DeepChem

Overview

DeepChem is a comprehensive Python library for applying machine learning to chemistry, materials science, and biology. Enable molecular property prediction, drug discovery, materials design, and biomolecule analysis through specialized neural networks, molecular featurization methods, and pretrained models.

When to Use This Skill

This skill should be used when:

• Loading and processing molecular data (SMILES strings, SDF files, protein sequences)
• Predicting molecular properties (solubility, toxicity, binding affinity, ADMET properties)
• Training models on chemical/biological datasets
• Using MoleculeNet benchmark datasets (Tox21, BBBP, Delaney, etc.)
• Converting molecules to ML-ready features (fingerprints, graph representations, descriptors)
• Implementing graph neural networks for molecules (GCN, GAT, MPNN, AttentiveFP)
• Applying transfer learning with pretrained models (ChemBERTa, GROVER, MolFormer)
• Predicting crystal/materials properties (bandgap, formation energy)
• Analyzing protein or DNA sequences

Core Capabilities

1. Molecular Data Loading and Processing

DeepChem provides specialized loaders for various chemical data formats:

import deepchem as dc

# Load CSV with SMILES
featurizer = dc.feat.CircularFingerprint(radius=2, size=2048)
loader = dc.data.CSVLoader(
    tasks=['solubility', 'toxicity'],
    feature_field='smiles',
    featurizer=featurizer
)
dataset = loader.create_dataset('molecules.csv')

# Load SDF files
loader = dc.data.SDFLoader(tasks=['activity'], featurizer=featurizer)
dataset = loader.create_dataset('compounds.sdf')

# Load protein sequences
loader = dc.data.FASTALoader()
dataset = loader.create_dataset('proteins.fasta')

Key Loaders:

• CSVLoader: Tabular data with molecular identifiers
• SDFLoader: Molecular structure files
• FASTALoader: Protein/DNA sequences
• ImageLoader: Molecular images
• JsonLoader: JSON-formatted datasets

2. Molecular Featurization

Convert molecules into numerical representations for ML models.

Decision Tree for Featurizer Selection

Is the model a graph neural network?
├─ YES → Use graph featurizers
│   ├─ Standard GNN → MolGraphConvFeaturizer
│   ├─ Message passing → DMPNNFeaturizer
│   └─ Pretrained → GroverFeaturizer
│
└─ NO → What type of model?
    ├─ Traditional ML (RF, XGBoost, SVM)
    │   ├─ Fast baseline → CircularFingerprint (ECFP)
    │   ├─ Interpretable → RDKitDescriptors
    │   └─ Maximum coverage → MordredDescriptors
    │
    ├─ Deep learning (non-graph)
    │   ├─ Dense networks → CircularFingerprint
    │   └─ CNN → SmilesToImage
    │
    ├─ Sequence models (LSTM, Transformer)
    │   └─ SmilesToSeq
    │
    └─ 3D structure analysis
        └─ CoulombMatrix

Example Featurization

# Fingerprints (for traditional ML)
fp = dc.feat.CircularFingerprint(radius=2, size=2048)

# Descriptors (for interpretable models)
desc = dc.feat.RDKitDescriptors()

# Graph features (for GNNs)
graph_feat = dc.feat.MolGraphConvFeaturizer()

# Apply featurization
features = fp.featurize(['CCO', 'c1ccccc1'])

Selection Guide:

• Small datasets (<1K): CircularFingerprint or RDKitDescriptors
• Medium datasets (1K-100K): CircularFingerprint or graph featurizers
• Large datasets (>100K): Graph featurizers (MolGraphConvFeaturizer, DMPNNFeaturizer)
• Transfer learning: Pretrained model featurizers (GroverFeaturizer)

See references/api_reference.md for complete featurizer documentation.

3. Data Splitting

Critical: For drug discovery tasks, use ScaffoldSplitter to prevent data leakage from similar molecular structures appearing in both training and test sets.

