GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation

Proteins constitute a significant portion of the functional apparatus of biological systems, yet accurate protein function annotation remains a fundamental challenge in computational biology. Conventional bioinformatics pipelines yield reliable annotations for well-characterized proteins but fail to generalise to novel sequences lacking close homologs in reference databases. Recent deep learning methods that leverage learned representations of protein sequence and structure have improved upon traditional homology-based approaches, yet remain limited by the completeness of their training labels. Here we introduce GOAgent, a large language model agent that invokes a curated suite of sequence- and structure-derived bioinformatics tools and reasons over their outputs to predict Gene Ontology (GO) terms. We further evaluate a variant in which,rather than calling pre-defined tool functions, the agent is given a sandboxed execution environment together with the relevant input files and must orchestrate the analysis itself. We show that tool-augmented reasoning improves GO annotation quality over a zero-shot LLM baseline. We further train GOAgent end-to-end on multi-turn tool-calling rollouts via a policy optimization objective, yielding improved performance under both standard GO term prediction and a harder per-domain annotation task in which the agent must associate each predicted term with a specific functional region. While GOAgent does not match the raw accuracy of dedicated multimodal models for protein function prediction, it offers complementary strengths: model-agnostic extensibility, auditable tool-grounded reasoning, and a modular architecture in which new bioinformatics tools can be easily incorporated. We also release the code used to train multi turn tool calling agents on modal labs.