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="chemberta_finetuned/",
)
# Fine-tune on downstream task (BBB permeability)
chemberta.fit(train, nb_epoch=10)
metric = dc.metrics.Metric(dc.metrics.roc_auc_score)
print(f"ChemBERTa ROC-AUC: {chemberta.evaluate(test, [metric])}")

Module 6: Predictions on New Molecules

Run inference on new molecules with a trained model.

import deepchem as dc
import numpy as np

# Assume trained model from Module 3
# Featurize new molecules using same featurizer
featurizer = dc.feat.CircularFingerprint(size=1024, radius=2)
new_smiles = ["c1cc(O)ccc1", "CC(=O)Nc1ccc(O)cc1", "OC(=O)c1ccccc1"]
new_features = featurizer.featurize(new_smiles)
new_dataset = dc.data.NumpyDataset(X=new_features)

predictions = model.predict(new_dataset)
for smi, pred in zip(new_smiles, predictions):
    print(f"{smi}: {pred[0]:.2f}")

# Ensemble predictions from multiple models for robustness
models = [model1, model2, model3]  # trained models
all_preds = np.array([m.predict(new_dataset) for m in models])
ensemble_mean = all_preds.mean(axis=0)
ensemble_std = all_preds.std(axis=0)
print(f"Ensemble prediction: {ensemble_mean[0][0]:.2f} +/- {ensemble_std[0][0]:.2f}")

Key Concepts

Unified API Pattern

All DeepChem workflows follow a consistent 5-step pattern:

Load Data → Featurize → Split → Train → Evaluate

• Load: CSVLoader, SDFLoader, or dc.molnet.load_*() (auto-loads MoleculeNet datasets)
• Featurize: Pass featurizer to loader, or call featurizer.featurize(smiles) directly
• Split: ScaffoldSplitter (recommended for drug discovery), RandomSplitter, ButinaSplitter
• Train: model.fit(train_dataset, nb_epoch=N)
• Evaluate: model.evaluate(test_dataset, metrics_list)

Model Selection Guide

Data Type	Model	Key Feature	Use When
SMILES + fingerprints	MultitaskRegressor	Fast, baseline	First attempt, small datasets
SMILES + fingerprints	MultitaskClassifier	Multi-label	Multi-task classification (Tox21)
Molecular graphs	GraphConvModel	Learned fingerprints	Medium datasets, general properties
Molecular graphs	GATModel	Attention mechanism	When atom importance matters
Molecular graphs	AttentiveFPModel	Graph + timestep attention	State-of-art molecular properties
Molecular graphs	MPNNModel	Message passing	Complex molecular interactions
Molecular graphs	DMPNNModel	Directed MPNN	Bond-level predictions
SMILES strings	ChemBERTaModel	Pretrained transformer	Low-data regime, transfer learning
SMILES strings	GROVERModel	Graph + transformer	Rich molecular representations
Crystal structures	CGCNNModel	Crystal graph CNN	Materials property prediction
Crystal structures	MEGNetModel	Graph networks	Materials and molecules
Protein sequences	ProteinLigandComplexModel	Complex modeling	Binding affinity prediction
Tabular features	XGBoostModel, RandomForestModel	Classical ML	Interpretability, baselines

Featurizer Selection Guide

Featurizer	Class	Output	Best For
ECFP/Morgan	CircularFingerprint	Binary vector (1024-2048)	General QSAR, fast baselines
MACCS Keys	MACCSKeysFingerprint	167-bit vector	Substructure filtering
RDKit 2D	RDKitDescriptors	200+ descriptors	Interpretable models
Mol2Vec	Mol2VecFingerprint	300-dim embedding	Similarity, clustering
ConvMol	ConvMolFeaturizer	Graph features	GraphConvModel input
MolGraph	MolGraphConvFeaturizer	Node + edge features	AttentiveFPModel, MPNNModel
Weave	WeaveFeaturizer	Pair features	WeaveModel input
Coulomb Matrix	CoulombMatrix	Atom-pair distances	QM property prediction
SMILES tokens	SmilesTokenizer	Token IDs	ChemBERTa, transformer models

Data Splitting Strategies

Splitter	Use Case	Why
ScaffoldSplitter	Drug discovery (default)	Tests generalization to new chemotypes
RandomSplitter	Quick experiments	Baseline, but overestimates performance
ButinaSplitter	Diversity-based	Clusters by Tanimoto similarity
FingerprintSplitter	Chemical similarity	Groups structurally similar molecules
MaxMinSplitter	Maximum diversity test	Extreme generalization test

Common Workflows

Workflow 1: QSAR from CSV Data

Goal: Build a property prediction model from a CSV file with SMILES and activity columns.

import deepchem as dc
import pandas as pd

# Step 1: Load and featurize CSV data
loader = dc.data.CSVLoader(
    tasks=["pIC50"],
    feature_field="smiles",
    featurizer=dc.feat.CircularFingerprint(size=2048, radius=3),
)
dataset = loader.create_dataset("bioactivity_data.csv")

