Literature · wanglab
BioReason-Pro RL
A reinforcement-learning-tuned language model for biological reasoning and protein function prediction.
What BioReason-Pro RL is
BioReason-Pro RL is a specialised language model developed by the Wang Lab to advance protein function prediction and biological deduction. It serves as an optimised checkpoint that builds upon the supervised fine-tuned (SFT) BioReason-Pro base.
The model addresses the challenge of extracting accurate functional annotations from complex biological information. By sharpening reasoning pathways across Gene Ontology terms, KEGG biological pathways, and genetic variant data, it helps researchers interpret protein behaviour and biological mechanisms.
How it works
BioReason-Pro RL uses the Qwen3 architecture and undergoes reinforcement learning via Group Relative Policy Optimization (GRPO) starting from the BioReason-Pro SFT checkpoint.
The policy optimisation process trains on the wanglab/bioreason-pro-rl-reasoning-data dataset, with objectives focused on enhancing reasoning quality and improving the accuracy of Gene Ontology (GO) term predictions.
Key facts
- Base architecture
- Qwen3
- Training algorithm
- Group Relative Policy Optimization (GRPO)
- Training dataset
- wanglab/bioreason-pro-rl-reasoning-data
- Base checkpoint
- BioReason-Pro SFT
- GPU memory requirement
- 24 GB
- Release date
- 20 March 2026
What you can use it for
- Gene Ontology term prediction
- Predicting and reasoning through applicable Gene Ontology (GO) functional annotations for proteins.
- KEGG pathway reasoning
- Analysing biological pathways and functional mechanisms described within the KEGG database.
- Variant impact assessment
- Interpreting genetic variant data and predicting functional consequences for targeted proteins.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
BioReason-Pro RL
Strengths and limits
Good at
- Optimised using Group Relative Policy Optimization to generate higher-quality biological reasoning chains than standard supervised fine-tuning.
- Demonstrates improved prediction accuracy for Gene Ontology (GO) terms.
- Provides reasoned deductions across diverse biological contexts, including KEGG pathways and variant datasets.
Watch out for
- Requires a minimum of 24 GB of GPU memory for inference execution.
- Operates through standard text-generation prompting rather than direct molecular structure inputs.
- Outputs must be verified before relying on results in practical scientific settings.
Inputs and outputs
You provide
Prompt
The model continues from here.
You get back
The model's completion.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use BioReason-Pro RL commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run BioReason-Pro RL without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What distinguishes BioReason-Pro RL from BioReason-Pro SFT?
BioReason-Pro RL builds upon the supervised fine-tuned (SFT) model using Group Relative Policy Optimization (GRPO) to improve both reasoning quality and Gene Ontology term prediction accuracy.
What input and output formats does the model accept?
The model functions as a text-generation pipeline, accepting text prompts as input and returning textual completions containing biological reasoning and predictions.
What hardware is needed to execute this model?
The model requires at least 24 GB of GPU memory to run.
What datasets was the model trained on?
Reinforcement learning was conducted using the wanglab/bioreason-pro-rl-reasoning-data dataset.
How to cite it
Adibvafa Fallahpour, Arman Seyed-Ahmadi, Parsa Idehpour, Omar Ibrahim, Purav Gupta, Jack Naimer, Kevin Zhu, Arnav Shah, Shihao Ma, Abhinav Adduri, Talu Güloglu, Nuo Liu, Haotian Cui, Arihant Jain, Max de Castro, Amirfaham Fallahpour, Antonio Cembellin-Prieto, John S. Stiles, Filip Nemčko, Alexander A. Nevue, Hyungseok C. Moon, Lucas Sosnick, Olivia Markham, Haonan Duan, Michelle Y. Y. Lee, Andrea F. M. Salvador, Chris J. Maddison, Christoph A. Thaiss, Chiara Ricci-Tam, Brian S. Plosky, Dave P. Burke, Patrick D. Hsu, Hani Goodarzi, and Bo Wang. BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning. bioRxiv (2026). DOI: 10.64898/2026.03.19.712954.
Written from BioReason-Pro RL's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.