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

OncoTraj harmonizes real-world clinical-genomic data into a single schema with frozen, leakage-audited splits, and ships reproducible baselines and an evaluation harness. Its purpose is to establish a public floor — and an honest ceiling — for what snapshot (single-timepoint) features can predict about acquired resistance.

Headline finding (read this first). On clean within-source evaluation, snapshot-feature models 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, not a better algorithm, are the precondition for above-chance performance. That negative result is the contribution: it tells the field where the signal must come from. See leaderboard.md for the full, CI-annotated numbers and the cross-source (v1.1) analysis.

v1 comprises 813 EGFR-mutant NSCLC patients on first-line osimertinib, harmonized from three real-world clinical-genomic sources:

Patient-level test split: n=122 (Task A landmark-evaluable n=110; Task B/C event-observed n=85).

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

Majority/floor, logistic regression, random forest, XGBoost, LSTM, a small Transformer, and a Cox proportional-hazards model (Task B). All baselines reuse a single leak-safe feature pipeline; the leakage audit lives in tests/test_taska_feature_leakage.py.

uv sync
uv run pytest          # test suite, incl. leakage audits
uv run ruff check

Evaluate a submission against the locked test split:

uv run oncotraj-eval --predictions yours.csv --split test \
    --submission-id <your-name> --output eval_reports_v1/<your-name>.json
uv run oncotraj-eval --refresh   # regenerate leaderboard.md

The prediction CSV schema is documented in src/oncotraj/eval/report.py (SCHEMA_DOC).

The harmonized dataset is not redistributed in this repository. The underlying sources (AACR GENIE BPC, MSK-CHORD) carry Data Use Agreements that restrict redistribution of patient-level data. Instead, OncoTraj ships the parsers and a build script that regenerate the harmonized Parquet tables from sources you obtain under their respective terms:

# after placing source files under data/raw/ (see docs/DATASET_SPEC.md and DATA.md)
uv run python scripts/build_dataset.py
uv run python scripts/build_splits.py

See DATA.md for how to obtain each source, and docs/DATASET_SPEC.md for the full harmonization schema (one row per patient/variant/treatment/outcome, explicit missingness sentinels, no silent imputation).

src/oncotraj/      # library: schemas, parsers, splits, metrics, eval, baselines
scripts/           # build_dataset, build_splits, train_*, make_figures
tests/             # pytest suite, incl. leakage audits
eval_reports/      # baseline metric reports (JSON)
eval_reports_v1/   # v1 + v1.1 cross-source reports (JSON)
leaderboard.md     # auto-generated, CI-annotated results
docs/              # dataset spec and reproduction docs
data/, models/     # gitignored — built locally, not redistributed

A preprint describing OncoTraj is available. Citation details will be added here once the preprint is posted. In the meantime, please cite this repository.

Code: MIT (see LICENSE). Data are not included; any data you build with these tools remain governed by the upstream sources' terms (AACR GENIE, MSK-CHORD, and the respective journal supplements).