FROGENT: An End-to-End Full-process Drug Design Multi-Agent System

TL;DR

Overall, FROGENT not only achieves substantial gains in efficiency and accuracy, but also demonstrates the potential of LLM-based agentic systems to autonomously orchestrate drug development pipelines, reducing, or even replacing, reliance on manual, experience-driven human intervention.

Abstract

Drug discovery is a complex, multi-step pipeline that remains heavily dependent on manual, experience-driven operations; meanwhile, existing customized artificial intelligence tools are fragmented across web applications, desktop software, and code libraries, resulting in incompatible interfaces and inefficient, burdensome workflows. To overcome these challenges, we propose FROGENT, a full-process drug design multi-agent system that leverages the planning, reasoning, and tool-use capabilities of large language models (LLMs) to unify drug discovery within a closed-loop and autonomous framework. FROGENT is a collaborative multi-agent system comprising a central Orchestrate Agent for strategic workflow coordination and three distributed agents, Retrieve, Forge, and Gauge, that employ dynamic biochemical databases, extensible tool libraries, and task-specific computational models via the Model Context Protocol. This architecture enables end-to-end execution of complex drug discovery pipelines, covering target identification, small-molecule generation, peptide optimization, and retrosynthetic planning. Across eight benchmarks spanning core drug discovery tasks, FROGENT consistently outperforms six increasingly advanced ReAct-style agents. Case studies further demonstrate its practicality and generalization across real-world small-molecule and peptide design scenarios. Overall, FROGENT not only achieves substantial gains in efficiency and accuracy, but also demonstrates the potential of LLM-based agentic systems to autonomously orchestrate drug development pipelines, reducing, or even replacing, reliance on manual, experience-driven human intervention.

Figures (20)

• Figure 1: Architectural Overview of the Frogent Multi-Agent System. The system comprises four distinct agents coordinated by a central Orchestrate agent that handles planning, delegation, and feedback loop control. The Retrieve agent gathers essential data from literature and biological databases. The Forge agent is responsible for all generative tasks, including de novo molecular design, optimization, and retrosynthesis planning. The Gauge agent performs quantitative validation for both small molecules and peptides. The bottom panel illustrates the comprehensive suite of 11 drug discovery capabilities that Frogent can execute autonomously.
• Figure 2: Conceptual workflow of Frogent. The Orchestrate agent coordinates a multi-stage drug discovery campaign initiated by a single user query. The workflow progresses from initial knowledge gathering and hit generation to the core iterative optimization cycles, where collaboration between the Forge agent and Gauge agent refines both molecular properties and synthetic feasibility before the Orchestrate agent synthesizes a final solution.
• Figure 3: Overall Performance of Each Agent on Benchmarks. Performance evaluation of Frogent against six baseline models across eight diverse benchmarks spanning the full drug discovery workflow. The evaluated tasks span the entire workflow, from foundational knowledge assessment and structured data retrieval to critical hit-to-lead activities and mechanistic analysis. The evaluation culminates with complex, end-stage challenges and ensuring synthetic feasibility. Frogent consistently outperforms all baselines, highlighting the effectiveness of its integrated architecture and specialized tool suite in navigating the complexities of drug discovery.
• Figure 4: End-to-End Drug Design for Cardiomegaly-congestive Heart Failure. Following a user query, the Orchestrate agent tasks the Retrieve agent to identify PPAR$\gamma$ as a relevant target, which provides the necessary context for the Forge agent to generate initial drug candidates. An iterative optimization cycle driven by collaboration between the Forge agent and Gauge agent refines these candidates. The multi-agent process successfully rediscovers the known active molecule Luteolin and designs a novel, superior candidate, Compound (a), complete with a viable retrosynthesis plan.
• Figure 5: Lead Optimization for Glucagon-Like Peptide 1 Receptor. The Orchestrate agent initiates the campaign by tasking the Retrieve agent to acquire the GLP1R structure (PDB: 4ZGM) and the Glucagon sequence. It supervises a collaborative optimization cycle. The Forge agent generates new peptide variants, and the Gauge agent evaluates these peptide variants by predicting their binding affinity. Feedback from this evaluation guides the Forge agent’s subsequent design rounds, culminating in two novel peptide sequences with predicted binding scores superior to known potent agonists.