mbzuai-ai4bio/BioMLBench-24-Public

BioMLBench — v0.1a (24 tasks), public side

This repository holds only what an agent is allowed to see. The held-out labels live in a separate dataset, mbzuai-ai4bio/BioMLBench-24-Private, so a training job pointed at this repo cannot reach them even by accident.

The full v0.1a task set of BioML-bench, prepared end-to-end and split into public/ and private/ following the BioXArena convention.

All 24 tasks are here — nothing was dropped. They span four upstream data sources with quite different characters, from 951-byte CSV assay tables to a 112 GB DICOM archive.

Tasks

proteingym-dms/ — protein fitness, 6 tasks

Deep mutational scanning; metric proteingym-metric (Spearman).

Task	Difficulty
SPIKE_SARS2_Starr_2020_binding	easy
SBI_STAAM_Tsuboyama_2023_2JVG	easy
CBX4_HUMAN_Tsuboyama_2023_2K28	medium
PSAE_PICP2_Tsuboyama_2023_1PSE	medium
Q8EG35_SHEON_Campbell_2022_indels	hard
CSN4_MOUSE_Tsuboyama_2023_1UFM_indels	hard

polarishub/ — small-molecule ADME/activity, 9 tasks

All medium, metric polaris-metric (the metric each Polaris benchmark declares).

polaris-adme-fang-hclint-1, polaris-adme-fang-hppb-1, polaris-adme-fang-solu-1, polaris-pkis2-egfr-wt-c-1, tdcommons-bbb-martins, tdcommons-caco2-wang, tdcommons-cyp2d6-substrate-carbonmangels, tdcommons-herg, tdcommons-lipophilicity-astrazeneca

manual/ — single-cell genomics (Open Problems), 5 tasks

Task	Domain	Metric	Difficulty
open-problems-predict-modality	multimodal prediction	RMSE	hard
open-problems-single-cell-perturbations	drug discovery	mean row-wise RMSE	hard
open-problems-cell-cell-communication-ligand-target	cell communication	odds ratio	hard
open-problems-label-projection	cell-type annotation	weighted F1	medium
open-problems-spatially-variable-genes	spatial transcriptomics	Kendall correlation	medium

kaggle/ — biomedical imaging, 4 tasks

Task	Metric	Raw size
rsna-miccai-brain-tumor-radiogenomic-classification	AUC-ROC	112 GB
osic-pulmonary-fibrosis-progression	modified Laplace log-likelihood	22 GB
histopathologic-cancer-detection	AUC-ROC	6 GB
uw-madison-gi-tract-image-segmentation	Dice + Hausdorff	2.4 GB

These four also appear in MLE-Bench's biomedical slice. They are published here only, so the two repos can be combined without double-counting; see mbzuai-ai4bio/MLE-Bench-Bio8.

Layout

<source>/<task>/public/
├── description.md
├── sample_submission.csv
├── train.csv | data.csv | train/ ...
└── test_features.csv | test/ ...      # labels removed

The matching -Private repo contributes <source>/<task>/private/answers.csv, so downloading both into one directory rebuilds the full layout.

The 20 non-Kaggle tasks are loose files — small enough to browse and preview in the Hub. The 4 Kaggle imaging tasks are <task>.tar.zst archives, because they hold 640k files between them:

hf download mbzuai-ai4bio/BioMLBench-24 kaggle/<task>.tar.zst --repo-type dataset --local-dir .
sha256sum -c SHA256SUMS --ignore-missing
tar --use-compress-program=unzstd -xf kaggle/<task>.tar.zst

No validation split

Every task ships train/test only. Carve your own validation fold out of train — holding out part of train is the intended workflow, and the test labels in private/ are meant for final grading only.

Grading

git clone https://github.com/biomlbench/biomlbench && cd biomlbench && pip install -e .
biomlbench grade --submission submission.csv --task-id <source>/<task>

Each task's grader is named in the table above and implemented in the upstream repo under biomlbench/tasks/<source>/<task>/grade.py.

Provenance

Prepared with BioML-bench's own prepare.py for each task, over the upstream sources (ProteinGym DMS releases, Polaris Hub / TDCommons, Open Problems S3, Kaggle). No rows were resampled or re-split.

Getting all 24 through preparation took several passes: the Open Problems tasks initially failed on missing AWS CLI and then on TLS chain verification against openproblems-data.s3.amazonaws.com, and the Polaris and ProteinGym tasks failed until their leaderboards were cached locally. All 24 have since been verified to produce prepared/{public,private} with real held-out labels.

Licensing — mixed, please read

• BioML-bench harness code: MIT.
• ProteinGym, Polaris/TDCommons, Open Problems: permissive upstream licences; cite the original dataset papers.
• The 4 kaggle/ tasks: Kaggle competition data. Competition rules generally restrict redistribution. This repository is private and intended for internal MBZUAI AI4Bio research use — do not make it public, and check each competition's rules page before resharing.

@article{biomlbench2025,
  title  = {BioML-bench: Evaluating AI Agents on Biomedical Machine Learning Tasks},
  author = {BioML-bench team},
  year   = {2025}
}

Packaged by MBZUAI AI4Bio.