Therapeutics Benchmark · maomlab
TDC
Curated datasets spanning 22 single-instance prediction tasks for machine learning in therapeutic discovery.
What TDC is
TDC (Therapeutics Data Commons), published by maomlab, provides curated benchmark datasets for evaluating machine learning methods in therapeutic discovery. Spanning chemistry, biology, and medicine, this repository serves as a mirror dedicated to single-instance prediction tasks from the broader Therapeutics Data Commons project.
The collection contains 7,949,637 data points across 22 tasks, covering tabular classification and tabular regression. Key problem domains include absorption, distribution, metabolism, and excretion (ADME), toxicology (Tox), high-throughput screening (HTS), and quantum mechanical (QM) calculations.
How it was made
The data is structured into category-specific Parquet files, such as ADME, Tox, and HTS, with each file holding multiple distinct target properties such as solubility or permeability. Because preprocessing requirements vary by task, each category follows its own schema and column definitions.
Molecular representations are curated prior to inclusion. SMILES strings in the ADME, Tox, and HTS datasets are standardised to canonical forms with salts removed, resulting in the exclusion of 2 invalid SMILES strings from ADME and 59 from HTS. For quantum mechanical tasks, structures are converted into atom-level records specifying atomic identities and three-dimensional coordinates rather than single molecular instances.
Key facts
- Total data points
- 7,949,637
- Tasks included
- 22 single-instance prediction tasks
- Task categories
- Tabular classification and tabular regression
- Primary domains
- ADME, Tox, HTS, and quantum mechanics (QM)
- Data format
- Parquet with a companion summary CSV
- Publisher
- maomlab
What you can use it for
- ADMET Property Prediction
- Benchmarking regression and classification models on pharmacological endpoints, including solubility and membrane permeability.
- Toxicity Profiling
- Training algorithms to identify toxicological hazards and adverse liabilities in chemical candidates.
- High-Throughput Screening Analysis
- Modelling bioactivity outcomes and hit identification across large compound libraries.
- Quantum Chemical Modelling
- Evaluating molecular representation methods using atom identities and 3D spatial coordinates.
Explore the data
No queryable preview
This dataset is not available in a queryable form — usually an unusual file format or a repository too large to convert. The original source may still be readable.Strengths and limits
Good at
- Standardises and desaturates SMILES strings to canonical forms across chemical bioactivity subsets.
- Includes a metadata summary file specifying leaderboard metrics, evaluation splits, compound numbers, and task descriptions.
- Provides millions of curated data points across diverse drug-discovery problem categories.
Watch out for
- Each category file follows an independent schema with varying column names and preprocessing approaches.
- Coverage is restricted to single-instance prediction tasks and does not include multi-instance or network-based workflows.
- A small number of invalid SMILES had to be discarded from the ADME and HTS datasets during sanitisation.
Questions
Can I look inside TDC without downloading it?
Not here: it is not available in a queryable form. Its description and statistics are still shown.
What licence is TDC under?
Its authors declare mit. Read the terms on the original page before reusing the data.
What chemical structure sanitisation steps were applied?
SMILES strings in the ADME, Tox, and HTS datasets were converted to canonical format, and salts were stripped. Two unparseable SMILES in ADME and 59 in HTS were excluded.
How are quantum mechanical datasets formatted?
Quantum mechanical structures are resolved into atom-level records containing atomic identities and three-dimensional Cartesian coordinates, which are not counted as independent molecular observations.
What information does the summary file contain?
The summary file details category names, tasks, dataset identifiers, compound counts, descriptions, original references, and benchmark leaderboard settings such as metrics and data splits.
Written from TDC's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.