Bacterial gene-expression competition (benchopt + Codabench)

A Codabench competition whose evaluation is a benchopt benchmark: participants run the exact same code locally that the server runs. The task is to predict RNA-seq expression counts for a gene × sample observation from its genome sequence, gene description and experimental (environment) metadata, and to generalize to held-out bacterial species.

See gene_expression_getting_started.ipynb for an end-to-end walkthrough.

The benchmark (benchmark/)

The self-contained benchopt benchmark — run it directly with benchopt run benchmark:

• benchmark/objective.py — the task + metric (Negative-Binomial NLL, MAE/R²) per region, on the public and hidden private halves; always trains.
• benchmark/datasets/ — simulated.py (zero-dependency smoke test) and bacterial_expression.py (downloads the real data from Zenodo in prepare and writes its train / A / B / C / D regions).
• benchmark/solvers/ — constant.py (floor), baseline_nn.py (NN).
• benchmark/benchmark_utils/ — the metric, the GeneCompetSolver submission contract, feature engineering, shared helpers.

The competition wrapper (repo root)

• competition.yaml — the single-phase competition (public columns + hidden private columns).
• ingestion_program/ — a thin benchopt run wrapper (runs the submission).
• scoring_program/ — parses the benchopt results into scores.json.
• phases/main/input_data/config.yaml — the phase config (a native benchopt run config + a scoring column map).
• pages/, solution/submission.py — participant pages and the sample submission.
• tools/ — setup_data.py (build the release), Dockerfile, requirements.txt, create_bundle.py, run_docker.py.

Quickstart

pip install -r requirements.txt

# download the data, fit on train and score on the eval regions
# (mode=train for all the public data)
benchopt prepare benchmark -d "BacterialExpression[mode=debug]"
benchopt run benchmark -d "BacterialExpression[mode=debug]" \
    -s solution/submission.py

# design-level tests + zero-dependency smoke test
benchopt test benchmark --skip-install
benchopt run benchmark -d Simulated -s Constant

Build the Codabench bundle

python tools/create_bundle.py     # -> bundle.zip (upload to Codabench)

The bundle ships the config, programs, pages and the benchmark (in the phase input_data, so a benchmark change needs no image rebuild). The data (including the sealed private split) is staged separately onto the compute worker at $BENCHOPT_DATA_HOME (/app/data), so it never enters the public artifact.