WillBolton/oncology-trial-strategy

Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents

Oncology trial-strategy decision episodes — the dataset for our ICML 2026 workshop paper.

Temporal dataset for offline policy training of clinical-trial-strategy decision agents, from our ICML 2026 workshop paper, accepted at two workshops:

• GenBio (Generative and Agentic AI for Biology) as "Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents".
• Offline2Online (Decision-Making from Offline Datasets to Online Adaptation) as "Offline Policy Learning for Clinical-Trial Strategy".

Code: https://github.com/WilliamBolton/clinical-trial-strategy-agents

Each example is a decision episode: given the date-gated state of an oncology drug program at a six-month window, predict the portfolio of trials launched next. 881 episodes, 45 programs, assembled from 31,737 raw public fetches, yielding 1,959 launch actions.

Tier 1 — ready-to-train splits (load_dataset)

Eight holdout configs (one per generalization axis). Each row is a chat example; the four training objectives share the same messages and differ only in a per-row weight column:

{"messages": [ {"role":"system",...}, {"role":"user", "..."}, {"role":"assistant","{\"trials\":[...]}"} ],
 "reward_weight_rw_bc": 1.02, "reward_weight_lr_bc": 1.34, "reward_weight_iql": 1.06}

user = date-gated program state (drug, trials, approvals, competitors + CMS spend, SEER incidence/5-yr survival, SEC revenue). assistant = ground-truth next-window portfolio (14-field trials). BC = unweighted; RW-BC/LR-BC/IQL = use the matching reward_weight_* column as the per-example sampling weight. test.jsonl is messages only (identical across objectives).

from datasets import load_dataset
ds = load_dataset("WillBolton/oncology-trial-strategy", "keytruda")   # holdout config
print(ds["test"][0]["messages"][1]["content"][:400])

Holdouts: tagrisso, gleevec, imbruvica, keytruda, opdivo (single-drug), astrazeneca (sponsor), checkpoint (drug-class PD-1/L1), temporal_split (post-Sep-2024 cutoff).

Tier 2 — structured source layer (source_layer/, rebuild the splits)

source_layer/ holds the derived, structured public-source data the splits are built from (~700 MB), laid out as:

source_layer/
├── derived/<program>/      # 45 programs: snapshots/ (CT.gov trial records), trial_outcomes.json
│                           #   (FDA-linked), revenue.json (SEC), cms_prescribing.json, lineage_*.json
└── external/cms_part_d/    # CMS Part-D competitor spending

No raw publisher full text is included. Stock/market-position features (yfinance) are omitted for licensing, so a rebuild differs marginally from the published Tier-1 splits (that one field aside).

To rebuild the splits from the code repo (https://github.com/WilliamBolton/clinical-trial-strategy-agents):

cp -r source_layer/derived/*  data/derived/
cp -r source_layer/external/* data/external/
BIO_ENRICH=1 uv run python scripts/gen_canonical_dataset.py

Provenance & licensing

Built by gen_canonical_dataset.py (BIO_ENRICH=1). Sources: ClinicalTrials.gov (public domain), Drugs@FDA / FDA reviews, SEC EDGAR, CMS Part-D, SEER — all public; observe each source's attribution. No PubMed/PMC, OpenAlex, news, or guideline full text is redistributed. License CC BY-NC 4.0 (non-commercial). Research only — not for clinical, regulatory, or investment use.

Citation

@inproceedings{bolton2026genbio,
  title={Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents},
  author={Bolton, William and Torr, Philip},
  booktitle={ICML 2026 Workshop on Generative and Agentic AI for Biology (GenBio)},
  year={2026}
}

@inproceedings{bolton2026offline,
  title={Offline Policy Learning for Clinical-Trial Strategy},
  author={Bolton, William and Torr, Philip},
  booktitle={ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation},
  year={2026}
}