.CircularFingerprint(size=1024, radius=2)
GraphConv: dc.feat.ConvMolFeaturizer() -- converts molecules to graph objects.
Weave: dc.feat.WeaveFeaturizer() -- atom and pair features.
RDKitDescriptors: dc.feat.RDKitDescriptors() -- 200+ physicochemical descriptors.

Typical Workflow

import deepchem as dc

# Load or create dataset from SMILES
featurizer = dc.feat.CircularFingerprint(size=1024, radius=2)
loader = dc.data.CSVLoader(tasks=["activity"], feature_field="smiles", featurizer=featurizer)
dataset = loader.create_dataset("data.csv")

# Split
splitter = dc.splits.ScaffoldSplitter()
train, valid, test = splitter.train_valid_test_split(dataset)

# Train model
model = dc.models.MultitaskClassifier(n_tasks=1, n_features=1024, layer_sizes=[512, 256])
model.fit(train, nb_epoch=50)

# Evaluate
metric = dc.metrics.Metric(dc.metrics.roc_auc_score)
print(model.evaluate(test, [metric]))

Graph Convolutional Models

featurizer = dc.feat.ConvMolFeaturizer()
model = dc.models.GraphConvModel(n_tasks=1, mode="classification")

Key Details

Splitters: RandomSplitter, ScaffoldSplitter (preferred for generalization), ButinaSplitter.
MoleculeNet benchmarks available via dc.molnet.load_*() (e.g., load_tox21(), load_bbbp()).
Install: pip install deepchem.