deepchem

DeepChem — Deep Learning for Drug Discovery

Overview

DeepChem is an open-source Python framework providing a unified API for molecular machine learning across drug discovery, materials science, and quantum chemistry. It wraps 60+ model architectures (graph neural networks, transformers, classical ML) with 50+ molecular featurizers and standardized datasets (MoleculeNet), enabling end-to-end workflows from SMILES strings to trained predictive models.

When to Use

• Predicting molecular properties (solubility, toxicity, binding affinity) from SMILES
• Benchmarking models on MoleculeNet standardized datasets (BBBP, Tox21, ESOL, FreeSolv, etc.)
• Training graph neural networks on molecular graphs (GCN, GAT, AttentiveFP, MPNN, DMPNN)
• Fine-tuning pretrained chemical language models (ChemBERTa, GROVER, MolFormer)
• Running hyperparameter optimization for molecular ML models
• Virtual screening and hit prioritization with trained models
• Materials property prediction from crystal structures (CGCNN, MEGNet)
• Protein-ligand interaction modeling and binding affinity prediction
• For fingerprint-based cheminformatics without deep learning, use rdkit-cheminformatics instead
• For featurization only (no model training), use molfeat-molecular-featurization instead

Prerequisites

• Python packages: deepchem (core), torch or tensorflow (backend-dependent models)
• GPU: Recommended for graph neural networks and transformer models; CPU sufficient for classical ML and fingerprint models
• Data: SMILES strings with property labels (CSV), or MoleculeNet datasets (auto-downloaded)

# Core installation (includes RDKit, scikit-learn, XGBoost)
pip install deepchem

# With PyTorch backend (GNN models)
pip install deepchem[torch]

# With TensorFlow backend (legacy models)
pip install deepchem[tensorflow]

# Full installation (all backends + extras)
pip install deepchem[all]

Quick Start

import deepchem as dc

# Load MoleculeNet dataset with featurization + scaffold split
tasks, datasets, transformers = dc.molnet.load_delaney(featurizer="ECFP")
train, valid, test = datasets

# Train and evaluate a multitask regressor
model = dc.models.MultitaskRegressor(n_tasks=1, n_features=1024, dropouts=0.2)
model.fit(train, nb_epoch=50)
metric = dc.metrics.Metric(dc.metrics.pearson_r2_score)
print(f"Test R2: {model.evaluate(test, [metric])}")  # {'pearson_r2_score': ~0.7}

Core API

Module 1: Data Loading and Processing

Load molecular data from CSV files or MoleculeNet benchmark datasets.

import deepchem as dc

# Load from CSV (SMILES + property columns)
loader = dc.data.CSVLoader(
    tasks=["measured_log_solubility"],
    feature_field="smiles",
    featurizer=dc.feat.CircularFingerprint(size=2048, radius=3)
)
dataset = loader.create_dataset("solubility_data.csv")
print(f"Samples: {dataset.X.shape[0]}, Features: {dataset.X.shape[1]}")
# Samples: 1128, Features: 2048

# Load from SDF (3D structures)
sdf_loader = dc.data.SDFLoader(
    tasks=["activity"],
    featurizer=dc.feat.CoulombMatrix(max_atoms=50)
)
dataset_3d = sdf_loader.create_dataset("molecules.sdf")

# Load MoleculeNet benchmark datasets (auto-download + featurize + split)
# Available: load_delaney, load_bbbp, load_tox21, load_hiv, load_qm7, load_qm9, etc.
tasks, datasets, transformers = dc.molnet.load_tox21(featurizer="ECFP", splitter="scaffold")
train, valid, test = datasets
print(f"Tasks: {len(tasks)}, Train: {len(train)}, Test: {len(test)}")
# Tasks: 12, Train: ~6264, Test: ~631

# Inverse-transform predictions back to original scale
y_pred = model.predict(test)
y_original = transformers[0].untransform(y_pred)

Module 2: Molecular Featurization

Convert molecules to numerical representations for ML. DeepChem provides 50+ featurizers spanning fingerprints, descriptors, graph features, and Coulomb matrices.

import deepchem as dc

smiles = ["CCO", "CC(=O)O", "c1ccccc1", "CC(C)O"]

# Fingerprints (most common for classical ML)
ecfp = dc.feat.CircularFingerprint(size=2048, radius=3)
fp_features = ecfp.featurize(smiles)
print(f"ECFP shape: {fp_features.shape}")  # (4, 2048)

