PRIME: A Multi-Agent Environment for Orchestrating Dynamic Computational Workflows in Protein Engineering

Abstract

Artificial intelligence (AI) is revolutionizing protein engineering, yet its practical application is often hindered by a fragmented toolchain and the specialized expertise required to orchestrate complex computational workflows. To address these challenges, we have developed PRIME, an autonomous protein engineering multi-agent system. The core innovation of PRIME is dynamic workflow synthesis: it interprets high-level engineering objectives, reasons over a curated library of more than 65 validated protein tools, and autonomously constructs custom computational pathways. Crucially, by grounding every step in verifiable tool execution, PRIME mitigates the risk of model hallucination. It also automates the complete pipeline for developing specialized AI models, from data acquisition through training. In a benchmark of 213 multi-step protein engineering tasks, PRIME exhibited superior performance, successfully completing the majority of tasks where state-of-the-art, general-purpose AI agents fail. We validated its capabilities in demanding real-world applications, including the fully autonomous training of a machine learning classifier and the de novo design of a therapeutic antibody against SARS-CoV-2. By abstracting technical complexity, PRIME empowers scientists to execute sophisticated computational experiments with unprecedented flexibility, establishing a new paradigm for autonomous scientific discovery. The PRIME agent and source code will be made publicly available.