pytdc

PyTDC - Therapeutics Data Commons Drug Discovery Dataset Skills

Skill Overview

PyTDC is an open-source machine learning data platform for drug discovery. It provides AI-ready drug discovery datasets and standardized benchmarking tools, covering end-to-end drug discovery tasks such as ADME property prediction, toxicity assessment, drug–target interaction prediction, and molecular generation.

Use Cases

1. Drug Discovery and Machine Learning Research

When you need high-quality, standardized training data for drug discovery projects, PyTDC offers well-curated drug discovery datasets. These include multiple task types such as molecular property prediction, drug–target interaction, and drug–drug interaction. Each dataset comes with recommended train/validation/test splitting strategies.

2. Drug Property Prediction and Model Benchmarking

When evaluating the performance of machine learning models on drug discovery tasks, PyTDC provides professional evaluation suites such as ADMET benchmark sets. These include 22 datasets covering absorption, distribution, metabolism, excretion, and toxicity. It also supports 5 random-seed evaluation protocols to ensure scientific validity and reproducibility of model assessments.

3. Molecular Generation and Drug Design

For target-directed molecular generation and drug optimization, PyTDC provides 17+ molecular property predictors (oracles). These can evaluate key metrics of generated molecules in real time, such as drug activity, selectivity, and ADMET properties. Together with large-scale molecular datasets like ChEMBL, PyTDC enables efficient screening and optimization of new drug molecules.

Core Features

1. AI-ready Drug Discovery Datasets

Provides standardized datasets covering the full drug discovery workflow, including single-instance prediction (ADME, toxicity, bioactivity), multi-instance prediction (drug–target interactions DTI, drug–drug interactions DDI, protein–protein interactions PPI), and generation tasks (molecular generation, retrosynthesis analysis). All datasets can be used directly for training machine learning models, with support for multiple data formats and splitting strategies.

2. Multiple Data Splitting Strategies and Standardized Evaluation Tools

Offers various data splitting methods such as scaffold (based on molecular scaffolds), random, cold_drug (cold-start drugs), and cold_target (cold-start targets). This helps ensure realistic model evaluation and generalization ability. It also includes standardized evaluation metrics such as ROC-AUC, RMSE, MAE, and R2, supporting both classification and regression task evaluation commonly used in drug discovery.

3. Molecular Property Predictors and Benchmark Sets

Integrates 17+ molecular property predictors, including drug–target activity (GSK3β, JNK3, DRD2, etc.), drug selectivity, and ADMET properties. These can be used for target-directed molecular generation and optimization. It also provides professional benchmark sets such as ADMET and drug–target interactions, using multi-random-seed evaluation protocols to enable systematic comparisons of model performance and research reproducibility.

Frequently Asked Questions

What is PyTDC, and what projects is it suitable for?

PyTDC (Therapeutics Data Commons) is an open-source data platform focused on drug discovery. It is suitable for machine learning projects such as drug property prediction (e.g., ADME, toxicity), drug–target interaction prediction, molecular generation and optimization, and drug–drug interaction prediction. Its advantages include high data quality, scientifically grounded splitting strategies, and unified evaluation standards. It is especially well-suited for machine learning research and model benchmarking in the drug discovery domain.

What types of datasets does PyTDC provide?

PyTDC datasets are divided into three main categories by task type: single-instance prediction (predicting properties of individual molecules such as ADME, toxicity, and quantum chemistry properties), multi-instance prediction (predicting interactions between molecules such as drug–target binding, drug–drug interactions, and protein–protein interactions), and generation tasks (molecular generation, retrosynthesis, and pairwise molecular generation). In total, it provides 77+ carefully curated datasets covering the main application scenarios in drug discovery.

How do I get started with PyTDC?

After installing with uv pip install PyTDC, import specific task modules using from tdc.<problem> import <Task>. For example, load the ADME dataset with from tdc.single_pred import ADME. Each dataset supports the data.get_split() method to retrieve training/validation/test data and supports multiple splitting strategies such as scaffold and random. The official documentation includes detailed guides and example scripts covering the complete workflow, including data loading, model training, and benchmark evaluation.