# RDKit descriptors (interpretable physicochemical properties)
rdkit_desc = dc.feat.RDKitDescriptors()
desc_features = rdkit_desc.featurize(smiles)
print(f"Descriptor shape: {desc_features.shape}")  # (4, 208)

# Graph features (for GNN models — returns ConvMol objects)
graph_feat = dc.feat.ConvMolFeaturizer()
graphs = graph_feat.featurize(smiles)
print(f"Atoms in first mol: {graphs[0].get_num_atoms()}")  # 3

# Mol2Vec embeddings (pretrained word2vec on molecular substructures)
mol2vec = dc.feat.Mol2VecFingerprint()
embeddings = mol2vec.featurize(smiles)
print(f"Mol2Vec shape: {embeddings.shape}")  # (4, 300)

Module 3: Model Training and Evaluation

DeepChem provides MultitaskRegressor and MultitaskClassifier as general-purpose models, plus specialized architectures for graph and sequence data.

import deepchem as dc

# Load dataset
tasks, datasets, transformers = dc.molnet.load_delaney(featurizer="ECFP")
train, valid, test = datasets

# Regression model (fingerprint input)
model = dc.models.MultitaskRegressor(
    n_tasks=1,
    n_features=1024,
    layer_sizes=[1000, 500],
    dropouts=0.25,
    learning_rate=0.001,
    batch_size=64,
)
model.fit(train, nb_epoch=100)

# Evaluate with multiple metrics
metrics = [
    dc.metrics.Metric(dc.metrics.pearson_r2_score),
    dc.metrics.Metric(dc.metrics.mean_absolute_error),
    dc.metrics.Metric(dc.metrics.rms_score),
]
results = model.evaluate(test, metrics)
print(f"R2: {results['pearson_r2_score']:.3f}, MAE: {results['mean_absolute_error']:.3f}")

# Classification model (e.g., Tox21 toxicity prediction)
tasks, datasets, transformers = dc.molnet.load_tox21(featurizer="ECFP")
train, valid, test = datasets

clf = dc.models.MultitaskClassifier(
    n_tasks=len(tasks),
    n_features=1024,
    layer_sizes=[1000, 500],
    dropouts=0.5,
    learning_rate=0.001,
)
clf.fit(train, nb_epoch=50)
roc_metric = dc.metrics.Metric(dc.metrics.roc_auc_score, np.mean)
print(f"Mean ROC-AUC: {clf.evaluate(test, [roc_metric])}")

Module 4: Graph Neural Networks

GNNs operate directly on molecular graphs (atoms as nodes, bonds as edges), avoiding information loss from fixed fingerprints.

import deepchem as dc

# Load with graph featurizer
tasks, datasets, transformers = dc.molnet.load_delaney(featurizer="GraphConv")
train, valid, test = datasets

# Graph Convolutional Network (Duvenaud et al.)
gcn_model = dc.models.GraphConvModel(
    n_tasks=1,
    mode="regression",
    graph_conv_layers=[64, 64],
    dense_layer_size=256,
    dropout=0.2,
    learning_rate=0.001,
    batch_size=64,
)
gcn_model.fit(train, nb_epoch=100)
metric = dc.metrics.Metric(dc.metrics.pearson_r2_score)
print(f"GCN R2: {gcn_model.evaluate(test, [metric])}")

# AttentiveFP (Xiong et al.) — attention-based GNN, strong on molecular properties
tasks, datasets, transformers = dc.molnet.load_delaney(
    featurizer=dc.feat.MolGraphConvFeaturizer(use_edges=True)
)
train, valid, test = datasets

attfp_model = dc.models.AttentiveFPModel(
    n_tasks=1,
    mode="regression",
    num_layers=2,
    graph_feat_size=200,
    num_timesteps=2,
    dropout=0.2,
    learning_rate=0.001,
    batch_size=64,
)
attfp_model.fit(train, nb_epoch=100)
print(f"AttentiveFP R2: {attfp_model.evaluate(test, [metric])}")

Module 5: Transfer Learning

Fine-tune pretrained chemical language models for downstream tasks with limited data.

import deepchem as dc
from deepchem.models.torch_models import ChemBERTaModel

# ChemBERTa — SMILES-based transformer (pretrained on 77M molecules)
tasks, datasets, transformers = dc.molnet.load_bbbp(featurizer=dc.feat.SmilesTokenizer())
train, valid, test = datasets

chemberta = ChemBERTaModel(
    task="classification",
    n_tasks=1,
    model_dir="chem