TDC ADMET benchmark snapshot (2026-03-24)

Description

This is a frozen copy of the Therapeutics Data Commons ADMET benchmark group as it stood on 2026-03-24, downloaded with PyTDC 0.3.8.

It contains all 22 endpoints, each as the official scaffold split returned by tdc.benchmark_group.admet_group. No resampling, filtering, or relabelling on our side.

Motivation

The public TDC benchmark isn't version-pinned, and the underlying data can change silently. Re-downloading through PyTDC at a later date may give you a different split, which makes results from older papers hard to reproduce.

This deposit is the exact data we used in the paper below, so anyone who wants to reproduce or build on those results starts from the same place we did.

Contents

One subdirectory per endpoint, each with a train_val.csv and a test.csv:

admet_group/
├── <endpoint_1>/
│   ├── train_val.csv
│   └── test.csv
├── <endpoint_2>/
│   ├── train_val.csv
│   └── test.csv
└── ... (22 endpoints total)

Every CSV has three columns:

• Drug_ID — TDC compound identifier
• Drug — SMILES string
• Y — endpoint value (regression target or binary label, depending on the endpoint)

How to use it

To reproduce the standard TDC evaluation: run the multi-seed protocol (5 seeds) on train_val.csv and evaluate once on test.csv.

The companion code repository at github.com/receptor-ai/tdc-admet-bench has a working example, plus the feature extraction, training, hyperparameter search, and third-party model wrappers (MapLight, MapLight+GNN, CaliciBoost) used in the paper.

Associated paper

Critical Assessment of ML Models for ADMET Prediction in TDC Leaderboards — Koleiev et al., Receptor.AI (2026).

Preprint: https://www.biorxiv.org/content/10.64898/2026.02.26.708193v1.full.

License and integrity

Released under CC BY 4.0, matching this deposit's metadata and the upstream TDC dataset licensing.

Archive SHA-256: 8ac217bd8c316d04d15ab2ef5173ef6e9a084e156dfcd16c0ac21e2ed6e4590b.