Jerrychen229/PhenoBench

mmp-bench

Twelve data-analysis tasks built from published Cell Painting and transcriptomic screens.

Each task hands an agent a real, unprocessed deposit and asks a question the original paper asked. Nothing is normalised, filtered or aggregated in advance: the preprocessing decisions are the task. A submission is graded by a frozen scorer against a threshold that was calibrated before any model ran, by measuring pipelines that do the work against pipelines that skip a step.

What is and is not in this release

Included: the input data, the prompts, the scorers, the frozen thresholds and their calibration, the answer keys, and a card per task explaining how each threshold was set.

Not included: model results and agent trajectories. This release is the benchmark, not an evaluation of anything.

Layout

tasks/<task_id>/
├── task.md       the prompt, verbatim as the agent receives it
├── CARD.md       for whoever is running or reviewing the task; never given to the agent
├── data/         mounted at /app — the agent's working directory, task.txt and README included
└── evaluator/    scorer, frozen threshold, answer key

data/ is byte-for-byte the environment the evaluated runs used, and so is task.md. task.md and data/task.txt overlap in content; that duplication is real and is preserved rather than tidied away, because merging them would ship a variant that was never evaluated.

The evaluator files are byte-for-byte too, with one documented exception: seven tasks carried a build-stage status field with five different values (frozen, not_final, frozen_v4_multireference, frozen_for_agent_pilot, final), and all are set to final for release. No scorer reads that field, so no score changes; it was checked rather than assumed. Thresholds, keys and every other value are untouched.

The twelve tasks

task	paper / deposit	endpoint	threshold	chance	data
rohban17-gene-pairs	Rohban 2017	joint overlap vs. any of 15 accepted endpoints	≥13/15 pairs and ≥5/7 modules	0 of 100,000 random draws	229M
rohban22-mapk14	Rohban 2022	signed set overlap, K=15+15	≥10/30, and CP ≤1.0	—	1.4G
lazar24-proxbias	Lazar 2024, Nat Genet	planted module recall in the extreme 10%	≥0.363, and CP in [0.45, 0.6201], and a guard	0.10	1.5G
tegtmeyer25-22q11	Tegtmeyer 2025	per-cell-type cross-validated accuracy	inside a per-cell-type band	—	83M
caicedo22-variant-impact	Caicedo 2022	separation of planted impactful from neutral	≥0.4764	—	109M
glunk23-genotype	Glunk 2023	AUROC of correctly signed planted features	≥0.666	0.500	7.0M
lacoste24-mislocalisation	Lacoste, Haghighi 2024	AUROC vs. planted truth	≥0.8169	0.500	137M
way22-two-assay	Way 2022	tail_margin against withheld class-mates	≥0.1158	—	359M
way21-cell-health	Way 2021	Spearman(held-out correlation, planted signal fraction)	≥0.746	0.096	197M
kelley23-bortezomib	Kelley 2023	P(dosed clone scores above vacuum clone)	≥0.8447	0.5037	2.6M
rxrx1-cpg0016	RxRx1 × cpg0016	annotation recall at n, five gene-function sets	≥0.0091	0.00125	1.3G
j1mm-crossmodal	CPJUMP1 × LINCS L1000	Spearman(morphology transfer, expression transfer)	inside [+0.21, +0.865], and CP in [+0.316, +0.642]	0.000 exactly	240M

Total 5.4 GB. Read the task's CARD.md before using its number: several thresholds are two-sided, three tasks grade more than one quantity, and six of the twelve are semi-synthetic in a way the card states exactly.

Running a scorer

The scorers depend only on numpy, pandas and pyarrow. They locate their keys relative to their own path, so a task directory can be moved anywhere. Two take their paths from the environment instead of the command line.

python tasks/rohban17-gene-pairs/evaluator/score.py        <sub>/discoveries.csv
python tasks/rohban22-mapk14/evaluator/score_submission.py --workspace tasks/rohban22-mapk14/data \
                                                           --submission <sub> --output score.json
VERIFIER_TESTS=tasks/lazar24-proxbias/evaluator VERIFIER_APP=<sub> \
  python tasks/lazar24-proxbias/evaluator/score.py
python tasks/tegtmeyer25-22q11/evaluator/EP/score.py       <sub>/effect.csv
python tasks/caicedo22-variant-impact/evaluator/EP/score.py <sub>/variant_calls.csv
python tasks/glunk23-genotype/evaluator/score.py           <sub>/genotype_effects.csv
PLANTED_TRUTH=tasks/lacoste24-mislocalisation/evaluator/planted_truth.csv \
  python tasks/lacoste24-mislocalisation/evaluator/score.py <sub>/predictions.csv
python tasks/way22-two-assay/evaluator/score.py            <sub>/consensus_morphology.parquet
python tasks/way21-cell-health/evaluator/score.py          <sub>/crispr_predictions.csv
python tasks/kelley23-bortezomib/evaluator/score.py        <sub>/clone_predictions.csv
python tasks/rxrx1-cpg0016/evaluator/score_submission.py   --workspace tasks/rxrx1-cpg0016/data \
                                                           --submission <sub> --output score.json
python tasks/j1mm-crossmodal/evaluator/score_submission.py --workspace tasks/j1mm-crossmodal/data \
                                                           --submission <sub> --output score.json

