Biological Reasoning Corpus · wanglab
bioreason-pro-sft-reasoning-data
A training corpus of biological reasoning traces and annotations for supervised fine-tuning of protein models.
What bioreason-pro-sft-reasoning-data is
bioreason-pro-sft-reasoning-data is a biological reasoning corpus developed by the Wang Lab. It provides supervised fine-tuning data for BioReason-Pro, a system developed to advance protein function prediction using multimodal biological reasoning.
Predicting protein function requires drawing connections across diverse biological evidence, including molecular domains, interaction partners, and pathway contexts. This dataset addresses the problem by pairing protein records with synthetic step-by-step reasoning traces that explain biological relationships and functional predictions.
How it was made
The corpus compiles protein metadata from UniProt alongside functional terms from the Gene Ontology, structural and functional protein domains from InterPro, and protein–protein interaction networks from the STRING database. Synthetic reasoning traces were generated by GPT-5 to form rationales linking these attributes to biological outcomes.
Key facts
- Format
- Parquet, optimized-parquet
- Modality
- Text
- Reasoning trace generator
- GPT-5
- Biological data sources
- UniProt, Gene Ontology, InterPro, STRING
- Language
- English
What you can use it for
- Supervised fine-tuning of reasoning models
- Training language and sequence models to output multi-step biological rationales alongside protein predictions.
- Protein function prediction research
- Developing and evaluating methods that associate protein sequences and interaction networks with Gene Ontology annotations.
- Analysis of multimodal biological evidence
- Studying how machine learning architectures synthesize domain annotations from InterPro and network data from STRING to resolve pathway problems.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from wanglab/bioreason-pro-sft-reasoning-data at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 117,002 rows of default/train
final_answer
string_text- min
- 0
- median
- 1,705
- mean
- 1792.49
- max
- 10,310
go_bp
list- min
- 2
- median
- 28
- mean
- 38.97
- max
- 708
34.6% empty
go_bp_leaf
string_text- min
- 0
- median
- 44
- mean
- 105.09
- max
- 6,486
go_cc
list- min
- 2
- median
- 11
- mean
- 13.05
- max
- 102
32.2% empty
go_cc_leaf
string_text- min
- 0
- median
- 24
- mean
- 36.50
- max
- 898
go_ids
list- min
- 2
- median
- 25
- mean
- 39.43
- max
- 815
go_mf
list- min
- 2
- median
- 7
- mean
- 8.65
- max
- 83
41.1% empty
go_mf_leaf
string_text- min
- 0
- median
- 26
- mean
- 40.10
- max
- 1,173
go_pred
string_text- min
- 280
- median
- 1,917
- mean
- 2427.08
- max
- 15,472
go_pred_leaf
string_text- min
- 140
- median
- 281
- mean
- 369.15
- max
- 6,030
interaction_partners
list- min
- 1
- median
- 14
- mean
- 31.83
- max
- 1,441
33.0% empty
interpro_formatted
string_text- min
- 32
- median
- 279
- mean
- 329.30
- max
- 1,792
interpro_ids
list- min
- 1
- median
- 4
- mean
- 4.63
- max
- 20
interpro_location
string_text- min
- 22
- median
- 96
- mean
- 112.65
- max
- 540
length
float- min
- 10
- median
- 420
- mean
- 564.91
- max
- 35,375
organism
string_text- min
- 9
- median
- 35
- mean
- 36.75
- max
- 178
ppi_formatted
string_text- min
- 0
- median
- 178
- mean
- 186.02
- max
- 893
protein_function
string_text- min
- 9
- median
- 90
- mean
- 241.83
- max
- 10,648
protein_id
string_text- min
- 6
- median
- 6
- mean
- 6.22
- max
- 10
protein_names
string_text- min
- 2
- median
- 58
- mean
- 75.69
- max
- 2,586
reasoning
string_text- min
- 0
- median
- 4,021
- mean
- 4130.78
- max
- 14,082
sequence
string_text- min
- 10
- median
- 420
- mean
- 564.91
- max
- 35,375
string_id
string_text- min
- 10
- median
- 16
- mean
- 16.90
- max
- 27
25.0% empty
structure_path
string_text- min
- 25
- median
- 25
- mean
- 25.16
- max
- 29
5.9% empty
subcellular_location
string_text- min
- 7
- median
- 88
- mean
- 158.92
- max
- 5,399
30.8% empty
Strengths and limits
Good at
- Combines data across standard biological resources, including UniProt, InterPro, STRING, and the Gene Ontology.
- Provides multi-step reasoning traces rather than simple label-annotation pairs for protein records.
- Delivered in high-performance Parquet formats compatible with standard data tools including Polars, Dask, and Datasets.
Watch out for
- The explanatory reasoning traces are synthetically generated by GPT-5 rather than manually curated by experimental biologists.
- All biological features and relationships are represented exclusively in textual form.
Questions
Can I look inside bioreason-pro-sft-reasoning-data without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is bioreason-pro-sft-reasoning-data under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What biological information is included in the dataset?
The dataset contains protein metadata from UniProt, Gene Ontology annotations, protein domain information from InterPro, and protein–protein interaction data from STRING.
How were the reasoning traces produced?
The rationales and reasoning traces accompanying each protein entry were synthetically generated using GPT-5.
What file format does the corpus use?
The corpus is distributed as text entries in standard and optimized Apache Parquet formats.
How to cite it
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 Güloglu, 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 Nemčko, 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. BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning. bioRxiv, 2026. DOI: 10.64898/2026.03.19.712954.
Written from bioreason-pro-sft-reasoning-data's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.