GO-GPT

GO-GPT is a generative model for protein function prediction that produces Gene Ontology (GO) annotations directly from an amino acid sequence. Function prediction is conventionally cast as multi-label classification over a fixed vocabulary of GO terms, which treats each term as an independent label and struggles to respect the ontology's hierarchical and cross-aspect structure. GO-GPT reframes the task as sequence generation: given a protein sequence, a decoder autoregressively emits the set of GO terms spanning the three GO aspects — Molecular Function, Biological Process, and Cellular Component.

Because it generates terms one after another, GO-GPT can condition each prediction on the terms it has already emitted, capturing the dependencies that link parent and child terms and that couple functions across the three GO aspects. This is well suited to a label space of tens of thousands of interrelated terms organized by subsumption and other relations, where the independence assumption of classifiers is a poor fit.

GO-GPT was developed by the Bo Wang Lab at the University of Toronto and the Vector Institute and released in March 2026. It serves both as a standalone function predictor and as the structured label-prior component of BioReason-Pro, the lab's multimodal reasoning system for protein annotation, where GO-GPT's generated terms supply candidate functions to the downstream reasoning pipeline.

#Key Features

• Generative GO prediction: GO-GPT treats function annotation as sequence generation rather than multi-label classification, autoregressively emitting GO terms so that each prediction is conditioned on those already produced.
• Cross-aspect coverage: A single model generates annotations across all three GO aspects — Molecular Function, Biological Process, and Cellular Component — capturing dependencies that span the aspects.
• ESM2-3B sequence encoder: Protein sequences are embedded with the ESM2-3B protein language model (facebook/esm2_t36_3B_UR50D), grounding generation in learned evolutionary and structural representations.
• Prefix causal decoder: A 12-layer GPT-style decoder with prefix causal attention conditions on the full encoded sequence while autoregressively generating the term set.
• Label prior for reasoning: GO-GPT supplies structured GO-term priors to BioReason-Pro, connecting a fast generative predictor to a downstream interpretable reasoning model.
• Open release: Model weights and the training dataset are released on HuggingFace under the Apache-2.0 license.

#Technical Details

GO-GPT couples an ESM2-3B encoder with a 12-layer decoder-only transformer that uses prefix causal attention, for roughly 3.2 billion parameters in total (the 3B-parameter ESM2 encoder plus an approximately 200M-parameter decoder). The encoder maps the input sequence to per-residue embeddings, and the decoder autoregressively generates the associated GO terms. Training uses the gogpt-training-data corpus of roughly 133,000 protein records (about 120,000 training and 13,300 validation), integrating protein sequences and functions from UniProt, GO annotations across the three aspects, InterPro domain annotations, and STRING interaction data. On GO term prediction, GO-GPT reaches a weighted F-max of 0.65-0.70.

#Applications

GO-GPT is aimed at researchers who need functional annotations for proteins that lack curated GO terms, including uncharacterized open reading frames from newly sequenced genomes and metagenomes. Its generative formulation makes it a natural source of candidate annotations for downstream pipelines: within BioReason-Pro it provides structured label priors that seed an interpretable reasoning model. Standalone, it can be used for proteome-scale preliminary annotation and for prioritizing proteins for experimental characterization.

#Impact

GO-GPT contributes to a shift in automated protein annotation from fixed-label classification toward generative modeling of the Gene Ontology, where the hierarchical and cross-aspect structure of GO is captured by the generation process itself rather than imposed as an external constraint. By releasing weights and training data under an open license on HuggingFace, the Bo Wang Lab lowers the barrier to reuse and to integration into larger systems such as BioReason-Pro. The model is documented in a bioRxiv preprint awaiting peer review; its GO term prediction results are reported in silico, and its code is distributed in a repository that currently carries no license.