pytdc

PyTDC (Therapeutics Data Commons)

Overview

PyTDC is an open-science platform providing AI-ready datasets and benchmarks for drug discovery and development. Access curated datasets spanning the entire therapeutics pipeline with standardized evaluation metrics and meaningful data splits, organized into three categories: single-instance prediction (molecular/protein properties), multi-instance prediction (drug-target interactions, DDI), and generation (molecule generation, retrosynthesis).

When to Use This Skill

This skill should be used when:

• Working with drug discovery or therapeutic ML datasets
• Benchmarking machine learning models on standardized pharmaceutical tasks
• Predicting molecular properties (ADME, toxicity, bioactivity)
• Predicting drug-target or drug-drug interactions
• Generating novel molecules with desired properties
• Accessing curated datasets with proper train/test splits (scaffold, cold-split)
• Using molecular oracles for property optimization

Installation & Setup

Install PyTDC using pip:

uv pip install PyTDC

To upgrade to the latest version:

uv pip install PyTDC --upgrade

Core dependencies (automatically installed):

• numpy, pandas, tqdm, seaborn, scikit_learn, fuzzywuzzy

Additional packages are installed automatically as needed for specific features.

Quick Start

The basic pattern for accessing any TDC dataset follows this structure:

from tdc.<problem> import <Task>
data = <Task>(name='<Dataset>')
split = data.get_split(method='scaffold', seed=1, frac=[0.7, 0.1, 0.2])
df = data.get_data(format='df')

Where:

• <problem>: One of single_pred, multi_pred, or generation
• <Task>: Specific task category (e.g., ADME, DTI, MolGen)
• <Dataset>: Dataset name within that task

Example - Loading ADME data:

from tdc.single_pred import ADME
data = ADME(name='Caco2_Wang')
split = data.get_split(method='scaffold')
# Returns dict with 'train', 'valid', 'test' DataFrames

Single-Instance Prediction Tasks

Single-instance prediction involves forecasting properties of individual biomedical entities (molecules, proteins, etc.).

Available Task Categories

1. ADME (Absorption, Distribution, Metabolism, Excretion)

Predict pharmacokinetic properties of drug molecules.

from tdc.single_pred import ADME
data = ADME(name='Caco2_Wang')  # Intestinal permeability
# Other datasets: HIA_Hou, Bioavailability_Ma, Lipophilicity_AstraZeneca, etc.

Common ADME datasets:

• Caco2 - Intestinal permeability
• HIA - Human intestinal absorption
• Bioavailability - Oral bioavailability
• Lipophilicity - Octanol-water partition coefficient
• Solubility - Aqueous solubility
• BBB - Blood-brain barrier penetration
• CYP - Cytochrome P450 metabolism

2. Toxicity (Tox)

Predict toxicity and adverse effects of compounds.

from tdc.single_pred import Tox
data = Tox(name='hERG')  # Cardiotoxicity
# Other datasets: AMES, DILI, Carcinogens_Lagunin, etc.

Common toxicity datasets:

• hERG - Cardiac toxicity
• AMES - Mutagenicity
• DILI - Drug-induced liver injury
• Carcinogens - Carcinogenicity
• ClinTox - Clinical trial toxicity

3. HTS (High-Throughput Screening)

Bioactivity predictions from screening data.

from tdc.single_pred import HTS
data = HTS(name='SARSCoV2_Vitro_Touret')

4. QM (Quantum Mechanics)

Quantum mechanical properties of molecules.

from tdc.single_pred import QM
data = QM(name='QM7')

5. Other Single Prediction Tasks

• Yields: Chemical reaction yield prediction
• Epitope: Epitope prediction for biologics
• Develop: Development-stage predictions
• CRISPROutcome: Gene editing outcome prediction

Data Format

Single prediction datasets typically return DataFrames with columns:

• Drug_ID or Compound_ID: Unique identifier
• Drug or X: SMILES string or molecular representation
• Y: Target label (continuous or binary)

Multi-Instance Prediction Tasks

Multi-instance prediction involves forecasting properties of interactions between multiple biomedical entities.

Available Task Categories

1. DTI (Drug-Target Interaction)

Predict binding affinity between drugs and protein targets.

from tdc.multi_pred import DTI
data = DTI(name='BindingDB_Kd')
split = data.get_split()

Available datasets:

• BindingDB_Kd - Dissociation constant (52,284 pairs)
• BindingDB_IC50 - Half-maximal inhibitory concentration (991,486 pairs)
• BindingDB_Ki - Inhibition constant (375,032 pairs)
• DAVIS, KIBA - Kinase binding datasets

Data format: Drug_ID, Target_ID, Drug (SMILES), Target (sequence), Y (binding affinity)

2. DDI (Drug-Drug Interaction)

Predict interactions between drug pairs.

from tdc.multi_pred import DDI
data = DDI(name='DrugBank')
split = data.get_split()

Multi-class classification task predicting interaction types. Dataset contains 191,808 DDI pairs with 1,706 drugs.

