span-ai-labs/oncotraj-v1

OncoTraj

A public benchmark for longitudinal resistance prediction in EGFR-mutant non-small-cell lung cancer (NSCLC) on first-line osimertinib.

Headline finding. On clean within-source evaluation, snapshot (single-timepoint) features do not beat chance on any of the three tasks at this cohort size. The binding constraint is the data modality — single-timepoint tissue/ctDNA snapshots — not the model. Serial ctDNA trajectories are the precondition for above-chance performance. That negative result is the contribution.

⚠️ This dataset is built from source, not redistributed here

The harmonized cohort is derived from AACR Project GENIE BPC, MSK-CHORD, and the FLAURA molecular-resistance supplement. GENIE BPC (registered access) and MSK-CHORD carry Data Use Agreements that restrict redistribution of patient-level data, so the patient-level tables are not hosted on this repository.

To reproduce the cohort, use the open-source code and build script:

• Code & build pipeline: https://github.com/span-ai-labs/oncotraj
• How to obtain each source: https://github.com/span-ai-labs/oncotraj/blob/main/DATA.md
• Harmonization schema: https://github.com/span-ai-labs/oncotraj/blob/main/docs/DATASET_SPEC.md

git clone https://github.com/span-ai-labs/oncotraj
cd oncotraj && uv sync
# place obtained source files under data/raw/ per DATA.md, then:
uv run python scripts/build_dataset.py
uv run python scripts/build_splits.py

Cohort

813 EGFR-mutant NSCLC patients on first-line osimertinib — MSK-CHORD (672), FLAURA supplement (107), AACR GENIE BPC NSCLC (34). Frozen, leakage-audited patient-level splits.

Tasks

• Task A — 12-month landmark progression (binary).
• Task B — time-to-progression (regression; MAE + C-index).
• Task C — resistance mechanism (6-class).

Schema

Four harmonized tables (patients, variants, treatments, outcomes), one row per entity, explicit missingness sentinels, no silent imputation. Full spec in docs/DATASET_SPEC.md.

Citation

A preprint describing OncoTraj is available; citation details will be added here once posted. Please cite the repository in the meantime.

License

Code: MIT. Data: governed by the upstream sources' terms (AACR GENIE, MSK-CHORD, and the respective journal supplements) — see DATA.md.