DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists
Paper • 2610.09558 • Published • 1
Accept the ACDC terms to access DrugTargetWorld assets
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains cine-MRI derived from anatomy in the ACDC (Automated Cardiac Diagnosis Challenge) database, which is distributed by Creatis under registration at https://humanheart-project.creatis.insa-lyon.fr. Access is limited to people who have registered for ACDC and agree to its terms of use, including not redistributing ACDC data or data derived from it, and who cite Bernard et al., IEEE Transactions on Medical Imaging 37(11):2514-2525, 2018, in any use of the imaging. The synthetic omics, health-record and outcome data in this dataset are DrugTargetWorld's own; see LICENSE.md.
Log in or Sign Up to review the conditions and access this dataset content.
DrugTargetWorld assets
Immutable per-world payloads for DrugTargetWorld, from the paper DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists (arXiv:2610.09558).
Paper · Full text (HTML) · Project website · Code · Benchmark
DrugTargetWorld is an environment for training and evaluating AI scientists using procedurally generated multimodal biobanks with known but concealed causal ground truth. It evaluates open-ended phenotype construction, causal target discovery, therapeutic direction and budgeted experimentation with verifiable reward.
Paper: Margolis et al., 2026. DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists. arXiv:2610.09558.
Each world is a synthetic biobank of 54,000 participants generated from a hidden structural causal model, with realistic causal traps (confounding, reverse causation, collider bias, pleiotropy, surrogate-outcome discordance) and budgeted virtual experiments. Methodology: scientific worlds · causal traps · open-ended science · synthetic biobank · experimentation · findings.
Research questions this dataset supports: Can AI scientists distinguish causal targets from merely associated proteins? Can agents construct useful phenotypes from raw multimodal data? Do agents allocate experimental resources efficiently? Can procedurally generated causal worlds train scientific strategies? How should AI scientists behave when causal relationships are non-identifiable?
You do not download this repository by hand: Harbor resolves the archives a task needs and verifies them against the frozen manifest before the agent starts.
Access is gated behind the ACDC terms (see below). Accept them once on this page, then give Harbor your Hugging Face token:
uv tool install harbor
export HF_TOKEN=hf_... # a token from an account that accepted the terms
harbor run -d drugtargetbench/drugtargetbench@v1.1 -a <agent> -m <model>
Layout
worlds/<world-id>/tables.tar omics, tabular, data dictionary
worlds/<world-id>/signals.tar ECG waveforms, coronary angiograms
worlds/<world-id>/oracle-state.tar experiment-service state
worlds/<world-id>/imaging-NNN.tar cine-MRI, ~4 GB shards
worlds/<world-id>/imaging-visit2-NNN.tar
worlds/<world-id>/metadata.json
manifests/assets.json
manifests/sha256sums.txt
Twenty worlds, one independently addressable payload each. The three Harbor budget regimes for a world reference the same payload, so a full 60-task sweep materialises 345 GB once rather than 1.03 TB.
Access and licensing
| Component | Bytes | Source | Terms |
|---|---|---|---|
Omics, genotypes, covariates, EHR, ECG, coronary, outcomes (tables.tar, signals.tar) |
72.5 GB | Generated by DrugTargetWorld from a structural causal model; no participant data | DrugTargetWorld's own, MIT |
Cine-MRI (imaging-*.tar, imaging-visit2-*.tar) |
272.8 GB | Warped from anatomy in the ACDC dataset | ACDC terms of use; derivative of registration-gated data |
Because the imaging is derived from ACDC, which is distributed under registration, the whole dataset is gated: access requires confirming that you have registered for ACDC, agree to its terms, will not redistribute the imaging, and will cite Bernard et al. (IEEE TMI 2018). Details per component are in LICENSE.md.
If you prefer to render the imaging from your own ACDC download instead, Harbor rebuilds it locally and verifies it against the same SHA-256 values, so a trial is byte-identical either way:
export DTB_ACDC_ROOT=/path/to/your/ACDC/training
Integrity
manifests/sha256sums.txt covers every archive. Harbor checks each one after
download and refuses to run on a mismatch rather than scoring a partial
world. Always pin an immutable commit sha; never run against main.
Citation
Margolis S, Schmiedmayer P, Huang A, et al. DrugTargetWorld: A Synthetic Biobank for Training and Benchmarking AI Scientists. arXiv:2610.09558 (2026).
@article{margolis2026drugtargetworld,
title = {{DrugTargetWorld}: A Synthetic Biobank for Training and
Benchmarking {AI} Scientists},
author = {Margolis, Samuel and Schmiedmayer, Paul and Huang, Alan and
Chen, Ethan and Bhattacharjee, Ishan and Shah, Atman and
Viggiano, Ben and Cao, Fang and Reddy, Shriya and Xia, Roger and
O'Sullivan, Jack and Katz, Daniel and Wheeler, Matthew and
Ashley, Euan and Gomes, Bruna},
journal = {arXiv preprint arXiv:2610.09558},
year = {2026},
doi = {10.48550/arXiv.2610.09558}
}
Cite ACDC separately if you use the real-anatomy imaging: Bernard et al., Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis, IEEE TMI 2018.
- Downloads last month
- 488