Each prints or writes JSON. final_pass — or passes, in the older scorers — is the verdict.

Every scorer in this package was checked against a real archived submission and reproduces the score that submission received when it was evaluated.

Running the agent side

The evaluated runs used Harbor 0.21.0 with the singularity backend and mini-swe-agent 2.4.6, mounting data/ at /app inside a container whose Python is 3.11 with numpy 2.4.6, pandas 2.3.3, pyarrow 24.0.0, scipy 1.17.1 and scikit-learn 1.9.0. No image is shipped here: any harness that presents task.md to an agent with data/ as its working directory and collects the files the prompt names will do. The scorers do not depend on Harbor.

Data sources and licences

This benchmark is non-commercial research and is distributed on that basis.

task	source	licence
rohban17-gene-pairs, rohban22-mapk14	Cell Painting Gallery; LINCS L1000	CC0 / CC-BY family
lazar24-proxbias, rxrx1-cpg0016 (CRISPR arm)	cpg0016, Cell Painting Gallery	CC0
rxrx1-cpg0016 (siRNA arm)	RxRx1, Recursion	non-commercial research use only
tegtmeyer25-22q11	authors' deposit	MIT
caicedo22-variant-impact	authors' deposit	CC0 / CC-BY family
glunk23-genotype	authors' GitHub repository	no LICENSE file — see below
lacoste24-mislocalisation	authors' deposit	MIT
way22-two-assay, way21-cell-health	authors' deposits	MIT
kelley23-bortezomib	authors' deposit	MIT
j1mm-crossmodal	CPJUMP1 (CC0) + LINCS L1000	morphology CC0; expression licence unconfirmed

Two entries need stating plainly rather than burying:

• glunk23-genotype is built from two CSVs in a GitHub repository that carries no LICENSE file, so redistribution rights are unclear. They are nevertheless the only machine-readable form of these experiments. Anyone reusing this task should check the upstream repository's terms for themselves.
• j1mm-crossmodal uses CMap/LINCS expression data. It has been released for unrestricted research use and is served without authentication, but the exact licence text was not located during the build, so it is flagged rather than asserted.

Nature supplementary tables underlying glunk23-genotype's key are evaluator-side derived statistics only and are not in any data/ directory.

How the thresholds were set

Every threshold in the table above was placed in a measured gap: pipelines that do the work on one side, pipelines that skip a necessary step or take a shortcut on the other, with bootstrap intervals that do not overlap. Each task's CARD.md gives the arms and the edges.

Two conventions are worth knowing:

• A step counts as necessary only if dropping it moves the answer. Where a step the source paper presents as required turns out not to matter on the shipped data, the card says so rather than filing the arm as broken. Four tasks have one: kelley23-bortezomib (the cell-count exclusion), way21-cell-health (per-plate normalisation, on the CRISPR side), way22-two-assay (whitening against the controls) and rxrx1-cpg0016 (batch correction, where global whitening subsumes it).
• A checkpoint is scored only where it is measured to catch something the endpoint cannot. j1mm-crossmodal is where that measurement came out positive — skipping replicate aggregation leaves the endpoint inside its interval and sends the checkpoint outside. On rxrx1-cpg0016 the same measurement came out the other way and that checkpoint is recorded but not scored.

Eight of the twelve are semi-synthetic: rohban22-mapk14, lazar24-proxbias, tegtmeyer25-22q11, caicedo22-variant-impact, glunk23-genotype, lacoste24-mislocalisation, way21-cell-health and kelley23-bortezomib. In each case the real deposit was tried first and could not separate defensible analyses from broken ones; the card gives the measurement that forced the decision and the exact construction, and in every one the underlying measurements stay real — what is constructed is which of them carries the signal.

The four graded against real truth are rohban17-gene-pairs, way22-two-assay, rxrx1-cpg0016 and j1mm-crossmodal. (rxrx1-cpg0016 had a planted predecessor; it was withdrawn, and the card says why.)