Literature · wanglab
GO-GPT
Generative model predicting Gene Ontology terms across all three ontology aspects directly from protein sequences.
What GO-GPT is
GO-GPT is a generative model developed by wanglab that predicts Gene Ontology (GO) functional annotations directly from primary protein sequences. Built as part of the BioReason-Pro project, it applies language model-style autoregressive decoding to functional characterisation in biology, genomics, and medicine.
Traditional automated function prediction relies heavily on discriminative classification frameworks. GO-GPT instead reframes functional annotation as a sequence generation problem, allowing the model to capture hierarchical relationships and cross-aspect dependencies between GO terms.
The system annotates proteins across all three primary Gene Ontology branches: Molecular Function (MF), Biological Process (BP), and Cellular Component (CC), achieving a weighted F_max score between 0.65 and 0.70.
How it works
GO-GPT couples a pre-trained protein language model encoder with an autoregressive transformer decoder. Protein sequences are encoded using ESM2-3B (facebook/esm2_t36_3B_UR50D), which generates sequence-level representation embeddings. These embeddings are fed to a 12-layer GPT decoder configured with prefix causal attention to generate GO annotations token by token.
The model contains approximately 3.2 billion parameters in total, comprising 3 billion parameters in the ESM2 encoder and 200 million parameters in the autoregressive decoder. It was trained on the dedicated dataset wanglab/gogpt-training-data.
Key facts
- Architecture
- ESM2-3B encoder coupled with a 12-layer GPT decoder using prefix causal attention
- Total parameters
- ~3.2 billion (~3B ESM2 encoder + 200M decoder)
- Protein encoder
- ESM2-3B (facebook/esm2_t36_3B_UR50D)
- Training dataset
- wanglab/gogpt-training-data
- Primary performance metric
- Weighted F_max of 0.65-0.70
What you can use it for
- Molecular Function Annotation
- Predict specific biochemical activities and molecular-level tasks carried out by uncharacterised protein sequences.
- Biological Process Prediction
- Identify the multi-step biological pathways and cellular programmes in which a protein participates.
- Cellular Component Localisation
- Determine the subcellular compartments, locations, and macromolecular complexes where a protein functions.
Run it here
We list this model but cannot run it yet: no library and no architectures — nothing can load it.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Formulates function prediction as an autoregressive generation task to capture hierarchical and cross-aspect dependencies among ontology terms.
- Achieves a weighted F_max performance of 0.65 to 0.70 across Gene Ontology evaluations.
- Jointly predicts all three major Gene Ontology categories: Molecular Function, Biological Process, and Cellular Component.
Watch out for
- Requires substantial compute resources to run inference due to its 3.2 billion parameter architecture based on ESM2-3B.
- Relies entirely on sequence embeddings from ESM2 without incorporating direct experimental structure or interaction network inputs.
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 · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use GO-GPT commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
Which Gene Ontology aspects does GO-GPT predict?
GO-GPT predicts terms across all three Gene Ontology aspects: Molecular Function (MF), Biological Process (BP), and Cellular Component (CC).
How does GO-GPT differ from standard discriminative annotation tools?
Rather than treating each functional term as an independent classification label, GO-GPT uses a generative autoregressive decoder with prefix causal attention to produce terms sequentially, capturing dependencies across the ontology hierarchy.
What underlying model provides the protein representations?
GO-GPT uses the 3-billion-parameter ESM2 model (facebook/esm2_t36_3B_UR50D) as its protein encoder.
How to cite it
Fallahpour, A., Seyed-Ahmadi, A., Idehpour, P., Ibrahim, O., Gupta, P., Naimer, J., Zhu, K., Shah, A., Ma, S., Adduri, A., Güloglu, T., Liu, N., Cui, H., Jain, A., de Castro, M., Fallahpour, A., Cembellin-Prieto, A., Stiles, J. S., Nemčko, F., Nevue, A. A., Moon, H. C., Sosnick, L., Markham, O., Duan, H., Lee, M. Y. Y., Salvador, A. F. M., Maddison, C. J., Thaiss, C. A., Ricci-Tam, C., Plosky, B. S., Burke, D. P., Hsu, P. D., Goodarzi, H., & Wang, B. (2026). BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning. bioRxiv. https://doi.org/10.64898/2026.03.19.712954
Written from GO-GPT'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.