3. PPI (Protein-Protein Interaction)

Predict protein-protein interactions.

from tdc.multi_pred import PPI
data = PPI(name='HuRI')

4. Other Multi-Prediction Tasks

• GDA: Gene-disease associations
• DrugRes: Drug resistance prediction
• DrugSyn: Drug synergy prediction
• PeptideMHC: Peptide-MHC binding
• AntibodyAff: Antibody affinity prediction
• MTI: miRNA-target interactions
• Catalyst: Catalyst prediction
• TrialOutcome: Clinical trial outcome prediction

Generation Tasks

Generation tasks involve creating novel biomedical entities with desired properties.

1. Molecular Generation (MolGen)

Generate diverse, novel molecules with desirable chemical properties.

from tdc.generation import MolGen
data = MolGen(name='ChEMBL_V29')
split = data.get_split()

Use with oracles to optimize for specific properties:

from tdc import Oracle
oracle = Oracle(name='GSK3B')
score = oracle('CC(C)Cc1ccc(cc1)C(C)C(O)=O')  # Evaluate SMILES

See references/oracles.md for all available oracle functions.

2. Retrosynthesis (RetroSyn)

Predict reactants needed to synthesize a target molecule.

from tdc.generation import RetroSyn
data = RetroSyn(name='USPTO')
split = data.get_split()

Dataset contains 1,939,253 reactions from USPTO database.

3. Paired Molecule Generation

Generate molecule pairs (e.g., prodrug-drug pairs).

from tdc.generation import PairMolGen
data = PairMolGen(name='Prodrug')

For detailed oracle documentation and molecular generation workflows, refer to references/oracles.md and scripts/molecular_generation.py.

Benchmark Groups

Benchmark groups provide curated collections of related datasets for systematic model evaluation.

ADMET Benchmark Group

from tdc.benchmark_group import admet_group
group = admet_group(path='data/')

# Get benchmark datasets
benchmark = group.get('Caco2_Wang')
predictions = {}

for seed in [1, 2, 3, 4, 5]:
    train, valid = benchmark['train'], benchmark['valid']
    # Train model here
    predictions[seed] = model.predict(benchmark['test'])

# Evaluate with required 5 seeds
results = group.evaluate(predictions)

ADMET Group includes 22 datasets covering absorption, distribution, metabolism, excretion, and toxicity.

Other Benchmark Groups

Available benchmark groups include collections for:

• ADMET properties
• Drug-target interactions
• Drug combination prediction
• And more specialized therapeutic tasks

For benchmark evaluation workflows, see scripts/benchmark_evaluation.py.

Data Functions

TDC provides comprehensive data processing utilities organized into four categories.

1. Dataset Splits

Retrieve train/validation/test partitions with various strategies:

# Scaffold split (default for most tasks)
split = data.get_split(method='scaffold', seed=1, frac=[0.7, 0.1, 0.2])

# Random split
split = data.get_split(method='random', seed=42, frac=[0.8, 0.1, 0.1])

# Cold split (for DTI/DDI tasks)
split = data.get_split(method='cold_drug', seed=1)  # Unseen drugs in test
split = data.get_split(method='cold_target', seed=1)  # Unseen targets in test

Available split strategies:

• random: Random shuffling
• scaffold: Scaffold-based (for chemical diversity)
• cold_drug, cold_target, cold_drug_target: For DTI tasks
• temporal: Time-based splits for temporal datasets

2. Model Evaluation

Use standardized metrics for evaluation:

from tdc import Evaluator

# For binary classification
evaluator = Evaluator(name='ROC-AUC')
score = evaluator(y_true, y_pred)

# For regression
evaluator = Evaluator(name='RMSE')
score = evaluator(y_true, y_pred)

Available metrics: ROC-AUC, PR-AUC, F1, Accuracy, RMSE, MAE, R2, Spearman, Pearson, and more.

3. Data Processing

TDC provides 11 key processing utilities:

from tdc.chem_utils import MolConvert

# Molecule format conversion
converter = MolConvert(src='SMILES', dst='PyG')
pyg_graph = converter('CC(C)Cc1ccc(cc1)C(C)C(O)=O')

Processing utilities include:

• Molecule format conversion (SMILES, SELFIES, PyG, DGL, ECFP, etc.)
• Molecule filters (PAINS, drug-likeness)
• Label binarization and unit conversion
• Data balancing (over/under-sampling)
• Negative sampling for pair data
• Graph transformation
• Entity retrieval (CID to SMILES, UniProt to sequence)

For comprehensive utilities documentation, see references/utilities.md.

4. Molecule Generation Oracles

TDC provides 17+ oracle functions for molecular optimization:

from tdc import Oracle

# Single oracle
oracle = Oracle(name='DRD2')
score = oracle('CC(C)Cc1ccc(cc1)C(C)C(O)=O')

# Multiple oracles
oracle = Oracle(name='JNK3')
scores = oracle(['SMILES1', 'SMILES2', 'SMILES3'])

For complete oracle documentation, see references/oracles.md.

Advanced Features

Retrieve Available Datasets

from tdc.utils imp

---

*Content truncated.*