pytdc

PyTDC (Therapeutics Data Commons) is a comprehensive open-science platform designed to facilitate machine learning in drug discovery and therapeutic development. It serves as a centralized hub for standardized datasets, evaluation metrics, and model benchmarks that are essential for researchers and AI agents working on pharmaceutical problems. By providing high-quality, curated data for the entire therapeutics pipeline, it enables seamless training, testing, and validation of models ranging from small-molecule property prediction to large-scale interaction networks.

• Access a vast repository of datasets for single-instance prediction, including ADME (Absorption, Distribution, Metabolism, Excretion), toxicity profiles (hERG, AMES, DILI), and High-Throughput Screening (HTS) bioactivity data.

• Handle complex multi-instance prediction tasks such as Drug-Target Interaction (DTI) using BindingDB or DAVIS, Drug-Drug Interaction (DDI) within DrugBank, and Protein-Protein Interaction (PPI) networks.

• Utilize generative capabilities for molecule discovery, including scaffold-based splitting to ensure robust model generalization and evaluation.

• Leverage molecular oracles for property-directed optimization, allowing users to score and refine novel molecules during the generation process.

• Support standardized data splitting methods including random, scaffold, and cold-start splits to simulate realistic clinical or experimental conditions.

• Install the library via pip (uv pip install PyTDC) and use a consistent programmatic pattern (from tdc.<problem> import <Task>) to fetch data frames.

• Typical inputs include dataset identifiers within categories like single_pred, multi_pred, or generation; outputs are provided as standardized pandas DataFrames.

• Ensure you are working within scientific or pharmaceutical ML contexts where specific pharmacokinetics or biochemical binding constraints apply.

• The tool is best suited for agents that need to benchmark model performance against state-of-the-art pharmaceutical datasets or iterate on molecule design workflows.