lokahq/bioreason-rl

BioReason RL

Dataset Summary

BioReason RL is a reward-filtered dataset for reinforcement-learning (GRPO) training and evaluation of protein Gene Ontology (GO) function prediction with reasoning. Each example is a protein record with GO term annotations across the three GO aspects, InterPro domains, STRING protein-protein interactions, baseline GO predictions, and a set of per-example reward signals recorded during curation.

The dataset is a curated subset of the BioReason-Pro data released by the Wang Lab. The train split is filtered from wanglab/bioreason-pro-rl-reasoning-data and the test split is filtered from wanglab/bioreason-pro-test-data. It is intended for Prime-RL BioReason ablations where train and eval must share the same schema and evidence channels.

Dataset Details

• Repository: lokahq/bioreason-rl
• Task: reinforcement learning (GRPO) for protein GO function prediction with reasoning
• Language: English
• Domain: computational biology, protein function annotation, bioinformatics
• Format: Parquet, one protein record per row
• Train source: wanglab/bioreason-pro-rl-reasoning-data (train split)
• Test source: wanglab/bioreason-pro-test-data (test split)
• Selection: per-example reward filtering of the source splits

Splits

split	rows	source dataset	source split	source rows
train	4,810	wanglab/bioreason-pro-rl-reasoning-data	train	9,197
test	3,844	wanglab/bioreason-pro-test-data	test	8,630
total	8,654

The train and test splits are drawn from two distinct source datasets and do not overlap. Both splits share the same column schema and evidence channels (GO annotations, InterPro domains, STRING PPIs, baseline GO predictions), so a model trained on train can be evaluated on test without any schema or feature mismatch.

Data Fields

Protein record fields

Carried from the source datasets:

• protein_id: protein identifier.
• protein_names: protein name(s).
• protein_function: free-text functional description.
• organism: source organism.
• length: sequence length.
• subcellular_location: subcellular localization annotation.
• sequence: amino-acid sequence.
• go_ids: list of Gene Ontology term IDs.
• go_bp: GO terms for the Biological Process aspect.
• go_mf: GO terms for the Molecular Function aspect.
• go_cc: GO terms for the Cellular Component aspect.
• structure_path: reference to the predicted structure artifact.
• interpro_location, interpro_formatted: InterPro domain annotations.
• ppi_formatted: STRING protein-protein interaction context.
• go_pred: baseline GO term predictions (populated for all rows in both splits).

Reward-filter fields

Per-example reward signals recorded during curation and used to select the published subset:

• reward_filter_source_index: row index within the source split.
• reward_filter_reward: aggregate reward score.
• reward_filter_go_f1_mean: mean GO F1 across aspects.
• reward_filter_go_pair_f1: pairwise GO F1.
• reward_filter_aspect_presence_accuracy: accuracy of predicting which GO aspects are present.
• reward_filter_json_format_score: validity score of the structured JSON output.
• reward_filter_go_ted_score: GO tree-edit-distance score.
• reward_filter_is_truncated: whether the scored generation hit the length cap.
• reward_filter_stop_condition: recorded stop condition for the scored generation.

Reward Filtering

The published splits are quality-filtered subsets of the source datasets, selected using the per-example reward signals above (train: 4,810 of 9,197; test: 3,844 of 8,630). Across all retained rows:

• reward_filter_json_format_score is 1.0 (every retained generation produced valid structured JSON).
• reward_filter_is_truncated is False (no truncated generations were retained).
• reward_filter_stop_condition is uniformly max_turns_reached.

The full set of reward_filter_* fields is preserved on every row so downstream consumers can re-threshold, stratify, or weight examples by reward component (overall reward, GO F1, pairwise GO F1, aspect presence accuracy, GO tree-edit-distance) without recomputing them.

Intended Use

This dataset is intended for:

• reinforcement-learning (GRPO) training of protein GO function prediction with reasoning
• evaluation on a held-out, schema-aligned test split
• Prime-RL BioReason ablations that require train and eval to share the same schema and evidence channels
• research on reward-filtered data curation for biological reasoning

Out-of-Scope Use

This dataset should not be used as clinical, diagnostic, or wet-lab decision support. GO annotations and baseline predictions may be incomplete or stale. Downstream systems should validate biological claims against current primary databases (UniProt, QuickGO, InterPro, STRING).

Loading

from datasets import load_dataset

dataset = load_dataset("lokahq/bioreason-rl")
train = dataset["train"]
test = dataset["test"]

Files

• data/train-00000-of-00001.parquet: training split (filtered from wanglab/bioreason-pro-rl-reasoning-data).
• data/test-00000-of-00001.parquet: test split (filtered from wanglab/bioreason-pro-test-data).

License

Released under the Apache 2.0 License, inherited from the upstream BioReason-Pro datasets.

Citation

This dataset is derived from BioReason-Pro. If you use it, please cite the upstream work:

@article{Fallahpour2026.03.19.712954,
    author = {Fallahpour, Adibvafa and Seyed-Ahmadi, Arman and Idehpour, Parsa and Ibrahim, Omar and Gupta, Purav and Naimer, Jack and Zhu, Kevin and Shah, Arnav and Ma, Shihao and Adduri, Abhinav and G{\"u}loglu, Talu and Liu, Nuo and Cui, Haotian and Jain, Arihant and de Castro, Max and Fallahpour, Amirfaham and Cembellin-Prieto, Antonio and Stiles, John S. and Nem{\v c}ko, Filip and Nevue, Alexander A. and Moon, Hyungseok C. and Sosnick, Lucas and Markham, Olivia and Duan, Haonan and Lee, Michelle Y. Y. and Salvador, Andrea F. M. and Maddison, Chris J. and Thaiss, Christoph A. and Ricci-Tam, Chiara and Plosky, Brian S. and Burke, Dave P. and Hsu, Patrick D. and Goodarzi, Hani and Wang, Bo},
    title = {BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning},
    elocation-id = {2026.03.19.712954},
    year = {2026},
    doi = {10.64898/2026.03.19.712954},
    publisher = {Cold Spring Harbor Laboratory},
    journal = {bioRxiv}
}

@misc{fallahpour2025bioreasonincentivizingmultimodalbiological,
      title={BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model},
      author={Adibvafa Fallahpour and Andrew Magnuson and Purav Gupta and Shihao Ma and Jack Naimer and Arnav Shah and Haonan Duan and Omar Ibrahim and Hani Goodarzi and Chris J. Maddison and Bo Wang},
      year={2025},
      eprint={2505.23579},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.23579}
}