TDC

Developed by Harvard researchers, TDC consolidates decades of scattered pharmaceutical data into a unified, standardized format that ML researchers can access through a Python APIcovering molecular property prediction, drug-target interaction, ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) prediction, and clinical trial outcome modeling.

The platform includes over 70 machine learning tasks derived from real therapeutic discovery problems, each with standardized train/validation/test splits, evaluation metrics, and baseline model results that enable fair comparison across approaches.

Data sources span small molecules, proteins, clinical observations, and multi-omics measurements, supporting research across the full spectrum from early-stage target identification through clinical outcome prediction.

TDC actively curates for data quality issues that have historically limited reproducibility in computational drug discovery research.

ML researchers working on drug discovery applications, computational chemists evaluating graph neural network approaches to molecular property prediction, and pharmaceutical companies benchmarking internal ML models against published baselines use TDC as a common evaluation framework.

The platform's standardized benchmark format has improved reproducibility in ML-for-drug-discovery research by reducing the ad-hoc dataset curation and splitting decisions that previously made result comparison across papers unreliable.

Its foundation model extensionproviding pre-computed embeddings and multimodal representationsmakes it increasingly relevant as large biological foundation models become central to modern computational drug discovery.