# Step 2: Normalize targets
transformer = dc.trans.NormalizationTransformer(
    transform_y=True, dataset=dataset
)
dataset = transformer.transform(dataset)

# Step 3: Scaffold split (realistic for drug discovery)
splitter = dc.splits.ScaffoldSplitter()
train, valid, test = splitter.train_valid_test_split(dataset)
print(f"Train: {len(train)}, Valid: {len(valid)}, Test: {len(test)}")

# Step 4: Train model
model = dc.models.MultitaskRegressor(
    n_tasks=1, n_features=2048,
    layer_sizes=[1000, 500], dropouts=0.25,
    learning_rate=0.001, batch_size=64,
)
model.fit(train, nb_epoch=100)

# Step 5: Evaluate
metrics = [
    dc.metrics.Metric(dc.metrics.pearson_r2_score),
    dc.metrics.Metric(dc.metrics.mean_absolute_error),
]
results = model.evaluate(test, metrics)
print(f"R2: {results['pearson_r2_score']:.3f}, MAE: {results['mean_absolute_error']:.3f}")

Workflow 2: MoleculeNet Benchmark Comparison

Goal: Compare multiple models on a MoleculeNet benchmark dataset.

import deepchem as dc

# Load dataset with graph featurizer (supports both fingerprint and GNN models)
tasks, datasets, transformers = dc.molnet.load_bbbp(
    featurizer="GraphConv", splitter="scaffold"
)
train, valid, test = datasets
metric = dc.metrics.Metric(dc.metrics.roc_auc_score)

# Model 1: Graph Convolutional Network
gcn = dc.models.GraphConvModel(n_tasks=1, mode="classification", dropout=0.2)
gcn.fit(train, nb_epoch=50)
gcn_score = gcn.evaluate(test, [metric])

# Model 2: Random Forest baseline (needs fingerprints)
tasks_fp, datasets_fp, _ = dc.molnet.load_bbbp(featurizer="ECFP", splitter="scaffold")
train_fp, _, test_fp = datasets_fp
rf = dc.models.SklearnModel(
    model=dc.models.sklearn_models.RandomForestClassifier(n_estimators=500),
    model_dir="rf_model/"
)
rf.fit(train_fp)
rf_score = rf.evaluate(test_fp, [metric])

print(f"GCN ROC-AUC: {gcn_score['roc_auc_score']:.3f}")
print(f"RF  ROC-AUC: {rf_score['roc_auc_score']:.3f}")

Workflow 3: Transfer Learning Pipeline

Goal: Fine-tune a pretrained model on a small dataset.

1. Load pretrained ChemBERTa model (see Module 5 for code)
2. Prepare downstream dataset with SmilesTokenizer featurizer
3. Fine-tune with reduced learning rate (1e-5 to 5e-5) for 5-15 epochs
4. Evaluate on held-out scaffold split — expect gains over fingerprint baselines when training data < 1000 samples
5. Save fine-tuned model: model.save_checkpoint()
6. See references/workflows_model_catalog.md Workflow 1 for complete hyperparameter optimization code

Key Parameters

Parameter	Module	Default	Range / Options	Effect
n_features	MultitaskRegressor/Classifier	Required	Matches featurizer output	Input feature dimension
layer_sizes	MultitaskRegressor/Classifier	[1000]	[256] to [1000, 500, 250]	Hidden layer dimensions
dropouts	All neural models	0.0	0.0-0.5	Regularization strength
learning_rate	All neural models	0.001	1e-5-0.01	Training step size
batch_size	All neural models	100	16-256	Samples per gradient update
nb_epoch	model.fit()	10	10-300	Training iterations
size	CircularFingerprint	2048	512-4096	Fingerprint bit length
radius	CircularFingerprint	2	2-4	Substructure neighborhood radius
graph_conv_layers	GraphConvModel	[64, 64]	[32] to [128, 128, 64]	Graph convolution widths
num_layers	AttentiveFPModel	2	1-5	GNN message passing depth
graph_feat_size	AttentiveFPModel	200	64-512	Graph feature dimension
splitter	dc.molnet.load_*()	"scaffold"	"scaffold", "random", "butina"	Data splitting strategy

Best Practices

1. Always use scaffold splitting for drug discovery: Random splits leak structural information and overestimate performance. Scaffold splits test generalization to novel chemotypes.

2. Normalize regression targets: Apply NormalizationTransformer(transform_y=True) before training. Remember to untransform() predictions for interpretable values.

3. Start with fingerprint baselines: Train MultitaskRegressor + ECFP first. Only move to GNNs if fingerprint baseline is insufficient — GNNs need more data and compute.

   # Baseline first
   baseline = dc.models.MultitaskRegressor(n_tasks=1, n_features=2048)
   

4. Match featurizer to model: GNN models require graph featurizers (ConvMolFeaturizer, MolGraphConvFeaturizer). Fingerprint models need CircularFingerprint. Mixing causes silent errors.