# Scaffold splitting (recommended for molecules)
splitter = dc.splits.ScaffoldSplitter()
train, valid, test = splitter.train_valid_test_split(
    dataset,
    frac_train=0.8,
    frac_valid=0.1,
    frac_test=0.1
)

# Random splitting (for non-molecular data)
splitter = dc.splits.RandomSplitter()
train, test = splitter.train_test_split(dataset)

# Stratified splitting (for imbalanced classification)
splitter = dc.splits.RandomStratifiedSplitter()
train, test = splitter.train_test_split(dataset)

Available Splitters:

• ScaffoldSplitter: Split by molecular scaffolds (prevents leakage)
• ButinaSplitter: Clustering-based molecular splitting
• MaxMinSplitter: Maximize diversity between sets
• RandomSplitter: Random splitting
• RandomStratifiedSplitter: Preserves class distributions

4. Model Selection and Training

Quick Model Selection Guide

Dataset Size	Task	Recommended Model	Featurizer
< 1K samples	Any	SklearnModel (RandomForest)	CircularFingerprint
1K-100K	Classification/Regression	GBDTModel or MultitaskRegressor	CircularFingerprint
> 100K	Molecular properties	GCNModel, AttentiveFPModel, DMPNNModel	MolGraphConvFeaturizer
Any (small preferred)	Transfer learning	ChemBERTa, GROVER, MolFormer	Model-specific
Crystal structures	Materials properties	CGCNNModel, MEGNetModel	Structure-based
Protein sequences	Protein properties	ProtBERT	Sequence-based

Example: Traditional ML

from sklearn.ensemble import RandomForestRegressor

# Wrap scikit-learn model
sklearn_model = RandomForestRegressor(n_estimators=100)
model = dc.models.SklearnModel(model=sklearn_model)
model.fit(train)

Example: Deep Learning

# Multitask regressor (for fingerprints)
model = dc.models.MultitaskRegressor(
    n_tasks=2,
    n_features=2048,
    layer_sizes=[1000, 500],
    dropouts=0.25,
    learning_rate=0.001
)
model.fit(train, nb_epoch=50)

Example: Graph Neural Networks

# Graph Convolutional Network
model = dc.models.GCNModel(
    n_tasks=1,
    mode='regression',
    batch_size=128,
    learning_rate=0.001
)
model.fit(train, nb_epoch=50)

# Graph Attention Network
model = dc.models.GATModel(n_tasks=1, mode='classification')
model.fit(train, nb_epoch=50)

# Attentive Fingerprint
model = dc.models.AttentiveFPModel(n_tasks=1, mode='regression')
model.fit(train, nb_epoch=50)

5. MoleculeNet Benchmarks

Quick access to 30+ curated benchmark datasets with standardized train/valid/test splits:

# Load benchmark dataset
tasks, datasets, transformers = dc.molnet.load_tox21(
    featurizer='GraphConv',  # or 'ECFP', 'Weave', 'Raw'
    splitter='scaffold',     # or 'random', 'stratified'
    reload=False
)
train, valid, test = datasets

# Train and evaluate
model = dc.models.GCNModel(n_tasks=len(tasks), mode='classification')
model.fit(train, nb_epoch=50)

metric = dc.metrics.Metric(dc.metrics.roc_auc_score)
test_score = model.evaluate(test, [metric])

Common Datasets:

• Classification: load_tox21(), load_bbbp(), load_hiv(), load_clintox()
• Regression: load_delaney(), load_freesolv(), load_lipo()
• Quantum properties: load_qm7(), load_qm8(), load_qm9()
• Materials: load_perovskite(), load_bandgap(), load_mp_formation_energy()

See references/api_reference.md for complete dataset list.

6. Transfer Learning

Leverage pretrained models for improved performance, especially on small datasets:

# ChemBERTa (BERT pretrained on 77M molecules)
model = dc.models.HuggingFaceModel(
    model='seyonec/ChemBERTa-zinc-base-v1',
    task='classification',
    n_tasks=1,
    learning_rate=2e-5  # Lower LR for fine-tuning
)
model.fit(train, nb_epoch=10)

# GROVER (graph transformer pretrained on 10M molecules)
model = dc.models.GroverModel(
    task='regression',
    n_tasks=1
)
model.fit(train, nb_epoch=20)

When to use transfer learning:

• Small datasets (< 1000 samples)
• Novel molecular scaffolds
• Limited computational resources
• Need for rapid prototyping

Use the scripts/transfer_learning.py script for guided transfer learning workflows.

