CRISPR-GPT for Agentic Automation of Gene-editing Experiments

Kaixuan Huang, Yuanhao Qu, Henry C. Cousins, William A. Johnson, Di Yin, Mihir Shah, Denny Zhou, Russ B. Altman, Mengdi Wang, Le Cong, Altman, Russ B., Zhou, Denny, Wang, Mengdi, Cong, Le

Abstract

The introduction of genome engineering technology has transformed biomedical research, making it possible to make precise changes to genetic information. However, creating an efficient gene-editing system requires a deep understanding of CRISPR technology and the complex experimental systems under investigation. While Large Language Models (LLMs) have shown promise in various tasks, they often lack specific knowledge and struggle to accurately solve biological design problems. In this work, we introduce CRISPR-GPT, an LLM agent augmented with domain knowledge and external tools to automate and enhance the design process of CRISPR-based gene-editing experiments. CRISPR-GPT leverages the reasoning ability of LLMs to facilitate the process of selecting CRISPR systems, designing guide RNAs, recommending cellular delivery methods, drafting protocols, and designing validation experiments to confirm editing outcomes. We showcase the potential of CRISPR-GPT for assisting non-expert researchers with gene-editing experiments from scratch and validate the agent's effectiveness in a real-world use case. Furthermore, we explore the ethical and regulatory considerations associated with automated gene-editing design, highlighting the need for responsible and transparent use of these tools. Our work aims to bridge the gap between beginner biological researchers and CRISPR genome engineering techniques and demonstrate the potential of LLM agents in facilitating complex biological discovery tasks.

Introduction

While CRISPR technology has revolutionized biology, designing an effective experiment remains a high-barrier task. Researchers must navigate a maze of decisions: selecting the right Cas protein (e.g., Cas9 vs. Cas12a), designing highly specific guide RNAs (gRNAs) to avoid off-target effects, choosing delivery methods (like lentivirus or electroporation), and planning validation steps like PCR and Next-Generation Sequencing (NGS).

Overview of the CRISPR-GPT framework for experiment design and automation.

General-purpose LLMs like GPT-4 show promise in reasoning but often fail in this specialized domain. They frequently "hallucinate" biological sequences—providing gRNA strings that do not exist in the target genome—and lack the up-to-date technical knowledge required for wet-lab success. CRISPR-GPT bridges this gap by turning the LLM from a simple chatbot into an orchestrator of specialized biological tools.

Methodology: A Modular Agentic Architecture

The researchers developed a modular architecture to ensure reliability and accuracy. The system is built around four core components that work in concert to guide the user from a high-level request to a validated experimental protocol.

The modular architecture of CRISPR-GPT, including the Planner, Executor, Tool Provider, and Agent interface.
  1. LLM Planner: Decomposes a high-level user request (e.g., "I want to knock out EGFR") into a sequence of manageable sub-tasks.
  2. Task Executor: Operates as a state machine. This ensures the experiment design follows a logical, robust path (e.g., ensuring a target sequence is selected before designing primers).
  3. Tool Provider: Connects the agent to the real world via external APIs and software, such as Primer3 for primer design, BLAST for sequence alignment, and validated gRNA libraries.
  4. LLM Agent: The primary interface that interacts with the user, providing reasoning and collecting feedback.

The system uses ReAct (Reasoning and Acting) prompting, allowing the model to think through a problem before executing a tool-based action. It also includes a safety layer that filters for human germline editing and prevents the leakage of private genomic data by blocking sequences longer than 20 base pairs.

Key Insights and Results

A significant insight from this work is that LLMs should not be used to generate genomic sequences directly. Instead, they should orchestrate specialized software that handles sequences, effectively eliminating the hallucination problem. Furthermore, by constraining the LLM within a state machine, the authors prevented the "conversational drift" common in long LLM sessions.

Performance and Validation

  • Superior Accuracy: In human expert evaluations, CRISPR-GPT significantly outperformed GPT-3.5 and GPT-4 across accuracy, reasoning, and completeness.
  • Wet-Lab Success: The authors used CRISPR-GPT to design knockout experiments for four genes (TGFBR1, SNAI1, BAX, BCL2L1) in human A375 cells. The agent-designed gRNAs and protocols resulted in high editing efficiency, confirmed by NGS.
  • Accessibility: The study demonstrated that a scientist unfamiliar with CRISPR could use the agent to design a successful experiment from scratch, from system selection to final validation.

Limitations and Future Outlook

Currently, the agent can design individual components (gRNAs, primers) but cannot yet generate full, complex plasmid maps or modular vector designs. While the current version uses a predefined list of tasks for robustness, future iterations could allow the LLM to dynamically create new task types. Ultimately, the authors envision a future where CRISPR-GPT is integrated with "cloud labs" or robotic platforms, allowing for a fully automated pipeline from digital design to physical execution.

Summary updated

Abstract

The introduction of genome engineering technology has transformed biomedical research, making it possible to make precise changes to genetic information. However, creating an efficient gene-editing system requires a deep understanding of CRISPR technology, and the complex experimental systems under investigation. While Large Language Models (LLMs) have shown promise in various tasks, they often lack specific knowledge and struggle to accurately solve biological design problems. In this work, we introduce CRISPR-GPT, an LLM agent augmented with domain knowledge and external tools to automate and enhance the design process of CRISPR-based gene-editing experiments. CRISPR-GPT leverages the reasoning ability of LLMs to facilitate the process of selecting CRISPR systems, designing guide RNAs, recommending cellular delivery methods, drafting protocols, and designing validation experiments to confirm editing outcomes. We showcase the potential of CRISPR-GPT for assisting non-expert researchers with gene-editing experiments from scratch and validate the agent's effectiveness in a real-world use case. Furthermore, we explore the ethical and regulatory considerations associated with automated gene-editing design, highlighting the need for responsible and transparent use of these tools. Our work aims to bridge the gap between beginner biological researchers and CRISPR genome engineering techniques, and demonstrate the potential of LLM agents in facilitating complex biological discovery tasks. The published version of this draft is available at https://www.nature.com/articles/s41551-025-01463-z.
BibTeX
@inproceedings{huang2025crispr,
  author = {Kaixuan Huang and Yuanhao Qu and Henry C. Cousins and William A. Johnson and Di Yin and Mihir Shah and Denny Zhou and Russ B. Altman and Mengdi Wang and Le Cong and Altman, Russ B. and Zhou, Denny and Wang, Mengdi and Cong, Le},
  title = {{CRISPR-GPT for Agentic Automation of Gene-editing Experiments}},
  booktitle = {Nature Biomedical Engineering},
  year = {2025},
  doi = {10.1038/s41551-025-01463-z},
}
Figure