5. Anti-pattern -- Do not use random split for drug discovery benchmarks: Results with RandomSplitter are not publishable for molecular property prediction. Reviewers expect scaffold or temporal splits.

6. Handle missing labels in multi-task datasets: Tox21 and many bioactivity datasets have missing values. DeepChem handles NaN labels automatically during training (masked loss), but verify with np.isnan(dataset.y).sum().

7. Use early stopping via validation set: Monitor validation loss to prevent overfitting, especially with GNN models.

Common Recipes

Recipe: Hyperparameter Search

When to use: Optimize model performance before final evaluation.

import deepchem as dc

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

# Define parameter grid
params = {
    "n_features": [1024],
    "layer_sizes": [[500], [1000, 500], [1000, 500, 250]],
    "dropouts": [0.1, 0.25, 0.5],
    "learning_rate": [0.001, 0.0005],
}

optimizer = dc.hyper.GridHyperparamOpt(lambda **p: dc.models.MultitaskRegressor(**p))
metric = dc.metrics.Metric(dc.metrics.pearson_r2_score)
best_model, best_params, all_results = optimizer.hyperparam_search(
    params, train, valid, metric, logdir="hyperparam_logs/"
)
print(f"Best params: {best_params}")
print(f"Best R2: {best_model.evaluate(test, [metric])}")

Recipe: Save and Reload Models

When to use: Deploy trained models or resume training.

# Save model checkpoint
model.save_checkpoint(model_dir="saved_model/")

# Reload model
loaded_model = dc.models.MultitaskRegressor(n_tasks=1, n_features=2048)
loaded_model.restore(model_dir="saved_model/")
predictions = loaded_model.predict(test)

Recipe: Custom Metric

When to use: Evaluate models with domain-specific metrics.

import deepchem as dc
import numpy as np

def enrichment_factor(y_true, y_pred, top_fraction=0.01):
    """Enrichment factor at top X% of ranked predictions."""
    n = len(y_true)
    n_top = max(int(n * top_fraction), 1)
    top_indices = np.argsort(y_pred.flatten())[-n_top:]
    hits_in_top = y_true.flatten()[top_indices].sum()
    expected = y_true.sum() * top_fraction
    return hits_in_top / expected if expected > 0 else 0.0

ef_metric = dc.metrics.Metric(enrichment_factor, mode="regression")
print(f"EF@1%: {model.evaluate(test, [ef_metric])}")

Troubleshooting

Problem	Cause	Solution
ModuleNotFoundError: torch	PyTorch not installed	pip install deepchem[torch] for GNN models
ValueError: n_features mismatch	Featurizer output size does not match model n_features	Check dataset.X.shape[1] and set n_features accordingly
NaN loss during training	Learning rate too high or unnormalized targets	Apply NormalizationTransformer, reduce learning rate to 1e-4
Low scaffold-split performance	Model memorizes scaffolds, not properties	Use more data, try GNN models, or add regularization (dropout 0.3-0.5)
RuntimeError: CUDA out of memory	Batch size too large for GPU	Reduce batch_size (32 or 16), or use CPU for small datasets
FeaturizationError on some SMILES	Invalid or complex SMILES strings	Pre-filter with RDKit: Chem.MolFromSmiles(smi) is not None
Model predicts constant values	Targets not normalized or too few epochs	Apply NormalizationTransformer, increase nb_epoch
Slow featurization	Large dataset with expensive featurizer	Use CircularFingerprint (fast) or parallelize with n_jobs parameter

Bundled Resources

• references/workflows_model_catalog.md -- Extended workflows (hyperparameter optimization with full code, MolGAN generative models, materials property prediction with CGCNN/MEGNet, protein-ligand modeling, custom model architecture) plus complete model catalog (60+ models organized by category) and complete featurizer catalog (50+ featurizers). Covers: workflows 4-8 from original, extended model and featurizer inventories, MoleculeNet dataset catalog. Relocated inline: top 3 workflows (QSAR, MoleculeNet benchmark, transfer learning) are in Common Workflows; core model/featurizer tables are in Key Concepts. Omitted: detailed installation troubleshooting for TensorFlow 1.x (deprecated) and Docker-specific setup (covered by official docs).

Related Skills

• rdkit-cheminformatics -- molecular manipulation, fingerprints, substructure search (upstream featurization)
• molfeat-molecular-featurization -- 100+ featurizers with scikit-learn API (featurization-only alternative)
• datamol-cheminformatics -- Pythonic molecular processing (upstream data prep)
• pytdc-therapeutics-data-commons -- curated ADMET/DTI datasets with standardized splits (complementary data source)
• torch-geometric-graph-neural-networks -- lower-level PyG for custom GNN architectures (alternative for advanced users)
• scikit-learn-machine-learning -- classical ML baselines that DeepChem wraps via SklearnModel

References

• DeepChem documentation -- official API docs and tutorials
• DeepChem GitHub -- source code, examples, issues
• MoleculeNet benchmark paper -- Wu et al. 2018, benchmark dataset descriptions
• DeepChem tutorials -- Jupyter notebook tutorials