7. Model Evaluation

# Define metrics
classification_metrics = [
    dc.metrics.Metric(dc.metrics.roc_auc_score, name='ROC-AUC'),
    dc.metrics.Metric(dc.metrics.accuracy_score, name='Accuracy'),
    dc.metrics.Metric(dc.metrics.f1_score, name='F1')
]

regression_metrics = [
    dc.metrics.Metric(dc.metrics.r2_score, name='R²'),
    dc.metrics.Metric(dc.metrics.mean_absolute_error, name='MAE'),
    dc.metrics.Metric(dc.metrics.root_mean_squared_error, name='RMSE')
]

# Evaluate
train_scores = model.evaluate(train, classification_metrics)
test_scores = model.evaluate(test, classification_metrics)

8. Making Predictions

# Predict on test set
predictions = model.predict(test)

# Predict on new molecules
new_smiles = ['CCO', 'c1ccccc1', 'CC(C)O']
new_features = featurizer.featurize(new_smiles)
new_dataset = dc.data.NumpyDataset(X=new_features)

# Apply same transformations as training
for transformer in transformers:
    new_dataset = transformer.transform(new_dataset)

predictions = model.predict(new_dataset)

Typical Workflows

Workflow A: Quick Benchmark Evaluation

For evaluating a model on standard benchmarks:

import deepchem as dc

# 1. Load benchmark
tasks, datasets, _ = dc.molnet.load_bbbp(
    featurizer='GraphConv',
    splitter='scaffold'
)
train, valid, test = datasets

# 2. Train model
model = dc.models.GCNModel(n_tasks=len(tasks), mode='classification')
model.fit(train, nb_epoch=50)

# 3. Evaluate
metric = dc.metrics.Metric(dc.metrics.roc_auc_score)
test_score = model.evaluate(test, [metric])
print(f"Test ROC-AUC: {test_score}")

Workflow B: Custom Data Prediction

For training on custom molecular datasets:

import deepchem as dc

# 1. Load and featurize data
featurizer = dc.feat.CircularFingerprint(radius=2, size=2048)
loader = dc.data.CSVLoader(
    tasks=['activity'],
    feature_field='smiles',
    featurizer=featurizer
)
dataset = loader.create_dataset('my_molecules.csv')

# 2. Split data (use ScaffoldSplitter for molecules!)
splitter = dc.splits.ScaffoldSplitter()
train, valid, test = splitter.train_valid_test_split(dataset)

# 3. Normalize (optional but recommended)
transformers = [dc.trans.NormalizationTransformer(
    transform_y=True, dataset=train
)]
for transformer in transformers:
    train = transformer.transform(train)
    valid = transformer.transform(valid)
    test = transformer.transform(test)

# 4. Train model
model = dc.models.MultitaskRegressor(
    n_tasks=1,
    n_features=2048,
    layer_sizes=[1000, 500],
    dropouts=0.25
)
model.fit(train, nb_epoch=50)

# 5. Evaluate
metric = dc.metrics.Metric(dc.metrics.r2_score)
test_score = model.evaluate(test, [metric])

Workflow C: Transfer Learning on Small Dataset

For leveraging pretrained models:

import deepchem as dc

# 1. Load data (pretrained models often need raw SMILES)
loader = dc.data.CSVLoader(
    tasks=['activity'],
    feature_field='smiles',
    featurizer=dc.feat.DummyFeaturizer()  # Model handles featurization
)
dataset = loader.create_dataset('small_dataset.csv')

# 2. Split data
splitter = dc.splits.ScaffoldSplitter()
train, test = splitter.train_test_split(dataset)

# 3. Load pretrained model
model = dc.models.HuggingFaceModel(
    model='seyonec/ChemBERTa-zinc-base-v1',
    task='classification',
    n_tasks=1,
    learning_rate=2e-5
)

# 4. Fine-tune
model.fit(train, nb_epoch=10)

# 5. Evaluate
predictions = model.predict(test)

See references/workflows.md for 8 detailed workflow examples covering molecular generation, materials science, protein analysis, and more.

Example Scripts

This skill includes three production-ready scripts in the scripts/ directory:

1. predict_solubility.py

Train and evaluate solubility prediction models. Works with Delaney benchmark or custom CSV data.

```bash

Use Delaney benchmark

python scripts/predict_solubility.py

Use custom data

python scripts/predict_solubility.py \

--data my_data.csv \